block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@suse.de>
   8  */
   9 #include <linux/config.h>
  10 #include <linux/module.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/hash.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16
  17 /*
  18  * tunables
  19  */
  20 static const int cfq_quantum = 4;               /* max queue in one round of service */
  21 static const int cfq_queued = 8;                /* minimum rq allocate limit per-queue*/
  22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  23 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
  24 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
  25
  26 static const int cfq_slice_sync = HZ / 10;
  27 static int cfq_slice_async = HZ / 25;
  28 static const int cfq_slice_async_rq = 2;
  29 static int cfq_slice_idle = HZ / 125;
  30
  31 #define CFQ_IDLE_GRACE          (HZ / 10)
  32 #define CFQ_SLICE_SCALE         (5)
  33
  34 #define CFQ_KEY_ASYNC           (0)
  35
  36 static DEFINE_SPINLOCK(cfq_exit_lock);
  37
  38 /*
  39  * for the hash of cfqq inside the cfqd
  40  */
  41 #define CFQ_QHASH_SHIFT         6
  42 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
  43 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
  44
  45 /*
  46  * for the hash of crq inside the cfqq
  47  */
  48 #define CFQ_MHASH_SHIFT         6
  49 #define CFQ_MHASH_BLOCK(sec)    ((sec) >> 3)
  50 #define CFQ_MHASH_ENTRIES       (1 << CFQ_MHASH_SHIFT)
  51 #define CFQ_MHASH_FN(sec)       hash_long(CFQ_MHASH_BLOCK(sec), CFQ_MHASH_SHIFT)
  52 #define rq_hash_key(rq)         ((rq)->sector + (rq)->nr_sectors)
  53 #define list_entry_hash(ptr)    hlist_entry((ptr), struct cfq_rq, hash)
  54
  55 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
  56 #define list_entry_fifo(ptr)    list_entry((ptr), struct request, queuelist)
  57
  58 #define RQ_DATA(rq)             (rq)->elevator_private
  59
  60 /*
  61  * rb-tree defines
  62  */
  63 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  64 #define RB_CLEAR(node)          do {    \
  65                 memset(node, 0, sizeof(*node)); \
  66 } while (0)
  67 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  68 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  69 #define rq_rb_key(rq)           (rq)->sector
  70
  71 static kmem_cache_t *crq_pool;
  72 static kmem_cache_t *cfq_pool;
  73 static kmem_cache_t *cfq_ioc_pool;
  74
  75 static atomic_t ioc_count = ATOMIC_INIT(0);
  76 static struct completion *ioc_gone;
  77
  78 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  79 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  80 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  81 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  82
  83 #define ASYNC                   (0)
  84 #define SYNC                    (1)
  85
  86 #define cfq_cfqq_dispatched(cfqq)       \
  87         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  88
  89 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  90
  91 #define cfq_cfqq_sync(cfqq)             \
  92         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  93
  94 #define sample_valid(samples)   ((samples) > 80)
  95
  96 /*
  97  * Per block device queue structure
  98  */
  99 struct cfq_data {
 100         request_queue_t *queue;
 101
 102         /*
 103          * rr list of queues with requests and the count of them
 104          */
 105         struct list_head rr_list[CFQ_PRIO_LISTS];
 106         struct list_head busy_rr;
 107         struct list_head cur_rr;
 108         struct list_head idle_rr;
 109         unsigned int busy_queues;
 110
 111         /*
 112          * non-ordered list of empty cfqq's
 113          */
 114         struct list_head empty_list;
 115
 116         /*
 117          * cfqq lookup hash
 118          */
 119         struct hlist_head *cfq_hash;
 120
 121         /*
 122          * global crq hash for all queues
 123          */
 124         struct hlist_head *crq_hash;
 125
 126         mempool_t *crq_pool;
 127
 128         int rq_in_driver;
 129         int hw_tag;
 130
 131         /*
 132          * schedule slice state info
 133          */
 134         /*
 135          * idle window management
 136          */
 137         struct timer_list idle_slice_timer;
 138         struct work_struct unplug_work;
 139
 140         struct cfq_queue *active_queue;
 141         struct cfq_io_context *active_cic;
 142         int cur_prio, cur_end_prio;
 143         unsigned int dispatch_slice;
 144
 145         struct timer_list idle_class_timer;
 146
 147         sector_t last_sector;
 148         unsigned long last_end_request;
 149
 150         unsigned int rq_starved;
 151
 152         /*
 153          * tunables, see top of file
 154          */
 155         unsigned int cfq_quantum;
 156         unsigned int cfq_queued;
 157         unsigned int cfq_fifo_expire[2];
 158         unsigned int cfq_back_penalty;
 159         unsigned int cfq_back_max;
 160         unsigned int cfq_slice[2];
 161         unsigned int cfq_slice_async_rq;
 162         unsigned int cfq_slice_idle;
 163
 164         struct list_head cic_list;
 165 };
 166
 167 /*
 168  * Per process-grouping structure
 169  */
 170 struct cfq_queue {
 171         /* reference count */
 172         atomic_t ref;
 173         /* parent cfq_data */
 174         struct cfq_data *cfqd;
 175         /* cfqq lookup hash */
 176         struct hlist_node cfq_hash;
 177         /* hash key */
 178         unsigned int key;
 179         /* on either rr or empty list of cfqd */
 180         struct list_head cfq_list;
 181         /* sorted list of pending requests */
 182         struct rb_root sort_list;
 183         /* if fifo isn't expired, next request to serve */
 184         struct cfq_rq *next_crq;
 185         /* requests queued in sort_list */
 186         int queued[2];
 187         /* currently allocated requests */
 188         int allocated[2];
 189         /* fifo list of requests in sort_list */
 190         struct list_head fifo;
 191
 192         unsigned long slice_start;
 193         unsigned long slice_end;
 194         unsigned long slice_left;
 195         unsigned long service_last;
 196
 197         /* number of requests that are on the dispatch list */
 198         int on_dispatch[2];
 199
 200         /* io prio of this group */
 201         unsigned short ioprio, org_ioprio;
 202         unsigned short ioprio_class, org_ioprio_class;
 203
 204         /* various state flags, see below */
 205         unsigned int flags;
 206 };
 207
 208 struct cfq_rq {
 209         struct rb_node rb_node;
 210         sector_t rb_key;
 211         struct request *request;
 212         struct hlist_node hash;
 213
 214         struct cfq_queue *cfq_queue;
 215         struct cfq_io_context *io_context;
 216
 217         unsigned int crq_flags;
 218 };
 219
 220 enum cfqq_state_flags {
 221         CFQ_CFQQ_FLAG_on_rr = 0,
 222         CFQ_CFQQ_FLAG_wait_request,
 223         CFQ_CFQQ_FLAG_must_alloc,
 224         CFQ_CFQQ_FLAG_must_alloc_slice,
 225         CFQ_CFQQ_FLAG_must_dispatch,
 226         CFQ_CFQQ_FLAG_fifo_expire,
 227         CFQ_CFQQ_FLAG_idle_window,
 228         CFQ_CFQQ_FLAG_prio_changed,
 229 };
 230
 231 #define CFQ_CFQQ_FNS(name)                                              \
 232 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 233 {                                                                       \
 234         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 235 }                                                                       \
 236 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 237 {                                                                       \
 238         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 239 }                                                                       \
 240 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 241 {                                                                       \
 242         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 243 }
 244
 245 CFQ_CFQQ_FNS(on_rr);
 246 CFQ_CFQQ_FNS(wait_request);
 247 CFQ_CFQQ_FNS(must_alloc);
 248 CFQ_CFQQ_FNS(must_alloc_slice);
 249 CFQ_CFQQ_FNS(must_dispatch);
 250 CFQ_CFQQ_FNS(fifo_expire);
 251 CFQ_CFQQ_FNS(idle_window);
 252 CFQ_CFQQ_FNS(prio_changed);
 253 #undef CFQ_CFQQ_FNS
 254
 255 enum cfq_rq_state_flags {
 256         CFQ_CRQ_FLAG_is_sync = 0,
 257 };
 258
 259 #define CFQ_CRQ_FNS(name)                                               \
 260 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 261 {                                                                       \
 262         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 263 }                                                                       \
 264 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 265 {                                                                       \
 266         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 267 }                                                                       \
 268 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 269 {                                                                       \
 270         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 271 }
 272
 273 CFQ_CRQ_FNS(is_sync);
 274 #undef CFQ_CRQ_FNS
 275
 276 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 277 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 278 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 279
 280 /*
 281  * lots of deadline iosched dupes, can be abstracted later...
