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 / 70;
  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         unsigned int max_queued;
 127
 128         mempool_t *crq_pool;
 129
 130         int rq_in_driver;
 131         int hw_tag;
 132
 133         /*
 134          * schedule slice state info
 135          */
 136         /*
 137          * idle window management
 138          */
 139         struct timer_list idle_slice_timer;
 140         struct work_struct unplug_work;
 141
 142         struct cfq_queue *active_queue;
 143         struct cfq_io_context *active_cic;
 144         int cur_prio, cur_end_prio;
 145         unsigned int dispatch_slice;
 146
 147         struct timer_list idle_class_timer;
 148
 149         sector_t last_sector;
 150         unsigned long last_end_request;
 151
 152         unsigned int rq_starved;
 153
 154         /*
 155          * tunables, see top of file
 156          */
 157         unsigned int cfq_quantum;
 158         unsigned int cfq_queued;
 159         unsigned int cfq_fifo_expire[2];
 160         unsigned int cfq_back_penalty;
 161         unsigned int cfq_back_max;
 162         unsigned int cfq_slice[2];
 163         unsigned int cfq_slice_async_rq;
 164         unsigned int cfq_slice_idle;
 165
 166         struct list_head cic_list;
 167 };
 168
 169 /*
 170  * Per process-grouping structure
 171  */
 172 struct cfq_queue {
 173         /* reference count */
 174         atomic_t ref;
 175         /* parent cfq_data */
 176         struct cfq_data *cfqd;
 177         /* cfqq lookup hash */
 178         struct hlist_node cfq_hash;
 179         /* hash key */
 180         unsigned int key;
 181         /* on either rr or empty list of cfqd */
 182         struct list_head cfq_list;
 183         /* sorted list of pending requests */
 184         struct rb_root sort_list;
 185         /* if fifo isn't expired, next request to serve */
 186         struct cfq_rq *next_crq;
 187         /* requests queued in sort_list */
 188         int queued[2];
 189         /* currently allocated requests */
 190         int allocated[2];
 191         /* fifo list of requests in sort_list */
 192         struct list_head fifo;
 193
 194         unsigned long slice_start;
 195         unsigned long slice_end;
 196         unsigned long slice_left;
 197         unsigned long service_last;
 198
 199         /* number of requests that are on the dispatch list */
 200         int on_dispatch[2];
 201
 202         /* io prio of this group */
 203         unsigned short ioprio, org_ioprio;
 204         unsigned short ioprio_class, org_ioprio_class;
 205
 206         /* various state flags, see below */
 207         unsigned int flags;
 208 };
 209
 210 struct cfq_rq {
 211         struct rb_node rb_node;
 212         sector_t rb_key;
 213         struct request *request;
 214         struct hlist_node hash;
 215
 216         struct cfq_queue *cfq_queue;
 217         struct cfq_io_context *io_context;
 218
 219         unsigned int crq_flags;
 220 };
 221
 222 enum cfqq_state_flags {
 223         CFQ_CFQQ_FLAG_on_rr = 0,
 224         CFQ_CFQQ_FLAG_wait_request,
 225         CFQ_CFQQ_FLAG_must_alloc,
 226         CFQ_CFQQ_FLAG_must_alloc_slice,
 227         CFQ_CFQQ_FLAG_must_dispatch,
 228         CFQ_CFQQ_FLAG_fifo_expire,
 229         CFQ_CFQQ_FLAG_idle_window,
 230         CFQ_CFQQ_FLAG_prio_changed,
 231 };
 232
 233 #define CFQ_CFQQ_FNS(name)                                              \
 234 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 235 {                                                                       \
 236         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 237 }                                                                       \
 238 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 239 {                                                                       \
 240         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 241 }                                                                       \
 242 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 243 {                                                                       \
 244         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 245 }
 246
 247 CFQ_CFQQ_FNS(on_rr);
 248 CFQ_CFQQ_FNS(wait_request);
 249 CFQ_CFQQ_FNS(must_alloc);
 250 CFQ_CFQQ_FNS(must_alloc_slice);
 251 CFQ_CFQQ_FNS(must_dispatch);
 252 CFQ_CFQQ_FNS(fifo_expire);
 253 CFQ_CFQQ_FNS(idle_window);
 254 CFQ_CFQQ_FNS(prio_changed);
 255 #undef CFQ_CFQQ_FNS
 256
 257 enum cfq_rq_state_flags {
 258         CFQ_CRQ_FLAG_is_sync = 0,
 259 };
 260
 261 #define CFQ_CRQ_FNS(name)                                               \
 262 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 263 {                                                                       \
 264         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 265 }                                                                       \
 266 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 267 {                                                                       \
 268         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 269 }                                                                       \
 270 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 271 {                                                                       \
 272         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 273 }
 274
 275 CFQ_CRQ_FNS(is_sync);
 276 #undef CFQ_CRQ_FNS
 277
 278 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 279 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 280 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 281
 282 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 283
 284 /*
 285  * lots of deadline iosched dupes, can be abstracted later...
 286  */
 287 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 288 {
 289         hlist_del_init(&crq->hash);
 290 }
 291
 292 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 293 {
 294         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 295
 296         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 297 }
 298
 299 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 300 {
 301         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 302         struct hlist_node *entry, *next;
 303
 304         hlist_for_each_safe(entry, next, hash_list) {
 305                 struct cfq_rq *crq = list_entry_hash(entry);
 306                 struct request *__rq = crq->request;
 307
 308                 if (!rq_mergeable(__rq)) {
 309                         cfq_del_crq_hash(crq);
 310                         continue;
 311                 }
 312
 313                 if (rq_hash_key(__rq) == offset)
 314                         return __rq;
 315         }
 316
 317         return NULL;
 318 }
 319
 320 /*
 321  * scheduler run of queue, if there are requests pending and no one in the
 322  * driver that will restart queueing
 323  */
 324 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 325 {
 326         if (cfqd->busy_queues)
 327                 kblockd_schedule_work(&cfqd->unplug_work);
 328 }
 329
 330 static int cfq_queue_empty(request_queue_t *q)
 331 {
 332         struct cfq_data *cfqd = q->elevator->elevator_data;
 333
 334         return !cfqd->busy_queues;
 335 }
 336
 337 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 338 {
 339         if (rw == READ || process_sync(task))
 340                 return task->pid;
 341
 342         return CFQ_KEY_ASYNC;
 343 }
 344
 345 /*
 346  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 347  * We choose the request that is closest to the head right now. Distance
 348  * behind the head is penalized and only allowed to a certain extent.
 349  */
 350 static struct cfq_rq *
 351 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 352 {
 353         sector_t last, s1, s2, d1 = 0, d2 = 0;
 354         unsigned long back_max;
 355 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 356 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 357         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 358
 359         if (crq1 == NULL || crq1 == crq2)
 360                 return crq2;
 361         if (crq2 == NULL)
 362                 return crq1;
 363
 364         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 365                 return crq1;
 366         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 367                 return crq2;
 368
 369         s1 = crq1->request->sector;
 370         s2 = crq2->request->sector;
 371
 372         last = cfqd->last_sector;
 373
 374         /*
 375          * by definition, 1KiB is 2 sectors
 376          */
 377         back_max = cfqd->cfq_back_max * 2;
 378
 379         /*
 380          * Strict one way elevator _except_ in the case where we allow
 381          * short backward seeks which are biased as twice the cost of a
 382          * similar forward seek.
 383          */
 384         if (s1 >= last)
 385                 d1 = s1 - last;
 386         else if (s1 + back_max >= last)
 387                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 388         else
 389                 wrap |= CFQ_RQ1_WRAP;
 390
 391         if (s2 >= last)
 392                 d2 = s2 - last;
 393         else if (s2 + back_max >= last)
 394                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 395         else
 396                 wrap |= CFQ_RQ2_WRAP;
 397
 398         /* Found required data */
 399
 400         /*
 401          * By doing switch() on the bit mask "wrap" we avoid having to
 402          * check two variables for all permutations: --> faster!
 403          */
 404         switch (wrap) {
 405         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 406                 if (d1 < d2)
 407                         return crq1;
 408                 else if (d2 < d1)
 409                         return crq2;
 410                 else {
 411                         if (s1 >= s2)
 412                                 return crq1;
 413                         else
 414                                 return crq2;
 415                 }
 416
 417         case CFQ_RQ2_WRAP:
 418                 return crq1;
 419         case CFQ_RQ1_WRAP:
 420                 return crq2;
 421         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 422         default:
 423                 /*
 424                  * Since both rqs are wrapped,
 425                  * start with the one that's further behind head
 426                  * (--> only *one* back seek required),
 427                  * since back seek takes more time than forward.
 428                  */
 429                 if (s1 <= s2)
 430                         return crq1;
 431                 else
 432                         return crq2;
 433         }
 434 }
 435
 436 /*
 437  * would be nice to take fifo expire time into account as well
 438  */
 439 static struct cfq_rq *
 440 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 441                   struct cfq_rq *last)
 442 {
 443         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 444         struct rb_node *rbnext, *rbprev;
 445
 446         if (!(rbnext = rb_next(&last->rb_node))) {
 447                 rbnext = rb_first(&cfqq->sort_list);
 448                 if (rbnext == &last->rb_node)
 449                         rbnext = NULL;
 450         }
 451
 452         rbprev = rb_prev(&last->rb_node);
 453
 454         if (rbprev)
 455                 crq_prev = rb_entry_crq(rbprev);
 456         if (rbnext)
 457                 crq_next = rb_entry_crq(rbnext);
 458
 459         return cfq_choose_req(cfqd, crq_next, crq_prev);
 460 }
 461
 462 static void cfq_update_next_crq(struct cfq_rq *crq)
 463 {
 464         struct cfq_queue *cfqq = crq->cfq_queue;
 465
 466         if (cfqq->next_crq == crq)
 467                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 468 }
 469
 470 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 471 {
 472         struct cfq_data *cfqd = cfqq->cfqd;
 473         struct list_head *list, *entry;
 474
 475         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 476
 477         list_del(&cfqq->cfq_list);
 478
 479         if (cfq_class_rt(cfqq))
 480                 list = &cfqd->cur_rr;
 481         else if (cfq_class_idle(cfqq))
 482                 list = &cfqd->idle_rr;
 483         else {
 484                 /*
 485                  * if cfqq has requests in flight, don't allow it to be
 486                  * found in cfq_set_active_queue before it has finished them.
 487                  * this is done to increase fairness between a process that
 488                  * has lots of io pending vs one that only generates one
 489                  * sporadically or synchronously
 490                  */
 491                 if (cfq_cfqq_dispatched(cfqq))
 492                         list = &cfqd->busy_rr;
 493                 else
 494                         list = &cfqd->rr_list[cfqq->ioprio];
 495         }
 496
 497         /*
 498          * if queue was preempted, just add to front to be fair. busy_rr
 499          * isn't sorted, but insert at the back for fairness.