 282  */
 283 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 284 {
 285         hlist_del_init(&crq->hash);
 286 }
 287
 288 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 289 {
 290         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 291
 292         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 293 }
 294
 295 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 296 {
 297         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 298         struct hlist_node *entry, *next;
 299
 300         hlist_for_each_safe(entry, next, hash_list) {
 301                 struct cfq_rq *crq = list_entry_hash(entry);
 302                 struct request *__rq = crq->request;
 303
 304                 if (!rq_mergeable(__rq)) {
 305                         cfq_del_crq_hash(crq);
 306                         continue;
 307                 }
 308
 309                 if (rq_hash_key(__rq) == offset)
 310                         return __rq;
 311         }
 312
 313         return NULL;
 314 }
 315
 316 /*
 317  * scheduler run of queue, if there are requests pending and no one in the
 318  * driver that will restart queueing
 319  */
 320 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 321 {
 322         if (cfqd->busy_queues)
 323                 kblockd_schedule_work(&cfqd->unplug_work);
 324 }
 325
 326 static int cfq_queue_empty(request_queue_t *q)
 327 {
 328         struct cfq_data *cfqd = q->elevator->elevator_data;
 329
 330         return !cfqd->busy_queues;
 331 }
 332
 333 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 334 {
 335         if (rw == READ || rw == WRITE_SYNC)
 336                 return task->pid;
 337
 338         return CFQ_KEY_ASYNC;
 339 }
 340
 341 /*
 342  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 343  * We choose the request that is closest to the head right now. Distance
 344  * behind the head is penalized and only allowed to a certain extent.
 345  */
 346 static struct cfq_rq *
 347 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 348 {
 349         sector_t last, s1, s2, d1 = 0, d2 = 0;
 350         unsigned long back_max;
 351 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 352 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 353         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 354
 355         if (crq1 == NULL || crq1 == crq2)
 356                 return crq2;
 357         if (crq2 == NULL)
 358                 return crq1;
 359
 360         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 361                 return crq1;
 362         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 363                 return crq2;
 364
 365         s1 = crq1->request->sector;
 366         s2 = crq2->request->sector;
 367
 368         last = cfqd->last_sector;
 369
 370         /*
 371          * by definition, 1KiB is 2 sectors
 372          */
 373         back_max = cfqd->cfq_back_max * 2;
 374
 375         /*
 376          * Strict one way elevator _except_ in the case where we allow
 377          * short backward seeks which are biased as twice the cost of a
 378          * similar forward seek.
 379          */
 380         if (s1 >= last)
 381                 d1 = s1 - last;
 382         else if (s1 + back_max >= last)
 383                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 384         else
 385                 wrap |= CFQ_RQ1_WRAP;
 386
 387         if (s2 >= last)
 388                 d2 = s2 - last;
 389         else if (s2 + back_max >= last)
 390                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 391         else
 392                 wrap |= CFQ_RQ2_WRAP;
 393
 394         /* Found required data */
 395
 396         /*
 397          * By doing switch() on the bit mask "wrap" we avoid having to
 398          * check two variables for all permutations: --> faster!
 399          */
 400         switch (wrap) {
 401         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 402                 if (d1 < d2)
 403                         return crq1;
 404                 else if (d2 < d1)
 405                         return crq2;
 406                 else {
 407                         if (s1 >= s2)
 408                                 return crq1;
 409                         else
 410                                 return crq2;
 411                 }
 412
 413         case CFQ_RQ2_WRAP:
 414                 return crq1;
 415         case CFQ_RQ1_WRAP:
 416                 return crq2;
 417         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 418         default:
 419                 /*
 420                  * Since both rqs are wrapped,
 421                  * start with the one that's further behind head
 422                  * (--> only *one* back seek required),
 423                  * since back seek takes more time than forward.
 424                  */
 425                 if (s1 <= s2)
 426                         return crq1;
 427                 else
 428                         return crq2;
 429         }
 430 }
 431
 432 /*
 433  * would be nice to take fifo expire time into account as well
 434  */
 435 static struct cfq_rq *
 436 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 437                   struct cfq_rq *last)
 438 {
 439         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 440         struct rb_node *rbnext, *rbprev;
 441
 442         if (!(rbnext = rb_next(&last->rb_node))) {
 443                 rbnext = rb_first(&cfqq->sort_list);
 444                 if (rbnext == &last->rb_node)
 445                         rbnext = NULL;
 446         }
 447
 448         rbprev = rb_prev(&last->rb_node);
 449
 450         if (rbprev)
 451                 crq_prev = rb_entry_crq(rbprev);
 452         if (rbnext)
 453                 crq_next = rb_entry_crq(rbnext);
 454
 455         return cfq_choose_req(cfqd, crq_next, crq_prev);
 456 }
 457
 458 static void cfq_update_next_crq(struct cfq_rq *crq)
 459 {
 460         struct cfq_queue *cfqq = crq->cfq_queue;
 461
 462         if (cfqq->next_crq == crq)
 463                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 464 }
 465
 466 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 467 {
 468         struct cfq_data *cfqd = cfqq->cfqd;
 469         struct list_head *list, *entry;
 470
 471         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 472
 473         list_del(&cfqq->cfq_list);
 474
 475         if (cfq_class_rt(cfqq))
 476                 list = &cfqd->cur_rr;
 477         else if (cfq_class_idle(cfqq))
 478                 list = &cfqd->idle_rr;
 479         else {
 480                 /*
 481                  * if cfqq has requests in flight, don't allow it to be
 482                  * found in cfq_set_active_queue before it has finished them.
 483                  * this is done to increase fairness between a process that
 484                  * has lots of io pending vs one that only generates one
 485                  * sporadically or synchronously
 486                  */
 487                 if (cfq_cfqq_dispatched(cfqq))
 488                         list = &cfqd->busy_rr;
 489                 else
 490                         list = &cfqd->rr_list[cfqq->ioprio];
 491         }
 492
 493         /*
 494          * if queue was preempted, just add to front to be fair. busy_rr
 495          * isn't sorted, but insert at the back for fairness.
 496          */
 497         if (preempted || list == &cfqd->busy_rr) {
 498                 if (preempted)
 499                         list = list->prev;
 500
 501                 list_add_tail(&cfqq->cfq_list, list);
 502                 return;
 503         }
 504
 505         /*
 506          * sort by when queue was last serviced
 507          */
 508         entry = list;
 509         while ((entry = entry->prev) != list) {
 510                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 511
 512                 if (!__cfqq->service_last)
 513                         break;
 514                 if (time_before(__cfqq->service_last, cfqq->service_last))
 515                         break;
 516         }
 517
 518         list_add(&cfqq->cfq_list, entry);
 519 }
 520
 521 /*
 522  * add to busy list of queues for service, trying to be fair in ordering
 523  * the pending list according to last request service
 524  */
 525 static inline void
 526 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 527 {
 528         BUG_ON(cfq_cfqq_on_rr(cfqq));
 529         cfq_mark_cfqq_on_rr(cfqq);
 530         cfqd->busy_queues++;
 531
 532         cfq_resort_rr_list(cfqq, 0);
 533 }
 534
 535 static inline void
 536 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 537 {
 538         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 539         cfq_clear_cfqq_on_rr(cfqq);
 540         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 541
 542         BUG_ON(!cfqd->busy_queues);
 543         cfqd->busy_queues--;
 544 }
 545
 546 /*
 547  * rb tree support functions
 548  */
 549 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 550 {
 551         struct cfq_queue *cfqq = crq->cfq_queue;
 552         struct cfq_data *cfqd = cfqq->cfqd;
 553         const int sync = cfq_crq_is_sync(crq);
 554
 555         BUG_ON(!cfqq->queued[sync]);
 556         cfqq->queued[sync]--;
 557
 558         cfq_update_next_crq(crq);
 559
 560         rb_erase(&crq->rb_node, &cfqq->sort_list);
 561
 562         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 563                 cfq_del_cfqq_rr(cfqd, cfqq);
 564 }
 565
 566 static struct cfq_rq *
 567 __cfq_add_crq_rb(struct cfq_rq *crq)
 568 {
 569         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 570         struct rb_node *parent = NULL;
 571         struct cfq_rq *__crq;
 572
 573         while (*p) {
 574                 parent = *p;
 575                 __crq = rb_entry_crq(parent);
 576
 577                 if (crq->rb_key < __crq->rb_key)
 578                         p = &(*p)->rb_left;
 579                 else if (crq->rb_key > __crq->rb_key)
 580                         p = &(*p)->rb_right;
 581                 else
 582                         return __crq;
 583         }
 584
 585         rb_link_node(&crq->rb_node, parent, p);
 586         return NULL;
 587 }
 588
 589 static void cfq_add_crq_rb(struct cfq_rq *crq)
 590 {
 591         struct cfq_queue *cfqq = crq->cfq_queue;
 592         struct cfq_data *cfqd = cfqq->cfqd;
 593         struct request *rq = crq->request;
 594         struct cfq_rq *__alias;
 595
 596         crq->rb_key = rq_rb_key(rq);
 597         cfqq->queued[cfq_crq_is_sync(crq)]++;
 598
 599         /*
 600          * looks a little odd, but the first insert might return an alias.