 500          */
 501         if (preempted || list == &cfqd->busy_rr) {
 502                 if (preempted)
 503                         list = list->prev;
 504
 505                 list_add_tail(&cfqq->cfq_list, list);
 506                 return;
 507         }
 508
 509         /*
 510          * sort by when queue was last serviced
 511          */
 512         entry = list;
 513         while ((entry = entry->prev) != list) {
 514                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 515
 516                 if (!__cfqq->service_last)
 517                         break;
 518                 if (time_before(__cfqq->service_last, cfqq->service_last))
 519                         break;
 520         }
 521
 522         list_add(&cfqq->cfq_list, entry);
 523 }
 524
 525 /*
 526  * add to busy list of queues for service, trying to be fair in ordering
 527  * the pending list according to last request service
 528  */
 529 static inline void
 530 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 531 {
 532         BUG_ON(cfq_cfqq_on_rr(cfqq));
 533         cfq_mark_cfqq_on_rr(cfqq);
 534         cfqd->busy_queues++;
 535
 536         cfq_resort_rr_list(cfqq, 0);
 537 }
 538
 539 static inline void
 540 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 541 {
 542         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 543         cfq_clear_cfqq_on_rr(cfqq);
 544         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 545
 546         BUG_ON(!cfqd->busy_queues);
 547         cfqd->busy_queues--;
 548 }
 549
 550 /*
 551  * rb tree support functions
 552  */
 553 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 554 {
 555         struct cfq_queue *cfqq = crq->cfq_queue;
 556         struct cfq_data *cfqd = cfqq->cfqd;
 557         const int sync = cfq_crq_is_sync(crq);
 558
 559         BUG_ON(!cfqq->queued[sync]);
 560         cfqq->queued[sync]--;
 561
 562         cfq_update_next_crq(crq);
 563
 564         rb_erase(&crq->rb_node, &cfqq->sort_list);
 565
 566         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 567                 cfq_del_cfqq_rr(cfqd, cfqq);
 568 }
 569
 570 static struct cfq_rq *
 571 __cfq_add_crq_rb(struct cfq_rq *crq)
 572 {
 573         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 574         struct rb_node *parent = NULL;
 575         struct cfq_rq *__crq;
 576
 577         while (*p) {
 578                 parent = *p;
 579                 __crq = rb_entry_crq(parent);
 580
 581                 if (crq->rb_key < __crq->rb_key)
 582                         p = &(*p)->rb_left;
 583                 else if (crq->rb_key > __crq->rb_key)
 584                         p = &(*p)->rb_right;
 585                 else
 586                         return __crq;
 587         }
 588
 589         rb_link_node(&crq->rb_node, parent, p);
 590         return NULL;
 591 }
 592
 593 static void cfq_add_crq_rb(struct cfq_rq *crq)
 594 {
 595         struct cfq_queue *cfqq = crq->cfq_queue;
 596         struct cfq_data *cfqd = cfqq->cfqd;
 597         struct request *rq = crq->request;
 598         struct cfq_rq *__alias;
 599
 600         crq->rb_key = rq_rb_key(rq);
 601         cfqq->queued[cfq_crq_is_sync(crq)]++;
 602
 603         /*
 604          * looks a little odd, but the first insert might return an alias.
 605          * if that happens, put the alias on the dispatch list
 606          */
 607         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 608                 cfq_dispatch_insert(cfqd->queue, __alias);
 609
 610         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 611
 612         if (!cfq_cfqq_on_rr(cfqq))
 613                 cfq_add_cfqq_rr(cfqd, cfqq);
 614
 615         /*
 616          * check if this request is a better next-serve candidate
 617          */
 618         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 619 }
 620
 621 static inline void
 622 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 623 {
 624         rb_erase(&crq->rb_node, &cfqq->sort_list);
 625         cfqq->queued[cfq_crq_is_sync(crq)]--;
 626
 627         cfq_add_crq_rb(crq);
 628 }
 629
 630 static struct request *
 631 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 632 {
 633         struct task_struct *tsk = current;
 634         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 635         struct cfq_queue *cfqq;
 636         struct rb_node *n;
 637         sector_t sector;
 638
 639         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 640         if (!cfqq)
 641                 goto out;
 642
 643         sector = bio->bi_sector + bio_sectors(bio);
 644         n = cfqq->sort_list.rb_node;
 645         while (n) {
 646                 struct cfq_rq *crq = rb_entry_crq(n);
 647
 648                 if (sector < crq->rb_key)
 649                         n = n->rb_left;
 650                 else if (sector > crq->rb_key)
 651                         n = n->rb_right;
 652                 else
 653                         return crq->request;
 654         }
 655
 656 out:
 657         return NULL;
 658 }
 659
 660 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 661 {
 662         struct cfq_data *cfqd = q->elevator->elevator_data;
 663
 664         cfqd->rq_in_driver++;
 665
 666         /*
 667          * If the depth is larger 1, it really could be queueing. But lets
 668          * make the mark a little higher - idling could still be good for
 669          * low queueing, and a low queueing number could also just indicate
 670          * a SCSI mid layer like behaviour where limit+1 is often seen.
 671          */
 672         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
 673                 cfqd->hw_tag = 1;
 674 }
 675
 676 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 677 {
 678         struct cfq_data *cfqd = q->elevator->elevator_data;
 679
 680         WARN_ON(!cfqd->rq_in_driver);
 681         cfqd->rq_in_driver--;
 682 }
 683
 684 static void cfq_remove_request(struct request *rq)
 685 {
 686         struct cfq_rq *crq = RQ_DATA(rq);
 687
 688         list_del_init(&rq->queuelist);
 689         cfq_del_crq_rb(crq);
 690         cfq_del_crq_hash(crq);
 691 }
 692
 693 static int
 694 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 695 {
 696         struct cfq_data *cfqd = q->elevator->elevator_data;
 697         struct request *__rq;
 698         int ret;
 699
 700         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 701         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 702                 ret = ELEVATOR_BACK_MERGE;
 703                 goto out;
 704         }
 705
 706         __rq = cfq_find_rq_fmerge(cfqd, bio);
 707         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 708                 ret = ELEVATOR_FRONT_MERGE;
 709                 goto out;
 710         }
 711
 712         return ELEVATOR_NO_MERGE;
 713 out:
 714         *req = __rq;
 715         return ret;
 716 }
 717
 718 static void cfq_merged_request(request_queue_t *q, struct request *req)
 719 {
 720         struct cfq_data *cfqd = q->elevator->elevator_data;
 721         struct cfq_rq *crq = RQ_DATA(req);
 722
 723         cfq_del_crq_hash(crq);
 724         cfq_add_crq_hash(cfqd, crq);
 725
 726         if (rq_rb_key(req) != crq->rb_key) {
 727                 struct cfq_queue *cfqq = crq->cfq_queue;
 728
 729                 cfq_update_next_crq(crq);
 730                 cfq_reposition_crq_rb(cfqq, crq);
 731         }
 732 }
 733
 734 static void
 735 cfq_merged_requests(request_queue_t *q, struct request *rq,
 736                     struct request *next)
 737 {
 738         cfq_merged_request(q, rq);
 739
 740         /*
 741          * reposition in fifo if next is older than rq
 742          */
 743         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 744             time_before(next->start_time, rq->start_time))
 745                 list_move(&rq->queuelist, &next->queuelist);
 746
 747         cfq_remove_request(next);
 748 }
 749
 750 static inline void
 751 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 752 {
 753         if (cfqq) {
 754                 /*
 755                  * stop potential idle class queues waiting service
 756                  */
 757                 del_timer(&cfqd->idle_class_timer);
 758
 759                 cfqq->slice_start = jiffies;
 760                 cfqq->slice_end = 0;
 761                 cfqq->slice_left = 0;
 762                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 763                 cfq_clear_cfqq_fifo_expire(cfqq);
 764         }
 765
 766         cfqd->active_queue = cfqq;
 767 }
 768
 769 /*
 770  * current cfqq expired its slice (or was too idle), select new one
 771  */
 772 static void
 773 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 774                     int preempted)
 775 {
 776         unsigned long now = jiffies;
 777
 778         if (cfq_cfqq_wait_request(cfqq))
 779                 del_timer(&cfqd->idle_slice_timer);
 780
 781         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 782                 cfqq->service_last = now;
 783                 cfq_schedule_dispatch(cfqd);
 784         }
 785
 786         cfq_clear_cfqq_must_dispatch(cfqq);
 787         cfq_clear_cfqq_wait_request(cfqq);
 788
 789         /*
 790          * store what was left of this slice, if the queue idled out
 791          * or was preempted
 792          */
 793         if (time_after(cfqq->slice_end, now))
 794                 cfqq->slice_left = cfqq->slice_end - now;
 795         else
 796                 cfqq->slice_left = 0;
 797
 798         if (cfq_cfqq_on_rr(cfqq))
 799                 cfq_resort_rr_list(cfqq, preempted);
 800
 801         if (cfqq == cfqd->active_queue)
 802                 cfqd->active_queue = NULL;
 803
 804         if (cfqd->active_cic) {
 805                 put_io_context(cfqd->active_cic->ioc);
 806                 cfqd->active_cic = NULL;
 807         }
 808
 809         cfqd->dispatch_slice = 0;
 810 }
 811
 812 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 813 {
 814         struct cfq_queue *cfqq = cfqd->active_queue;
 815
 816         if (cfqq)
 817                 __cfq_slice_expired(cfqd, cfqq, preempted);
 818 }
 819
 820 /*
 821  * 0
 822  * 0,1
 823  * 0,1,2
 824  * 0,1,2,3
 825  * 0,1,2,3,4
 826  * 0,1,2,3,4,5
 827  * 0,1,2,3,4,5,6
 828  * 0,1,2,3,4,5,6,7
 829  */
 830 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 831 {
 832         int prio, wrap;
 833
 834         prio = -1;
 835         wrap = 0;
 836         do {
 837                 int p;
 838
 839                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 840                         if (!list_empty(&cfqd->rr_list[p])) {
 841                                 prio = p;
 842                                 break;
 843                         }
 844                 }
 845
 846                 if (prio != -1)
 847                         break;
 848                 cfqd->cur_prio = 0;
 849                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 850                         cfqd->cur_end_prio = 0;
 851                         if (wrap)
 852                                 break;
 853                         wrap = 1;
 854                 }
 855         } while (1);
 856
 857         if (unlikely(prio == -1))
 858                 return -1;
 859
 860         BUG_ON(prio >= CFQ_PRIO_LISTS);
 861
 862         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 863
 864         cfqd->cur_prio = prio + 1;
 865         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 866                 cfqd->cur_end_prio = cfqd->cur_prio;
 867                 cfqd->cur_prio = 0;
 868         }
 869         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 870                 cfqd->cur_prio = 0;
 871                 cfqd->cur_end_prio = 0;
 872         }
 873
 874         return prio;
 875 }
 876
 877 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 878 {
 879         struct cfq_queue *cfqq = NULL;
 880
 881         /*
 882          * if current list is non-empty, grab first entry. if it is empty,
 883          * get next prio level and grab first entry then if any are spliced
 884          */
 885         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 886                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 887
 888         /*
 889          * If no new queues are available, check if the busy list has some
 890          * before falling back to idle io.