 601          * if that happens, put the alias on the dispatch list
 602          */
 603         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 604                 cfq_dispatch_insert(cfqd->queue, __alias);
 605
 606         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 607
 608         if (!cfq_cfqq_on_rr(cfqq))
 609                 cfq_add_cfqq_rr(cfqd, cfqq);
 610
 611         /*
 612          * check if this request is a better next-serve candidate
 613          */
 614         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 615 }
 616
 617 static inline void
 618 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 619 {
 620         rb_erase(&crq->rb_node, &cfqq->sort_list);
 621         cfqq->queued[cfq_crq_is_sync(crq)]--;
 622
 623         cfq_add_crq_rb(crq);
 624 }
 625
 626 static struct request *
 627 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 628 {
 629         struct task_struct *tsk = current;
 630         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 631         struct cfq_queue *cfqq;
 632         struct rb_node *n;
 633         sector_t sector;
 634
 635         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 636         if (!cfqq)
 637                 goto out;
 638
 639         sector = bio->bi_sector + bio_sectors(bio);
 640         n = cfqq->sort_list.rb_node;
 641         while (n) {
 642                 struct cfq_rq *crq = rb_entry_crq(n);
 643
 644                 if (sector < crq->rb_key)
 645                         n = n->rb_left;
 646                 else if (sector > crq->rb_key)
 647                         n = n->rb_right;
 648                 else
 649                         return crq->request;
 650         }
 651
 652 out:
 653         return NULL;
 654 }
 655
 656 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 657 {
 658         struct cfq_data *cfqd = q->elevator->elevator_data;
 659
 660         cfqd->rq_in_driver++;
 661
 662         /*
 663          * If the depth is larger 1, it really could be queueing. But lets
 664          * make the mark a little higher - idling could still be good for
 665          * low queueing, and a low queueing number could also just indicate
 666          * a SCSI mid layer like behaviour where limit+1 is often seen.
 667          */
 668         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 669                 cfqd->hw_tag = 1;
 670 }
 671
 672 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 673 {
 674         struct cfq_data *cfqd = q->elevator->elevator_data;
 675
 676         WARN_ON(!cfqd->rq_in_driver);
 677         cfqd->rq_in_driver--;
 678 }
 679
 680 static void cfq_remove_request(struct request *rq)
 681 {
 682         struct cfq_rq *crq = RQ_DATA(rq);
 683
 684         list_del_init(&rq->queuelist);
 685         cfq_del_crq_rb(crq);
 686         cfq_del_crq_hash(crq);
 687 }
 688
 689 static int
 690 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 691 {
 692         struct cfq_data *cfqd = q->elevator->elevator_data;
 693         struct request *__rq;
 694         int ret;
 695
 696         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 697         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 698                 ret = ELEVATOR_BACK_MERGE;
 699                 goto out;
 700         }
 701
 702         __rq = cfq_find_rq_fmerge(cfqd, bio);
 703         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 704                 ret = ELEVATOR_FRONT_MERGE;
 705                 goto out;
 706         }
 707
 708         return ELEVATOR_NO_MERGE;
 709 out:
 710         *req = __rq;
 711         return ret;
 712 }
 713
 714 static void cfq_merged_request(request_queue_t *q, struct request *req)
 715 {
 716         struct cfq_data *cfqd = q->elevator->elevator_data;
 717         struct cfq_rq *crq = RQ_DATA(req);
 718
 719         cfq_del_crq_hash(crq);
 720         cfq_add_crq_hash(cfqd, crq);
 721
 722         if (rq_rb_key(req) != crq->rb_key) {
 723                 struct cfq_queue *cfqq = crq->cfq_queue;
 724
 725                 cfq_update_next_crq(crq);
 726                 cfq_reposition_crq_rb(cfqq, crq);
 727         }
 728 }
 729
 730 static void
 731 cfq_merged_requests(request_queue_t *q, struct request *rq,
 732                     struct request *next)
 733 {
 734         cfq_merged_request(q, rq);
 735
 736         /*
 737          * reposition in fifo if next is older than rq
 738          */
 739         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 740             time_before(next->start_time, rq->start_time))
 741                 list_move(&rq->queuelist, &next->queuelist);
 742
 743         cfq_remove_request(next);
 744 }
 745
 746 static inline void
 747 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 748 {
 749         if (cfqq) {
 750                 /*
 751                  * stop potential idle class queues waiting service
 752                  */
 753                 del_timer(&cfqd->idle_class_timer);
 754
 755                 cfqq->slice_start = jiffies;
 756                 cfqq->slice_end = 0;
 757                 cfqq->slice_left = 0;
 758                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 759                 cfq_clear_cfqq_fifo_expire(cfqq);
 760         }
 761
 762         cfqd->active_queue = cfqq;
 763 }
 764
 765 /*
 766  * current cfqq expired its slice (or was too idle), select new one
 767  */
 768 static void
 769 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 770                     int preempted)
 771 {
 772         unsigned long now = jiffies;
 773
 774         if (cfq_cfqq_wait_request(cfqq))
 775                 del_timer(&cfqd->idle_slice_timer);
 776
 777         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 778                 cfqq->service_last = now;
 779                 cfq_schedule_dispatch(cfqd);
 780         }
 781
 782         cfq_clear_cfqq_must_dispatch(cfqq);
 783         cfq_clear_cfqq_wait_request(cfqq);
 784
 785         /*
 786          * store what was left of this slice, if the queue idled out
 787          * or was preempted
 788          */
 789         if (time_after(cfqq->slice_end, now))
 790                 cfqq->slice_left = cfqq->slice_end - now;
 791         else
 792                 cfqq->slice_left = 0;
 793
 794         if (cfq_cfqq_on_rr(cfqq))
 795                 cfq_resort_rr_list(cfqq, preempted);
 796
 797         if (cfqq == cfqd->active_queue)
 798                 cfqd->active_queue = NULL;
 799
 800         if (cfqd->active_cic) {
 801                 put_io_context(cfqd->active_cic->ioc);
 802                 cfqd->active_cic = NULL;
 803         }
 804
 805         cfqd->dispatch_slice = 0;
 806 }
 807
 808 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 809 {
 810         struct cfq_queue *cfqq = cfqd->active_queue;
 811
 812         if (cfqq)
 813                 __cfq_slice_expired(cfqd, cfqq, preempted);
 814 }
 815
 816 /*
 817  * 0
 818  * 0,1
 819  * 0,1,2
 820  * 0,1,2,3
 821  * 0,1,2,3,4
 822  * 0,1,2,3,4,5
 823  * 0,1,2,3,4,5,6
 824  * 0,1,2,3,4,5,6,7
 825  */
 826 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 827 {
 828         int prio, wrap;
 829
 830         prio = -1;
 831         wrap = 0;
 832         do {
 833                 int p;
 834
 835                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 836                         if (!list_empty(&cfqd->rr_list[p])) {
 837                                 prio = p;
 838                                 break;
 839                         }
 840                 }
 841
 842                 if (prio != -1)
 843                         break;
 844                 cfqd->cur_prio = 0;
 845                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 846                         cfqd->cur_end_prio = 0;
 847                         if (wrap)
 848                                 break;
 849                         wrap = 1;
 850                 }
 851         } while (1);
 852
 853         if (unlikely(prio == -1))
 854                 return -1;
 855
 856         BUG_ON(prio >= CFQ_PRIO_LISTS);
 857
 858         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 859
 860         cfqd->cur_prio = prio + 1;
 861         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 862                 cfqd->cur_end_prio = cfqd->cur_prio;
 863                 cfqd->cur_prio = 0;
 864         }
 865         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 866                 cfqd->cur_prio = 0;
 867                 cfqd->cur_end_prio = 0;
 868         }
 869
 870         return prio;
 871 }
 872
 873 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 874 {
 875         struct cfq_queue *cfqq = NULL;
 876
 877         /*
 878          * if current list is non-empty, grab first entry. if it is empty,
 879          * get next prio level and grab first entry then if any are spliced
 880          */
 881         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 882                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 883
 884         /*
 885          * If no new queues are available, check if the busy list has some
 886          * before falling back to idle io.