 891          */
 892         if (!cfqq && !list_empty(&cfqd->busy_rr))
 893                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 894
 895         /*
 896          * if we have idle queues and no rt or be queues had pending
 897          * requests, either allow immediate service if the grace period
 898          * has passed or arm the idle grace timer
 899          */
 900         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 901                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 902
 903                 if (time_after_eq(jiffies, end))
 904                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 905                 else
 906                         mod_timer(&cfqd->idle_class_timer, end);
 907         }
 908
 909         __cfq_set_active_queue(cfqd, cfqq);
 910         return cfqq;
 911 }
 912
 913 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 914
 915 {
 916         struct cfq_io_context *cic;
 917         unsigned long sl;
 918
 919         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 920         WARN_ON(cfqq != cfqd->active_queue);
 921
 922         /*
 923          * idle is disabled, either manually or by past process history
 924          */
 925         if (!cfqd->cfq_slice_idle)
 926                 return 0;
 927         if (!cfq_cfqq_idle_window(cfqq))
 928                 return 0;
 929         /*
 930          * task has exited, don't wait
 931          */
 932         cic = cfqd->active_cic;
 933         if (!cic || !cic->ioc->task)
 934                 return 0;
 935
 936         cfq_mark_cfqq_must_dispatch(cfqq);
 937         cfq_mark_cfqq_wait_request(cfqq);
 938
 939         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 940
 941         /*
 942          * we don't want to idle for seeks, but we do want to allow
 943          * fair distribution of slice time for a process doing back-to-back
 944          * seeks. so allow a little bit of time for him to submit a new rq
 945          */
 946         if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
 947                 sl = 2;
 948
 949         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 950         return 1;
 951 }
 952
 953 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 954 {
 955         struct cfq_data *cfqd = q->elevator->elevator_data;
 956         struct cfq_queue *cfqq = crq->cfq_queue;
 957
 958         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 959         cfq_remove_request(crq->request);
 960         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 961         elv_dispatch_sort(q, crq->request);
 962 }
 963
 964 /*
 965  * return expired entry, or NULL to just start from scratch in rbtree
 966  */
 967 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 968 {
 969         struct cfq_data *cfqd = cfqq->cfqd;
 970         struct request *rq;
 971         struct cfq_rq *crq;
 972
 973         if (cfq_cfqq_fifo_expire(cfqq))
 974                 return NULL;
 975
 976         if (!list_empty(&cfqq->fifo)) {
 977                 int fifo = cfq_cfqq_class_sync(cfqq);
 978
 979                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 980                 rq = crq->request;
 981                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 982                         cfq_mark_cfqq_fifo_expire(cfqq);
 983                         return crq;
 984                 }
 985         }
 986
 987         return NULL;
 988 }
 989
 990 /*
 991  * Scale schedule slice based on io priority. Use the sync time slice only
 992  * if a queue is marked sync and has sync io queued. A sync queue with async
 993  * io only, should not get full sync slice length.
 994  */
 995 static inline int
 996 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 997 {
 998         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 999
1000         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1001
1002         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
1003 }
1004
1005 static inline void
1006 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1007 {
1008         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1009 }
1010
1011 static inline int
1012 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1013 {
1014         const int base_rq = cfqd->cfq_slice_async_rq;
1015
1016         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1017
1018         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1019 }
1020
1021 /*
1022  * get next queue for service
1023  */
1024 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1025 {
1026         unsigned long now = jiffies;
1027         struct cfq_queue *cfqq;
1028
1029         cfqq = cfqd->active_queue;
1030         if (!cfqq)
1031                 goto new_queue;
1032
1033         /*
1034          * slice has expired
1035          */
1036         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1037                 goto expire;
1038
1039         /*
1040          * if queue has requests, dispatch one. if not, check if
1041          * enough slice is left to wait for one
1042          */
1043         if (!RB_EMPTY(&cfqq->sort_list))
1044                 goto keep_queue;
1045         else if (cfq_cfqq_class_sync(cfqq) &&
1046                  time_before(now, cfqq->slice_end)) {
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. if at least one request was
1096          * sync, use the sync time slice value
1097          */
1098         if (!cfqq->slice_end)
1099                 cfq_set_prio_slice(cfqd, cfqq);
1100
1101         /*
1102          * expire an async queue immediately if it has used up its slice. idle
1103          * queue always expire after 1 dispatch round.
1104          */
1105         if ((!cfq_cfqq_sync(cfqq) &&
1106             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1107             cfq_class_idle(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         int dispatched = 0;
1117         struct cfq_queue *cfqq, *next;
1118         struct cfq_rq *crq;
1119
1120         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1121                 while ((crq = cfqq->next_crq)) {
1122                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1123                         dispatched++;
1124                 }
1125                 BUG_ON(!list_empty(&cfqq->fifo));
1126         }
1127         return dispatched;
1128 }
1129
1130 static int
1131 cfq_forced_dispatch(struct cfq_data *cfqd)
1132 {
1133         int i, dispatched = 0;
1134
1135         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1136                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1137
1138         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1139         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1140         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1141
1142         cfq_slice_expired(cfqd, 0);
1143
1144         BUG_ON(cfqd->busy_queues);
1145
1146         return dispatched;
1147 }
1148
1149 static int
1150 cfq_dispatch_requests(request_queue_t *q, int force)
1151 {
1152         struct cfq_data *cfqd = q->elevator->elevator_data;
1153         struct cfq_queue *cfqq;
1154
1155         if (!cfqd->busy_queues)
1156                 return 0;
1157
1158         if (unlikely(force))
1159                 return cfq_forced_dispatch(cfqd);
1160
1161         cfqq = cfq_select_queue(cfqd);
1162         if (cfqq) {
1163                 int max_dispatch;
1164
1165                 cfq_clear_cfqq_must_dispatch(cfqq);
1166                 cfq_clear_cfqq_wait_request(cfqq);
1167                 del_timer(&cfqd->idle_slice_timer);
1168
1169                 max_dispatch = cfqd->cfq_quantum;
1170                 if (cfq_class_idle(cfqq))
1171                         max_dispatch = 1;
1172
1173                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1174         }
1175
1176         return 0;
1177 }
1178
1179 /*
1180  * task holds one reference to the queue, dropped when task exits. each crq
1181  * in-flight on this queue also holds a reference, dropped when crq is freed.