 887          */
 888         if (!cfqq && !list_empty(&cfqd->busy_rr))
 889                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 890
 891         /*
 892          * if we have idle queues and no rt or be queues had pending
 893          * requests, either allow immediate service if the grace period
 894          * has passed or arm the idle grace timer
 895          */
 896         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 897                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 898
 899                 if (time_after_eq(jiffies, end))
 900                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 901                 else
 902                         mod_timer(&cfqd->idle_class_timer, end);
 903         }
 904
 905         __cfq_set_active_queue(cfqd, cfqq);
 906         return cfqq;
 907 }
 908
 909 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
 910
 911 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 912
 913 {
 914         struct cfq_io_context *cic;
 915         unsigned long sl;
 916
 917         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 918         WARN_ON(cfqq != cfqd->active_queue);
 919
 920         /*
 921          * idle is disabled, either manually or by past process history
 922          */
 923         if (!cfqd->cfq_slice_idle)
 924                 return 0;
 925         if (!cfq_cfqq_idle_window(cfqq))
 926                 return 0;
 927         /*
 928          * task has exited, don't wait
 929          */
 930         cic = cfqd->active_cic;
 931         if (!cic || !cic->ioc->task)
 932                 return 0;
 933
 934         cfq_mark_cfqq_must_dispatch(cfqq);
 935         cfq_mark_cfqq_wait_request(cfqq);
 936
 937         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 938
 939         /*
 940          * we don't want to idle for seeks, but we do want to allow
 941          * fair distribution of slice time for a process doing back-to-back
 942          * seeks. so allow a little bit of time for him to submit a new rq
 943          */
 944         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 945                 sl = 2;
 946
 947         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 948         return 1;
 949 }
 950
 951 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 952 {
 953         struct cfq_data *cfqd = q->elevator->elevator_data;
 954         struct cfq_queue *cfqq = crq->cfq_queue;
 955
 956         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 957         cfq_remove_request(crq->request);
 958         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 959         elv_dispatch_sort(q, crq->request);
 960 }
 961
 962 /*
 963  * return expired entry, or NULL to just start from scratch in rbtree
 964  */
 965 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 966 {
 967         struct cfq_data *cfqd = cfqq->cfqd;
 968         struct request *rq;
 969         struct cfq_rq *crq;
 970
 971         if (cfq_cfqq_fifo_expire(cfqq))
 972                 return NULL;
 973
 974         if (!list_empty(&cfqq->fifo)) {
 975                 int fifo = cfq_cfqq_class_sync(cfqq);
 976
 977                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 978                 rq = crq->request;
 979                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 980                         cfq_mark_cfqq_fifo_expire(cfqq);
 981                         return crq;
 982                 }
 983         }
 984
 985         return NULL;
 986 }
 987
 988 /*
 989  * Scale schedule slice based on io priority. Use the sync time slice only
 990  * if a queue is marked sync and has sync io queued. A sync queue with async
 991  * io only, should not get full sync slice length.
 992  */
 993 static inline int
 994 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 995 {
 996         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 997
 998         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 999
1000         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1001 }
1002
1003 static inline void
1004 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1005 {
1006         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1007 }
1008
1009 static inline int
1010 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1011 {
1012         const int base_rq = cfqd->cfq_slice_async_rq;
1013
1014         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1015
1016         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1017 }
1018
1019 /*
1020  * get next queue for service
1021  */
1022 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1023 {
1024         unsigned long now = jiffies;
1025         struct cfq_queue *cfqq;
1026
1027         cfqq = cfqd->active_queue;
1028         if (!cfqq)
1029                 goto new_queue;
1030
1031         /*
1032          * slice has expired
1033          */
1034         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1035                 goto expire;
1036
1037         /*
1038          * if queue has requests, dispatch one. if not, check if
1039          * enough slice is left to wait for one
1040          */
1041         if (!RB_EMPTY(&cfqq->sort_list))
1042                 goto keep_queue;
1043         else if (cfq_cfqq_dispatched(cfqq)) {
1044                 cfqq = NULL;
1045                 goto keep_queue;
1046         } else if (cfq_cfqq_class_sync(cfqq)) {
1047                 if (cfq_arm_slice_timer(cfqd, cfqq))
1048                         return NULL;
1049         }
1050
1051 expire:
1052         cfq_slice_expired(cfqd, 0);
1053 new_queue:
1054         cfqq = cfq_set_active_queue(cfqd);
1055 keep_queue:
1056         return cfqq;
1057 }
1058
1059 static int
1060 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1061                         int max_dispatch)
1062 {
1063         int dispatched = 0;
1064
1065         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1066
1067         do {
1068                 struct cfq_rq *crq;
1069
1070                 /*
1071                  * follow expired path, else get first next available
1072                  */
1073                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1074                         crq = cfqq->next_crq;
1075
1076                 /*
1077                  * finally, insert request into driver dispatch list
1078                  */
1079                 cfq_dispatch_insert(cfqd->queue, crq);
1080
1081                 cfqd->dispatch_slice++;
1082                 dispatched++;
1083
1084                 if (!cfqd->active_cic) {
1085                         atomic_inc(&crq->io_context->ioc->refcount);
1086                         cfqd->active_cic = crq->io_context;
1087                 }
1088
1089                 if (RB_EMPTY(&cfqq->sort_list))
1090                         break;
1091
1092         } while (dispatched < max_dispatch);
1093
1094         /*
1095          * if slice end isn't set yet, set it.
1096          */
1097         if (!cfqq->slice_end)
1098                 cfq_set_prio_slice(cfqd, cfqq);
1099
1100         /*
1101          * expire an async queue immediately if it has used up its slice. idle
1102          * queue always expire after 1 dispatch round.
1103          */
1104         if ((!cfq_cfqq_sync(cfqq) &&
1105             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1106             cfq_class_idle(cfqq) ||
1107             !cfq_cfqq_idle_window(cfqq))
1108                 cfq_slice_expired(cfqd, 0);
1109
1110         return dispatched;
1111 }
1112
1113 static int
1114 cfq_forced_dispatch_cfqqs(struct list_head *list)
1115 {
1116         struct cfq_queue *cfqq, *next;
1117         struct cfq_rq *crq;
1118         int dispatched;
1119
1120         dispatched = 0;
1121         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1122                 while ((crq = cfqq->next_crq)) {
1123                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1124                         dispatched++;
1125                 }
1126                 BUG_ON(!list_empty(&cfqq->fifo));
1127         }
1128
1129         return dispatched;
1130 }
1131
1132 static int
1133 cfq_forced_dispatch(struct cfq_data *cfqd)
1134 {
1135         int i, dispatched = 0;
1136
1137         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1138                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1139
1140         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1141         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1142         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1143
1144         cfq_slice_expired(cfqd, 0);
1145
1146         BUG_ON(cfqd->busy_queues);
1147
1148         return dispatched;
1149 }
1150
1151 static int
1152 cfq_dispatch_requests(request_queue_t *q, int force)
1153 {
1154         struct cfq_data *cfqd = q->elevator->elevator_data;
1155         struct cfq_queue *cfqq, *prev_cfqq;
1156         int dispatched;
1157
1158         if (!cfqd->busy_queues)
1159                 return 0;
1160
1161         if (unlikely(force))
1162                 return cfq_forced_dispatch(cfqd);
1163
1164         dispatched = 0;
1165         prev_cfqq = NULL;
1166         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1167                 int max_dispatch;
1168
1169                 /*
1170                  * Don't repeat dispatch from the previous queue.
1171                  */
1172                 if (prev_cfqq == cfqq)
1173                         break;
1174
1175                 cfq_clear_cfqq_must_dispatch(cfqq);
1176                 cfq_clear_cfqq_wait_request(cfqq);
1177                 del_timer(&cfqd->idle_slice_timer);
1178
1179                 max_dispatch = cfqd->cfq_quantum;
1180                 if (cfq_class_idle(cfqq))
1181                         max_dispatch = 1;
1182
1183                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1184
1185                 /*
1186                  * If the dispatch cfqq has idling enabled and is still
1187                  * the active queue, break out.
1188                  */
1189                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1190                         break;
1191
1192                 prev_cfqq = cfqq;
1193         }
1194
1195         return dispatched;
1196 }
1197
1198 /*
1199  * task holds one reference to the queue, dropped when task exits. each crq
1200  * in-flight on this queue also holds a reference, dropped when crq is freed.
1201  *
1202  * queue lock must be held here.
1203  */
1204 static void cfq_put_queue(struct cfq_queue *cfqq)
1205 {
1206         struct cfq_data *cfqd = cfqq->cfqd;
1207
1208         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1209
1210         if (!atomic_dec_and_test(&cfqq->ref))
1211                 return;
1212
1213         BUG_ON(rb_first(&cfqq->sort_list));
1214         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1215         BUG_ON(cfq_cfqq_on_rr(cfqq));
1216
1217         if (unlikely(cfqd->active_queue == cfqq))
1218                 __cfq_slice_expired(cfqd, cfqq, 0);
1219
1220         /*
1221          * it's on the empty list and still hashed
1222          */
1223         list_del(&cfqq->cfq_list);
1224         hlist_del(&cfqq->cfq_hash);
1225         kmem_cache_free(cfq_pool, cfqq);
1226 }
1227
1228 static inline struct cfq_queue *
1229 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1230                     const int hashval)
1231 {
1232         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1233         struct hlist_node *entry;
1234         struct cfq_queue *__cfqq;
1235
1236         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1237                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1238
1239                 if (__cfqq->key == key && (__p == prio || !prio))
1240                         return __cfqq;
1241         }
1242
1243         return NULL;
1244 }
1245
1246 static struct cfq_queue *
1247 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1248 {
1249         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1250 }
1251
1252 static void cfq_free_io_context(struct io_context *ioc)
1253 {
1254         struct cfq_io_context *__cic;
1255         struct rb_node *n;
1256         int freed = 0;
1257
1258         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1259                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1260                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1261                 kmem_cache_free(cfq_ioc_pool, __cic);
1262                 freed++;
1263         }
1264
1265         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1266                 complete(ioc_gone);
1267 }
1268
1269 static void cfq_trim(struct io_context *ioc)
1270 {
1271         ioc->set_ioprio = NULL;
1272         cfq_free_io_context(ioc);
1273 }
1274
1275 /*
1276  * Called with interrupts disabled
1277  */
1278 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1279 {
1280         struct cfq_data *cfqd = cic->key;
1281         request_queue_t *q;
1282
1283         if (!cfqd)
1284                 return;
1285
1286         q = cfqd->queue;
1287
1288         WARN_ON(!irqs_disabled());
1289
1290         spin_lock(q->queue_lock);
1291
1292         if (cic->cfqq[ASYNC]) {
1293                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1294                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1295                 cfq_put_queue(cic->cfqq[ASYNC]);
1296                 cic->cfqq[ASYNC] = NULL;
1297         }
1298
1299         if (cic->cfqq[SYNC]) {
1300                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1301                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1302                 cfq_put_queue(cic->cfqq[SYNC]);
1303                 cic->cfqq[SYNC] = NULL;
1304         }
1305
1306         cic->key = NULL;
1307         list_del_init(&cic->queue_list);
1308         spin_unlock(q->queue_lock);
1309 }
1310
1311 static void cfq_exit_io_context(struct io_context *ioc)
1312 {
1313         struct cfq_io_context *__cic;
1314         unsigned long flags;
1315         struct rb_node *n;
1316
1317         /*
1318          * put the reference this task is holding to the various queues
1319          */
1320         spin_lock_irqsave(&cfq_exit_lock, flags);
1321
1322         n = rb_first(&ioc->cic_root);
1323         while (n != NULL) {
1324                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1325
1326                 cfq_exit_single_io_context(__cic);
1327                 n = rb_next(n);
1328         }
1329
1330         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1331 }
1332
1333 static struct cfq_io_context *
1334 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1335 {
1336         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1337
1338         if (cic) {
1339                 memset(cic, 0, sizeof(*cic));
1340                 cic->last_end_request = jiffies;
1341                 INIT_LIST_HEAD(&cic->queue_list);
1342                 cic->dtor = cfq_free_io_context;
1343                 cic->exit = cfq_exit_io_context;
1344                 atomic_inc(&ioc_count);
1345         }
1346
1347         return cic;
1348 }
1349
1350 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1351 {
1352         struct task_struct *tsk = current;
1353         int ioprio_class;
1354
1355         if (!cfq_cfqq_prio_changed(cfqq))
1356                 return;
1357
1358         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1359         switch (ioprio_class) {
1360                 default:
1361                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1362                 case IOPRIO_CLASS_NONE:
1363                         /*
1364                          * no prio set, place us in the middle of the BE classes
1365                          */
1366                         cfqq->ioprio = task_nice_ioprio(tsk);
1367                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1368                         break;
1369                 case IOPRIO_CLASS_RT:
1370                         cfqq->ioprio = task_ioprio(tsk);
1371                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1372                         break;
1373                 case IOPRIO_CLASS_BE:
1374                         cfqq->ioprio = task_ioprio(tsk);
1375                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1376                         break;
1377                 case IOPRIO_CLASS_IDLE:
1378                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1379                         cfqq->ioprio = 7;
1380                         cfq_clear_cfqq_idle_window(cfqq);
1381                         break;
1382         }
1383
1384         /*
1385          * keep track of original prio settings in case we have to temporarily
1386          * elevate the priority of this queue
1387          */
1388         cfqq->org_ioprio = cfqq->ioprio;
1389         cfqq->org_ioprio_class = cfqq->ioprio_class;
1390
1391         if (cfq_cfqq_on_rr(cfqq))
1392                 cfq_resort_rr_list(cfqq, 0);
1393
1394         cfq_clear_cfqq_prio_changed(cfqq);
1395 }
1396
1397 static inline void changed_ioprio(struct cfq_io_context *cic)
1398 {
1399         struct cfq_data *cfqd = cic->key;
1400         struct cfq_queue *cfqq;
1401
1402         if (unlikely(!cfqd))
1403                 return;
1404
1405         spin_lock(cfqd->queue->queue_lock);
1406
1407         cfqq = cic->cfqq[ASYNC];
1408         if (cfqq) {
1409                 struct cfq_queue *new_cfqq;
1410                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1411                                          GFP_ATOMIC);
1412                 if (new_cfqq) {
1413                         cic->cfqq[ASYNC] = new_cfqq;
1414                         cfq_put_queue(cfqq);
1415                 }
1416         }
1417
1418         cfqq = cic->cfqq[SYNC];
1419         if (cfqq)
1420                 cfq_mark_cfqq_prio_changed(cfqq);
1421
1422         spin_unlock(cfqd->queue->queue_lock);
1423 }
1424
1425 /*
1426  * callback from sys_ioprio_set, irqs are disabled
1427  */
1428 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1429 {
1430         struct cfq_io_context *cic;
1431         struct rb_node *n;
1432
1433         spin_lock(&cfq_exit_lock);
1434
1435         n = rb_first(&ioc->cic_root);
1436         while (n != NULL) {
1437                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1438
1439                 changed_ioprio(cic);
1440                 n = rb_next(n);
1441         }
1442
1443         spin_unlock(&cfq_exit_lock);
1444
1445         return 0;
1446 }
1447
1448 static struct cfq_queue *
1449 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1450               gfp_t gfp_mask)
1451 {
1452         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1453         struct cfq_queue *cfqq, *new_cfqq = NULL;
1454         unsigned short ioprio;
1455
1456 retry:
1457         ioprio = tsk->ioprio;
1458         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1459
1460         if (!cfqq) {
1461                 if (new_cfqq) {
1462                         cfqq = new_cfqq;
1463                         new_cfqq = NULL;
1464                 } else if (gfp_mask & __GFP_WAIT) {
1465                         spin_unlock_irq(cfqd->queue->queue_lock);
1466                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1467                         spin_lock_irq(cfqd->queue->queue_lock);
1468                         goto retry;
1469                 } else {
1470                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1471                         if (!cfqq)
1472                                 goto out;
1473                 }
1474
1475                 memset(cfqq, 0, sizeof(*cfqq));
1476
1477                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1478                 INIT_LIST_HEAD(&cfqq->cfq_list);
1479                 RB_CLEAR_ROOT(&cfqq->sort_list);
1480                 INIT_LIST_HEAD(&cfqq->fifo);
1481
1482                 cfqq->key = key;
1483                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1484                 atomic_set(&cfqq->ref, 0);
1485                 cfqq->cfqd = cfqd;
1486                 cfqq->service_last = 0;
1487                 /*
1488                  * set ->slice_left to allow preemption for a new process
1489                  */
1490                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1491                 cfq_mark_cfqq_idle_window(cfqq);
1492                 cfq_mark_cfqq_prio_changed(cfqq);
1493                 cfq_init_prio_data(cfqq);
1494         }
1495
1496         if (new_cfqq)
1497                 kmem_cache_free(cfq_pool, new_cfqq);
1498
1499         atomic_inc(&cfqq->ref);
1500 out:
1501         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1502         return cfqq;
1503 }
1504
1505 static void
1506 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1507 {
1508         spin_lock(&cfq_exit_lock);
1509         rb_erase(&cic->rb_node, &ioc->cic_root);
1510         list_del_init(&cic->queue_list);
1511         spin_unlock(&cfq_exit_lock);
1512         kmem_cache_free(cfq_ioc_pool, cic);
1513         atomic_dec(&ioc_count);
1514 }
1515
1516 static struct cfq_io_context *
1517 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1518 {
1519         struct rb_node *n;
1520         struct cfq_io_context *cic;
1521         void *k, *key = cfqd;
1522
1523 restart:
1524         n = ioc->cic_root.rb_node;
1525         while (n) {
1526                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1527                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1528                 k = cic->key;
1529                 if (unlikely(!k)) {
1530                         cfq_drop_dead_cic(ioc, cic);
1531                         goto restart;
1532                 }
1533
1534                 if (key < k)
1535                         n = n->rb_left;
1536                 else if (key > k)
1537                         n = n->rb_right;
1538                 else
1539                         return cic;
1540         }
1541
1542         return NULL;
1543 }
1544
1545 static inline void
1546 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1547              struct cfq_io_context *cic)
1548 {
1549         struct rb_node **p;
1550         struct rb_node *parent;
1551         struct cfq_io_context *__cic;
1552         void *k;
1553
1554         cic->ioc = ioc;
1555         cic->key = cfqd;
1556
1557         ioc->set_ioprio = cfq_ioc_set_ioprio;
1558 restart:
1559         parent = NULL;
1560         p = &ioc->cic_root.rb_node;
1561         while (*p) {
1562                 parent = *p;
1563                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1564                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1565                 k = __cic->key;
1566                 if (unlikely(!k)) {
1567                         cfq_drop_dead_cic(ioc, cic);
1568                         goto restart;
1569                 }
1570
1571                 if (cic->key < k)
1572                         p = &(*p)->rb_left;
1573                 else if (cic->key > k)
1574                         p = &(*p)->rb_right;
1575                 else
1576                         BUG();
1577         }
1578
1579         spin_lock(&cfq_exit_lock);
1580         rb_link_node(&cic->rb_node, parent, p);
1581         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1582         list_add(&cic->queue_list, &cfqd->cic_list);
1583         spin_unlock(&cfq_exit_lock);
1584 }
1585
1586 /*
1587  * Setup general io context and cfq io context. There can be several cfq
1588  * io contexts per general io context, if this process is doing io to more
1589  * than one device managed by cfq.