1182  *
1183  * queue lock must be held here.
1184  */
1185 static void cfq_put_queue(struct cfq_queue *cfqq)
1186 {
1187         struct cfq_data *cfqd = cfqq->cfqd;
1188
1189         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1190
1191         if (!atomic_dec_and_test(&cfqq->ref))
1192                 return;
1193
1194         BUG_ON(rb_first(&cfqq->sort_list));
1195         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1196         BUG_ON(cfq_cfqq_on_rr(cfqq));
1197
1198         if (unlikely(cfqd->active_queue == cfqq))
1199                 __cfq_slice_expired(cfqd, cfqq, 0);
1200
1201         /*
1202          * it's on the empty list and still hashed
1203          */
1204         list_del(&cfqq->cfq_list);
1205         hlist_del(&cfqq->cfq_hash);
1206         kmem_cache_free(cfq_pool, cfqq);
1207 }
1208
1209 static inline struct cfq_queue *
1210 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1211                     const int hashval)
1212 {
1213         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1214         struct hlist_node *entry;
1215         struct cfq_queue *__cfqq;
1216
1217         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1218                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1219
1220                 if (__cfqq->key == key && (__p == prio || !prio))
1221                         return __cfqq;
1222         }
1223
1224         return NULL;
1225 }
1226
1227 static struct cfq_queue *
1228 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1229 {
1230         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1231 }
1232
1233 static void cfq_free_io_context(struct io_context *ioc)
1234 {
1235         struct cfq_io_context *__cic;
1236         struct rb_node *n;
1237         int freed = 0;
1238
1239         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1240                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1241                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1242                 kmem_cache_free(cfq_ioc_pool, __cic);
1243                 freed++;
1244         }
1245
1246         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1247                 complete(ioc_gone);
1248 }
1249
1250 static void cfq_trim(struct io_context *ioc)
1251 {
1252         ioc->set_ioprio = NULL;
1253         cfq_free_io_context(ioc);
1254 }
1255
1256 /*
1257  * Called with interrupts disabled
1258  */
1259 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1260 {
1261         struct cfq_data *cfqd = cic->key;
1262         request_queue_t *q;
1263
1264         if (!cfqd)
1265                 return;
1266
1267         q = cfqd->queue;
1268
1269         WARN_ON(!irqs_disabled());
1270
1271         spin_lock(q->queue_lock);
1272
1273         if (cic->cfqq[ASYNC]) {
1274                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1275                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1276                 cfq_put_queue(cic->cfqq[ASYNC]);
1277                 cic->cfqq[ASYNC] = NULL;
1278         }
1279
1280         if (cic->cfqq[SYNC]) {
1281                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1282                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1283                 cfq_put_queue(cic->cfqq[SYNC]);
1284                 cic->cfqq[SYNC] = NULL;
1285         }
1286
1287         cic->key = NULL;
1288         list_del_init(&cic->queue_list);
1289         spin_unlock(q->queue_lock);
1290 }
1291
1292 static void cfq_exit_io_context(struct io_context *ioc)
1293 {
1294         struct cfq_io_context *__cic;
1295         unsigned long flags;
1296         struct rb_node *n;
1297
1298         /*
1299          * put the reference this task is holding to the various queues
1300          */
1301         spin_lock_irqsave(&cfq_exit_lock, flags);
1302
1303         n = rb_first(&ioc->cic_root);
1304         while (n != NULL) {
1305                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1306
1307                 cfq_exit_single_io_context(__cic);
1308                 n = rb_next(n);
1309         }
1310
1311         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1312 }
1313
1314 static struct cfq_io_context *
1315 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1316 {
1317         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1318
1319         if (cic) {
1320                 memset(cic, 0, sizeof(*cic));
1321                 cic->last_end_request = jiffies;
1322                 INIT_LIST_HEAD(&cic->queue_list);
1323                 cic->dtor = cfq_free_io_context;
1324                 cic->exit = cfq_exit_io_context;
1325                 atomic_inc(&ioc_count);
1326         }
1327
1328         return cic;
1329 }
1330
1331 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1332 {
1333         struct task_struct *tsk = current;
1334         int ioprio_class;
1335
1336         if (!cfq_cfqq_prio_changed(cfqq))
1337                 return;
1338
1339         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1340         switch (ioprio_class) {
1341                 default:
1342                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1343                 case IOPRIO_CLASS_NONE:
1344                         /*
1345                          * no prio set, place us in the middle of the BE classes
1346                          */
1347                         cfqq->ioprio = task_nice_ioprio(tsk);
1348                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1349                         break;
1350                 case IOPRIO_CLASS_RT:
1351                         cfqq->ioprio = task_ioprio(tsk);
1352                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1353                         break;
1354                 case IOPRIO_CLASS_BE:
1355                         cfqq->ioprio = task_ioprio(tsk);
1356                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1357                         break;
1358                 case IOPRIO_CLASS_IDLE:
1359                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1360                         cfqq->ioprio = 7;
1361                         cfq_clear_cfqq_idle_window(cfqq);
1362                         break;
1363         }
1364
1365         /*
1366          * keep track of original prio settings in case we have to temporarily
1367          * elevate the priority of this queue
1368          */
1369         cfqq->org_ioprio = cfqq->ioprio;
1370         cfqq->org_ioprio_class = cfqq->ioprio_class;
1371
1372         if (cfq_cfqq_on_rr(cfqq))
1373                 cfq_resort_rr_list(cfqq, 0);
1374
1375         cfq_clear_cfqq_prio_changed(cfqq);
1376 }
1377
1378 static inline void changed_ioprio(struct cfq_io_context *cic)
1379 {