1590  */
1591 static struct cfq_io_context *
1592 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1593 {
1594         struct io_context *ioc = NULL;
1595         struct cfq_io_context *cic;
1596
1597         might_sleep_if(gfp_mask & __GFP_WAIT);
1598
1599         ioc = get_io_context(gfp_mask);
1600         if (!ioc)
1601                 return NULL;
1602
1603         cic = cfq_cic_rb_lookup(cfqd, ioc);
1604         if (cic)
1605                 goto out;
1606
1607         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1608         if (cic == NULL)
1609                 goto err;
1610
1611         cfq_cic_link(cfqd, ioc, cic);
1612 out:
1613         return cic;
1614 err:
1615         put_io_context(ioc);
1616         return NULL;
1617 }
1618
1619 static void
1620 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1621 {
1622         unsigned long elapsed, ttime;
1623
1624         /*
1625          * if this context already has stuff queued, thinktime is from
1626          * last queue not last end
1627          */
1628 #if 0
1629         if (time_after(cic->last_end_request, cic->last_queue))
1630                 elapsed = jiffies - cic->last_end_request;
1631         else
1632                 elapsed = jiffies - cic->last_queue;
1633 #else
1634                 elapsed = jiffies - cic->last_end_request;
1635 #endif
1636
1637         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1638
1639         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1640         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1641         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1642 }
1643
1644 static void
1645 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1646                        struct cfq_rq *crq)
1647 {
1648         sector_t sdist;
1649         u64 total;
1650
1651         if (cic->last_request_pos < crq->request->sector)
1652                 sdist = crq->request->sector - cic->last_request_pos;
1653         else
1654                 sdist = cic->last_request_pos - crq->request->sector;
1655
1656         /*
1657          * Don't allow the seek distance to get too large from the
1658          * odd fragment, pagein, etc
1659          */
1660         if (cic->seek_samples <= 60) /* second&third seek */
1661                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1662         else
1663                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1664
1665         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1666         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1667         total = cic->seek_total + (cic->seek_samples/2);
1668         do_div(total, cic->seek_samples);
1669         cic->seek_mean = (sector_t)total;
1670 }
1671
1672 /*
1673  * Disable idle window if the process thinks too long or seeks so much that
1674  * it doesn't matter
1675  */
1676 static void
1677 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1678                        struct cfq_io_context *cic)
1679 {
1680         int enable_idle = cfq_cfqq_idle_window(cfqq);
1681
1682         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1683             (cfqd->hw_tag && CIC_SEEKY(cic)))
1684                 enable_idle = 0;
1685         else if (sample_valid(cic->ttime_samples)) {
1686                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1687                         enable_idle = 0;
1688                 else
1689                         enable_idle = 1;
1690         }
1691
1692         if (enable_idle)
1693                 cfq_mark_cfqq_idle_window(cfqq);
1694         else
1695                 cfq_clear_cfqq_idle_window(cfqq);
1696 }
1697
1698
1699 /*
1700  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1701  * no or if we aren't sure, a 1 will cause a preempt.
1702  */
1703 static int
1704 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1705                    struct cfq_rq *crq)
1706 {
1707         struct cfq_queue *cfqq = cfqd->active_queue;
1708
1709         if (cfq_class_idle(new_cfqq))
1710                 return 0;
1711
1712         if (!cfqq)
1713                 return 0;
1714
1715         if (cfq_class_idle(cfqq))
1716                 return 1;
1717         if (!cfq_cfqq_wait_request(new_cfqq))
1718                 return 0;
1719         /*
1720          * if it doesn't have slice left, forget it
1721          */
1722         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1723                 return 0;
1724         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1725                 return 1;
1726
1727         return 0;
1728 }
1729
1730 /*
1731  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1732  * let it have half of its nominal slice.
1733  */
1734 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1735 {
1736         struct cfq_queue *__cfqq, *next;
1737
1738         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1739                 cfq_resort_rr_list(__cfqq, 1);
1740
1741         if (!cfqq->slice_left)
1742                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1743
1744         cfqq->slice_end = cfqq->slice_left + jiffies;
1745         cfq_slice_expired(cfqd, 1);
1746         __cfq_set_active_queue(cfqd, cfqq);
1747 }
1748
1749 /*
1750  * should really be a ll_rw_blk.c helper
1751  */
1752 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1753 {
1754         request_queue_t *q = cfqd->queue;
1755
1756         if (!blk_queue_plugged(q))
1757                 q->request_fn(q);
1758         else
1759                 __generic_unplug_device(q);
1760 }
1761
1762 /*
1763  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1764  * something we should do about it
1765  */
1766 static void
1767 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1768                  struct cfq_rq *crq)
1769 {
1770         struct cfq_io_context *cic;
1771
1772         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1773
1774         cic = crq->io_context;
1775
1776         /*
1777          * we never wait for an async request and we don't allow preemption
1778          * of an async request. so just return early
1779          */
1780         if (!cfq_crq_is_sync(crq)) {
1781                 /*
1782                  * sync process issued an async request, if it's waiting
1783                  * then expire it and kick rq handling.
1784                  */
1785                 if (cic == cfqd->active_cic &&
1786                     del_timer(&cfqd->idle_slice_timer)) {
1787                         cfq_slice_expired(cfqd, 0);
1788                         cfq_start_queueing(cfqd, cfqq);
1789                 }
1790                 return;
1791         }
1792
1793         cfq_update_io_thinktime(cfqd, cic);
1794         cfq_update_io_seektime(cfqd, cic, crq);
1795         cfq_update_idle_window(cfqd, cfqq, cic);
1796
1797         cic->last_queue = jiffies;
1798         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1799
1800         if (cfqq == cfqd->active_queue) {
1801                 /*
1802                  * if we are waiting for a request for this queue, let it rip
1803                  * immediately and flag that we must not expire this queue
1804                  * just now
1805                  */
1806                 if (cfq_cfqq_wait_request(cfqq)) {
1807                         cfq_mark_cfqq_must_dispatch(cfqq);
1808                         del_timer(&cfqd->idle_slice_timer);
1809                         cfq_start_queueing(cfqd, cfqq);
1810                 }
1811         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1812                 /*
1813                  * not the active queue - expire current slice if it is
1814                  * idle and has expired it's mean thinktime or this new queue
1815                  * has some old slice time left and is of higher priority
1816                  */
1817                 cfq_preempt_queue(cfqd, cfqq);
1818                 cfq_mark_cfqq_must_dispatch(cfqq);
1819                 cfq_start_queueing(cfqd, cfqq);
1820         }
1821 }
1822
1823 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1824 {
1825         struct cfq_data *cfqd = q->elevator->elevator_data;
1826         struct cfq_rq *crq = RQ_DATA(rq);
1827         struct cfq_queue *cfqq = crq->cfq_queue;
1828
1829         cfq_init_prio_data(cfqq);
1830
1831         cfq_add_crq_rb(crq);
1832
1833         list_add_tail(&rq->queuelist, &cfqq->fifo);
1834
1835         if (rq_mergeable(rq))
1836                 cfq_add_crq_hash(cfqd, crq);
1837
1838         cfq_crq_enqueued(cfqd, cfqq, crq);
1839 }
1840
1841 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1842 {
1843         struct cfq_rq *crq = RQ_DATA(rq);
1844         struct cfq_queue *cfqq = crq->cfq_queue;
1845         struct cfq_data *cfqd = cfqq->cfqd;
1846         const int sync = cfq_crq_is_sync(crq);
1847         unsigned long now;
1848
1849         now = jiffies;
1850
1851         WARN_ON(!cfqd->rq_in_driver);
1852         WARN_ON(!cfqq->on_dispatch[sync]);
1853         cfqd->rq_in_driver--;
1854         cfqq->on_dispatch[sync]--;
1855
1856         if (!cfq_class_idle(cfqq))
1857                 cfqd->last_end_request = now;
1858
1859         if (!cfq_cfqq_dispatched(cfqq)) {
1860                 if (cfq_cfqq_on_rr(cfqq)) {
1861                         cfqq->service_last = now;
1862                         cfq_resort_rr_list(cfqq, 0);
1863                 }
1864         }
1865
1866         if (sync)
1867                 crq->io_context->last_end_request = now;
1868
1869         /*
1870          * If this is the active queue, check if it needs to be expired,
1871          * or if we want to idle in case it has no pending requests.