1380         struct cfq_data *cfqd = cic->key;
1381         struct cfq_queue *cfqq;
1382         if (cfqd) {
1383                 spin_lock(cfqd->queue->queue_lock);
1384                 cfqq = cic->cfqq[ASYNC];
1385                 if (cfqq) {
1386                         struct cfq_queue *new_cfqq;
1387                         new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1388                                                 cic->ioc->task, GFP_ATOMIC);
1389                         if (new_cfqq) {
1390                                 cic->cfqq[ASYNC] = new_cfqq;
1391                                 cfq_put_queue(cfqq);
1392                         }
1393                 }
1394                 cfqq = cic->cfqq[SYNC];
1395                 if (cfqq) {
1396                         cfq_mark_cfqq_prio_changed(cfqq);
1397                         cfq_init_prio_data(cfqq);
1398                 }
1399                 spin_unlock(cfqd->queue->queue_lock);
1400         }
1401 }
1402
1403 /*
1404  * callback from sys_ioprio_set, irqs are disabled
1405  */
1406 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1407 {
1408         struct cfq_io_context *cic;
1409         struct rb_node *n;
1410
1411         spin_lock(&cfq_exit_lock);
1412
1413         n = rb_first(&ioc->cic_root);
1414         while (n != NULL) {
1415                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1416
1417                 changed_ioprio(cic);
1418                 n = rb_next(n);
1419         }
1420
1421         spin_unlock(&cfq_exit_lock);
1422
1423         return 0;
1424 }
1425
1426 static struct cfq_queue *
1427 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1428               gfp_t gfp_mask)
1429 {
1430         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1431         struct cfq_queue *cfqq, *new_cfqq = NULL;
1432         unsigned short ioprio;
1433
1434 retry:
1435         ioprio = tsk->ioprio;
1436         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1437
1438         if (!cfqq) {
1439                 if (new_cfqq) {
1440                         cfqq = new_cfqq;
1441                         new_cfqq = NULL;
1442                 } else if (gfp_mask & __GFP_WAIT) {
1443                         spin_unlock_irq(cfqd->queue->queue_lock);
1444                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1445                         spin_lock_irq(cfqd->queue->queue_lock);
1446                         goto retry;
1447                 } else {
1448                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1449                         if (!cfqq)
1450                                 goto out;
1451                 }
1452
1453                 memset(cfqq, 0, sizeof(*cfqq));
1454
1455                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1456                 INIT_LIST_HEAD(&cfqq->cfq_list);
1457                 RB_CLEAR_ROOT(&cfqq->sort_list);
1458                 INIT_LIST_HEAD(&cfqq->fifo);
1459
1460                 cfqq->key = key;
1461                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1462                 atomic_set(&cfqq->ref, 0);
1463                 cfqq->cfqd = cfqd;
1464                 cfqq->service_last = 0;
1465                 /*
1466                  * set ->slice_left to allow preemption for a new process
1467                  */
1468                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1469                 if (!cfqd->hw_tag)
1470                         cfq_mark_cfqq_idle_window(cfqq);
1471                 cfq_mark_cfqq_prio_changed(cfqq);
1472                 cfq_init_prio_data(cfqq);
1473         }
1474
1475         if (new_cfqq)
1476                 kmem_cache_free(cfq_pool, new_cfqq);
1477
1478         atomic_inc(&cfqq->ref);
1479 out:
1480         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1481         return cfqq;
1482 }
1483
1484 static void
1485 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1486 {
1487         spin_lock(&cfq_exit_lock);
1488         rb_erase(&cic->rb_node, &ioc->cic_root);
1489         list_del_init(&cic->queue_list);
1490         spin_unlock(&cfq_exit_lock);
1491         kmem_cache_free(cfq_ioc_pool, cic);
1492         atomic_dec(&ioc_count);
1493 }
1494
1495 static struct cfq_io_context *
1496 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1497 {
1498         struct rb_node *n;
1499         struct cfq_io_context *cic;
1500         void *k, *key = cfqd;
1501
1502 restart:
1503         n = ioc->cic_root.rb_node;
1504         while (n) {
1505                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1506                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1507                 k = cic->key;
1508                 if (unlikely(!k)) {
1509                         cfq_drop_dead_cic(ioc, cic);
1510                         goto restart;
1511                 }
1512
1513                 if (key < k)
1514                         n = n->rb_left;
1515                 else if (key > k)
1516                         n = n->rb_right;
1517                 else
1518                         return cic;
1519         }
1520
1521         return NULL;
1522 }
1523
1524 static inline void
1525 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1526              struct cfq_io_context *cic)
1527 {
1528         struct rb_node **p;
1529         struct rb_node *parent;
1530         struct cfq_io_context *__cic;
1531         void *k;
1532
1533         cic->ioc = ioc;
1534         cic->key = cfqd;
1535
1536         ioc->set_ioprio = cfq_ioc_set_ioprio;
1537 restart:
1538         parent = NULL;
1539         p = &ioc->cic_root.rb_node;
1540         while (*p) {
1541                 parent = *p;
1542                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1543                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1544                 k = __cic->key;
1545                 if (unlikely(!k)) {
1546                         cfq_drop_dead_cic(ioc, cic);
1547                         goto restart;
1548                 }
1549
1550                 if (cic->key < k)
1551                         p = &(*p)->rb_left;
1552                 else if (cic->key > k)
1553                         p = &(*p)->rb_right;
1554                 else
1555                         BUG();
1556         }
1557
1558         spin_lock(&cfq_exit_lock);
1559         rb_link_node(&cic->rb_node, parent, p);
1560         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1561         list_add(&cic->queue_list, &cfqd->cic_list);
1562         spin_unlock(&cfq_exit_lock);
1563 }
1564
1565 /*
1566  * Setup general io context and cfq io context. There can be several cfq
1567  * io contexts per general io context, if this process is doing io to more
1568  * than one device managed by cfq.