1872          */
1873         if (cfqd->active_queue == cfqq) {
1874                 if (time_after(now, cfqq->slice_end))
1875                         cfq_slice_expired(cfqd, 0);
1876                 else if (sync && RB_EMPTY(&cfqq->sort_list)) {
1877                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1878                                 cfq_schedule_dispatch(cfqd);
1879                 }
1880         }
1881 }
1882
1883 static struct request *
1884 cfq_former_request(request_queue_t *q, struct request *rq)
1885 {
1886         struct cfq_rq *crq = RQ_DATA(rq);
1887         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1888
1889         if (rbprev)
1890                 return rb_entry_crq(rbprev)->request;
1891
1892         return NULL;
1893 }
1894
1895 static struct request *
1896 cfq_latter_request(request_queue_t *q, struct request *rq)
1897 {
1898         struct cfq_rq *crq = RQ_DATA(rq);
1899         struct rb_node *rbnext = rb_next(&crq->rb_node);
1900
1901         if (rbnext)
1902                 return rb_entry_crq(rbnext)->request;
1903
1904         return NULL;
1905 }
1906
1907 /*
1908  * we temporarily boost lower priority queues if they are holding fs exclusive
1909  * resources. they are boosted to normal prio (CLASS_BE/4)
1910  */
1911 static void cfq_prio_boost(struct cfq_queue *cfqq)
1912 {
1913         const int ioprio_class = cfqq->ioprio_class;
1914         const int ioprio = cfqq->ioprio;
1915
1916         if (has_fs_excl()) {
1917                 /*
1918                  * boost idle prio on transactions that would lock out other
1919                  * users of the filesystem
1920                  */
1921                 if (cfq_class_idle(cfqq))
1922                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1923                 if (cfqq->ioprio > IOPRIO_NORM)
1924                         cfqq->ioprio = IOPRIO_NORM;
1925         } else {
1926                 /*
1927                  * check if we need to unboost the queue
1928                  */
1929                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1930                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1931                 if (cfqq->ioprio != cfqq->org_ioprio)
1932                         cfqq->ioprio = cfqq->org_ioprio;
1933         }
1934
1935         /*
1936          * refile between round-robin lists if we moved the priority class
1937          */
1938         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1939             cfq_cfqq_on_rr(cfqq))
1940                 cfq_resort_rr_list(cfqq, 0);
1941 }
1942
1943 static inline int
1944 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1945                 struct task_struct *task, int rw)
1946 {
1947         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1948             !cfq_cfqq_must_alloc_slice(cfqq)) {
1949                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1950                 return ELV_MQUEUE_MUST;
1951         }
1952
1953         return ELV_MQUEUE_MAY;
1954 }
1955
1956 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1957 {
1958         struct cfq_data *cfqd = q->elevator->elevator_data;
1959         struct task_struct *tsk = current;
1960         struct cfq_queue *cfqq;
1961
1962         /*
1963          * don't force setup of a queue from here, as a call to may_queue
1964          * does not necessarily imply that a request actually will be queued.
1965          * so just lookup a possibly existing queue, or return 'may queue'
1966          * if that fails
1967          */
1968         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1969         if (cfqq) {
1970                 cfq_init_prio_data(cfqq);
1971                 cfq_prio_boost(cfqq);
1972
1973                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1974         }
1975
1976         return ELV_MQUEUE_MAY;
1977 }
1978
1979 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1980 {
1981         struct cfq_data *cfqd = q->elevator->elevator_data;
1982
1983         if (unlikely(cfqd->rq_starved)) {
1984                 struct request_list *rl = &q->rq;
1985
1986                 smp_mb();
1987                 if (waitqueue_active(&rl->wait[READ]))
1988                         wake_up(&rl->wait[READ]);
1989                 if (waitqueue_active(&rl->wait[WRITE]))
1990                         wake_up(&rl->wait[WRITE]);
1991         }
1992 }
1993
1994 /*
1995  * queue lock held here
1996  */
1997 static void cfq_put_request(request_queue_t *q, struct request *rq)
1998 {
1999         struct cfq_data *cfqd = q->elevator->elevator_data;
2000         struct cfq_rq *crq = RQ_DATA(rq);
2001
2002         if (crq) {
2003                 struct cfq_queue *cfqq = crq->cfq_queue;
2004                 const int rw = rq_data_dir(rq);
2005
2006                 BUG_ON(!cfqq->allocated[rw]);
2007                 cfqq->allocated[rw]--;
2008
2009                 put_io_context(crq->io_context->ioc);
2010
2011                 mempool_free(crq, cfqd->crq_pool);
2012                 rq->elevator_private = NULL;
2013
2014                 cfq_check_waiters(q, cfqq);
2015                 cfq_put_queue(cfqq);
2016         }
2017 }
2018
2019 /*
2020  * Allocate cfq data structures associated with this request.
2021  */
2022 static int
2023 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2024                 gfp_t gfp_mask)
2025 {
2026         struct cfq_data *cfqd = q->elevator->elevator_data;
2027         struct task_struct *tsk = current;
2028         struct cfq_io_context *cic;
2029         const int rw = rq_data_dir(rq);
2030         pid_t key = cfq_queue_pid(tsk, rw);
2031         struct cfq_queue *cfqq;
2032         struct cfq_rq *crq;
2033         unsigned long flags;
2034         int is_sync = key != CFQ_KEY_ASYNC;
2035
2036         might_sleep_if(gfp_mask & __GFP_WAIT);
2037
2038         cic = cfq_get_io_context(cfqd, gfp_mask);
2039
2040         spin_lock_irqsave(q->queue_lock, flags);
2041
2042         if (!cic)
2043                 goto queue_fail;
2044
2045         if (!cic->cfqq[is_sync]) {
2046                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2047                 if (!cfqq)
2048                         goto queue_fail;
2049
2050                 cic->cfqq[is_sync] = cfqq;
2051         } else
2052                 cfqq = cic->cfqq[is_sync];
2053
2054         cfqq->allocated[rw]++;
2055         cfq_clear_cfqq_must_alloc(cfqq);
2056         cfqd->rq_starved = 0;
2057         atomic_inc(&cfqq->ref);
2058         spin_unlock_irqrestore(q->queue_lock, flags);
2059
2060         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2061         if (crq) {
2062                 RB_CLEAR(&crq->rb_node);
2063                 crq->rb_key = 0;
2064                 crq->request = rq;
2065                 INIT_HLIST_NODE(&crq->hash);
2066                 crq->cfq_queue = cfqq;
2067                 crq->io_context = cic;
2068
2069                 if (is_sync)
2070                         cfq_mark_crq_is_sync(crq);
2071                 else
2072                         cfq_clear_crq_is_sync(crq);
2073
2074                 rq->elevator_private = crq;
2075                 return 0;
2076         }
2077
2078         spin_lock_irqsave(q->queue_lock, flags);
2079         cfqq->allocated[rw]--;
2080         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2081                 cfq_mark_cfqq_must_alloc(cfqq);
2082         cfq_put_queue(cfqq);
2083 queue_fail:
2084         if (cic)
2085                 put_io_context(cic->ioc);
2086         /*
2087          * mark us rq allocation starved. we need to kickstart the process
2088          * ourselves if there are no pending requests that can do it for us.
2089          * that would be an extremely rare OOM situation
2090          */
2091         cfqd->rq_starved = 1;
2092         cfq_schedule_dispatch(cfqd);
2093         spin_unlock_irqrestore(q->queue_lock, flags);
2094         return 1;
2095 }
2096
2097 static void cfq_kick_queue(void *data)
2098 {
2099         request_queue_t *q = data;
2100         struct cfq_data *cfqd = q->elevator->elevator_data;
2101         unsigned long flags;
2102
2103         spin_lock_irqsave(q->queue_lock, flags);
2104
2105         if (cfqd->rq_starved) {
2106                 struct request_list *rl = &q->rq;
2107
2108                 /*
2109                  * we aren't guaranteed to get a request after this, but we
2110                  * have to be opportunistic
2111                  */
2112                 smp_mb();
2113                 if (waitqueue_active(&rl->wait[READ]))
2114                         wake_up(&rl->wait[READ]);
2115                 if (waitqueue_active(&rl->wait[WRITE]))
2116                         wake_up(&rl->wait[WRITE]);
2117         }
2118
2119         blk_remove_plug(q);
2120         q->request_fn(q);
2121         spin_unlock_irqrestore(q->queue_lock, flags);
2122 }
2123
2124 /*
2125  * Timer running if the active_queue is currently idling inside its time slice
2126  */
2127 static void cfq_idle_slice_timer(unsigned long data)
2128 {
2129         struct cfq_data *cfqd = (struct cfq_data *) data;
2130         struct cfq_queue *cfqq;
2131         unsigned long flags;
2132
2133         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2134
2135         if ((cfqq = cfqd->active_queue) != NULL) {
2136                 unsigned long now = jiffies;
2137
2138                 /*
2139                  * expired
2140                  */
2141                 if (time_after(now, cfqq->slice_end))
2142                         goto expire;
2143
2144                 /*
2145                  * only expire and reinvoke request handler, if there are
2146                  * other queues with pending requests
2147                  */
2148                 if (!cfqd->busy_queues)
2149                         goto out_cont;
2150
2151                 /*
2152                  * not expired and it has a request pending, let it dispatch
2153                  */
2154                 if (!RB_EMPTY(&cfqq->sort_list)) {
2155                         cfq_mark_cfqq_must_dispatch(cfqq);
2156                         goto out_kick;
2157                 }
2158         }
2159 expire:
2160         cfq_slice_expired(cfqd, 0);
2161 out_kick:
2162         cfq_schedule_dispatch(cfqd);
2163 out_cont:
2164         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2165 }
2166
2167 /*
2168  * Timer running if an idle class queue is waiting for service
2169  */
2170 static void cfq_idle_class_timer(unsigned long data)
2171 {
2172         struct cfq_data *cfqd = (struct cfq_data *) data;
2173         unsigned long flags, end;
2174
2175         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2176
2177         /*
2178          * race with a non-idle queue, reset timer
2179          */
2180         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2181         if (!time_after_eq(jiffies, end))
2182                 mod_timer(&cfqd->idle_class_timer, end);
2183         else
2184                 cfq_schedule_dispatch(cfqd);
2185
2186         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2187 }
2188
2189 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2190 {
2191         del_timer_sync(&cfqd->idle_slice_timer);
2192         del_timer_sync(&cfqd->idle_class_timer);
2193         blk_sync_queue(cfqd->queue);
2194 }
2195
2196 static void cfq_exit_queue(elevator_t *e)
2197 {
2198         struct cfq_data *cfqd = e->elevator_data;
2199         request_queue_t *q = cfqd->queue;
2200
2201         cfq_shutdown_timer_wq(cfqd);
2202
2203         spin_lock(&cfq_exit_lock);
2204         spin_lock_irq(q->queue_lock);
2205
2206         if (cfqd->active_queue)
2207                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2208
2209         while (!list_empty(&cfqd->cic_list)) {
2210                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2211                                                         struct cfq_io_context,
2212                                                         queue_list);
2213                 if (cic->cfqq[ASYNC]) {
2214                         cfq_put_queue(cic->cfqq[ASYNC]);
2215                         cic->cfqq[ASYNC] = NULL;
2216                 }
2217                 if (cic->cfqq[SYNC]) {
2218                         cfq_put_queue(cic->cfqq[SYNC]);
2219                         cic->cfqq[SYNC] = NULL;
2220                 }
2221                 cic->key = NULL;
2222                 list_del_init(&cic->queue_list);
2223         }
2224
2225         spin_unlock_irq(q->queue_lock);
2226         spin_unlock(&cfq_exit_lock);
2227
2228         cfq_shutdown_timer_wq(cfqd);
2229
2230         mempool_destroy(cfqd->crq_pool);
2231         kfree(cfqd->crq_hash);
2232         kfree(cfqd->cfq_hash);
2233         kfree(cfqd);
2234 }
2235
2236 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
2237 {
2238         struct cfq_data *cfqd;
2239         int i;
2240
2241         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2242         if (!cfqd)
2243                 return NULL;
2244
2245         memset(cfqd, 0, sizeof(*cfqd));
2246
2247         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2248                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2249
2250         INIT_LIST_HEAD(&cfqd->busy_rr);
2251         INIT_LIST_HEAD(&cfqd->cur_rr);
2252         INIT_LIST_HEAD(&cfqd->idle_rr);
2253         INIT_LIST_HEAD(&cfqd->empty_list);
2254         INIT_LIST_HEAD(&cfqd->cic_list);
2255
2256         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2257         if (!cfqd->crq_hash)
2258                 goto out_crqhash;
2259
2260         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2261         if (!cfqd->cfq_hash)
2262                 goto out_cfqhash;
2263
2264         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2265         if (!cfqd->crq_pool)
2266                 goto out_crqpool;
2267
2268         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2269                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2270         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2271                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2272
2273         cfqd->queue = q;
2274
2275         init_timer(&cfqd->idle_slice_timer);
2276         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2277         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2278
2279         init_timer(&cfqd->idle_class_timer);
2280         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2281         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2282
2283         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2284
2285         cfqd->cfq_queued = cfq_queued;
2286         cfqd->cfq_quantum = cfq_quantum;
2287         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2288         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2289         cfqd->cfq_back_max = cfq_back_max;
2290         cfqd->cfq_back_penalty = cfq_back_penalty;
2291         cfqd->cfq_slice[0] = cfq_slice_async;
2292         cfqd->cfq_slice[1] = cfq_slice_sync;
2293         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2294         cfqd->cfq_slice_idle = cfq_slice_idle;
2295
2296         return cfqd;
2297 out_crqpool:
2298         kfree(cfqd->cfq_hash);
2299 out_cfqhash:
2300         kfree(cfqd->crq_hash);
2301 out_crqhash:
2302         kfree(cfqd);
2303         return NULL;
2304 }
2305
2306 static void cfq_slab_kill(void)
2307 {
2308         if (crq_pool)
2309                 kmem_cache_destroy(crq_pool);
2310         if (cfq_pool)
2311                 kmem_cache_destroy(cfq_pool);
2312         if (cfq_ioc_pool)
2313                 kmem_cache_destroy(cfq_ioc_pool);
2314 }
2315
2316 static int __init cfq_slab_setup(void)
2317 {
2318         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2319                                         NULL, NULL);
2320         if (!crq_pool)
2321                 goto fail;
2322
2323         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2324                                         NULL, NULL);
2325         if (!cfq_pool)
2326                 goto fail;
2327
2328         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2329                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2330         if (!cfq_ioc_pool)
2331                 goto fail;
2332
2333         return 0;
2334 fail:
2335         cfq_slab_kill();
2336         return -ENOMEM;
2337 }
2338
2339 /*
2340  * sysfs parts below -->
2341  */
2342
2343 static ssize_t
2344 cfq_var_show(unsigned int var, char *page)
2345 {
2346         return sprintf(page, "%d\n", var);
2347 }
2348
2349 static ssize_t
2350 cfq_var_store(unsigned int *var, const char *page, size_t count)
2351 {
2352         char *p = (char *) page;
2353
2354         *var = simple_strtoul(p, &p, 10);
2355         return count;
2356 }
2357
2358 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2359 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2360 {                                                                       \
2361         struct cfq_data *cfqd = e->elevator_data;                       \
2362         unsigned int __data = __VAR;                                    \
2363         if (__CONV)                                                     \
2364                 __data = jiffies_to_msecs(__data);                      \
2365         return cfq_var_show(__data, (page));                            \
2366 }
2367 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2368 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2369 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2370 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2371 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2372 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2373 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2374 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2375 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2376 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2377 #undef SHOW_FUNCTION
2378
2379 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2380 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2381 {                                                                       \
2382         struct cfq_data *cfqd = e->elevator_data;                       \
2383         unsigned int __data;                                            \
2384         int ret = cfq_var_store(&__data, (page), count);                \
2385         if (__data < (MIN))                                             \
2386                 __data = (MIN);                                         \
2387         else if (__data > (MAX))                                        \
2388                 __data = (MAX);                                         \
2389         if (__CONV)                                                     \
2390                 *(__PTR) = msecs_to_jiffies(__data);                    \
2391         else                                                            \
2392                 *(__PTR) = __data;                                      \
2393         return ret;                                                     \
2394 }
2395 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2396 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2397 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2398 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2399 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2400 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2401 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2402 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2403 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2404 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2405 #undef STORE_FUNCTION
2406
2407 #define CFQ_ATTR(name) \
2408         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2409
2410 static struct elv_fs_entry cfq_attrs[] = {
2411         CFQ_ATTR(quantum),
2412         CFQ_ATTR(queued),
2413         CFQ_ATTR(fifo_expire_sync),
2414         CFQ_ATTR(fifo_expire_async),
2415         CFQ_ATTR(back_seek_max),
2416         CFQ_ATTR(back_seek_penalty),
2417         CFQ_ATTR(slice_sync),
2418         CFQ_ATTR(slice_async),
2419         CFQ_ATTR(slice_async_rq),
2420         CFQ_ATTR(slice_idle),
2421         __ATTR_NULL
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_dispatch_fn =         cfq_dispatch_requests,
2430                 .elevator_add_req_fn =          cfq_insert_request,
2431                 .elevator_activate_req_fn =     cfq_activate_request,
2432                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2433                 .elevator_queue_empty_fn =      cfq_queue_empty,
2434                 .elevator_completed_req_fn =    cfq_completed_request,
2435                 .elevator_former_req_fn =       cfq_former_request,
2436                 .elevator_latter_req_fn =       cfq_latter_request,
2437                 .elevator_set_req_fn =          cfq_set_request,
2438                 .elevator_put_req_fn =          cfq_put_request,
2439                 .elevator_may_queue_fn =        cfq_may_queue,
2440                 .elevator_init_fn =             cfq_init_queue,
2441                 .elevator_exit_fn =             cfq_exit_queue,
2442                 .trim =                         cfq_trim,
2443         },
2444         .elevator_attrs =       cfq_attrs,
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         DECLARE_COMPLETION(all_gone);
2474         elv_unregister(&iosched_cfq);
2475         ioc_gone = &all_gone;
2476         /* ioc_gone's update must be visible before reading ioc_count */
2477         smp_wmb();
2478         if (atomic_read(&ioc_count))
2479                 wait_for_completion(ioc_gone);
2480         synchronize_rcu();
2481         cfq_slab_kill();
2482 }
2483
2484 module_init(cfq_init);
2485 module_exit(cfq_exit);
2486
2487 MODULE_AUTHOR("Jens Axboe");
2488 MODULE_LICENSE("GPL");
2489 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");