1569  */
1570 static struct cfq_io_context *
1571 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1572 {
1573         struct io_context *ioc = NULL;
1574         struct cfq_io_context *cic;
1575
1576         might_sleep_if(gfp_mask & __GFP_WAIT);
1577
1578         ioc = get_io_context(gfp_mask);
1579         if (!ioc)
1580                 return NULL;
1581
1582         cic = cfq_cic_rb_lookup(cfqd, ioc);
1583         if (cic)
1584                 goto out;
1585
1586         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1587         if (cic == NULL)
1588                 goto err;
1589
1590         cfq_cic_link(cfqd, ioc, cic);
1591 out:
1592         return cic;
1593 err:
1594         put_io_context(ioc);
1595         return NULL;
1596 }
1597
1598 static void
1599 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1600 {
1601         unsigned long elapsed, ttime;
1602
1603         /*
1604          * if this context already has stuff queued, thinktime is from
1605          * last queue not last end
1606          */
1607 #if 0
1608         if (time_after(cic->last_end_request, cic->last_queue))
1609                 elapsed = jiffies - cic->last_end_request;
1610         else
1611                 elapsed = jiffies - cic->last_queue;
1612 #else
1613                 elapsed = jiffies - cic->last_end_request;
1614 #endif
1615
1616         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1617
1618         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1619         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1620         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1621 }
1622
1623 static void
1624 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1625                        struct cfq_rq *crq)
1626 {
1627         sector_t sdist;
1628         u64 total;
1629
1630         if (cic->last_request_pos < crq->request->sector)
1631                 sdist = crq->request->sector - cic->last_request_pos;
1632         else
1633                 sdist = cic->last_request_pos - crq->request->sector;
1634
1635         /*
1636          * Don't allow the seek distance to get too large from the
1637          * odd fragment, pagein, etc
1638          */
1639         if (cic->seek_samples <= 60) /* second&third seek */
1640                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1641         else
1642                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1643
1644         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1645         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1646         total = cic->seek_total + (cic->seek_samples/2);
1647         do_div(total, cic->seek_samples);
1648         cic->seek_mean = (sector_t)total;
1649 }
1650
1651 /*
1652  * Disable idle window if the process thinks too long or seeks so much that
1653  * it doesn't matter
1654  */
1655 static void
1656 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1657                        struct cfq_io_context *cic)
1658 {
1659         int enable_idle = cfq_cfqq_idle_window(cfqq);
1660
1661         if (!cic->ioc->task || !cfqd->cfq_slice_idle || cfqd->hw_tag)
1662                 enable_idle = 0;
1663         else if (sample_valid(cic->ttime_samples)) {
1664                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1665                         enable_idle = 0;
1666                 else
1667                         enable_idle = 1;
1668         }
1669
1670         if (enable_idle)
1671                 cfq_mark_cfqq_idle_window(cfqq);
1672         else
1673                 cfq_clear_cfqq_idle_window(cfqq);
1674 }
1675
1676
1677 /*
1678  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1679  * no or if we aren't sure, a 1 will cause a preempt.
1680  */
1681 static int
1682 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1683                    struct cfq_rq *crq)
1684 {
1685         struct cfq_queue *cfqq = cfqd->active_queue;
1686
1687         if (cfq_class_idle(new_cfqq))
1688                 return 0;
1689
1690         if (!cfqq)
1691                 return 1;
1692
1693         if (cfq_class_idle(cfqq))
1694                 return 1;
1695         if (!cfq_cfqq_wait_request(new_cfqq))
1696                 return 0;
1697         /*
1698          * if it doesn't have slice left, forget it
1699          */
1700         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1701                 return 0;
1702         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1703                 return 1;
1704
1705         return 0;
1706 }
1707
1708 /*
1709  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1710  * let it have half of its nominal slice.
1711  */
1712 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1713 {
1714         struct cfq_queue *__cfqq, *next;
1715
1716         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1717                 cfq_resort_rr_list(__cfqq, 1);
1718
1719         if (!cfqq->slice_left)
1720                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1721
1722         cfqq->slice_end = cfqq->slice_left + jiffies;
1723         __cfq_slice_expired(cfqd, cfqq, 1);
1724         __cfq_set_active_queue(cfqd, cfqq);
1725 }
1726
1727 /*
1728  * should really be a ll_rw_blk.c helper
1729  */
1730 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1731 {
1732         request_queue_t *q = cfqd->queue;
1733
1734         if (!blk_queue_plugged(q))
1735                 q->request_fn(q);
1736         else
1737                 __generic_unplug_device(q);
1738 }
1739
1740 /*
1741  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1742  * something we should do about it
1743  */
1744 static void
1745 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1746                  struct cfq_rq *crq)
1747 {
1748         struct cfq_io_context *cic;
1749
1750         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1751
1752         cic = crq->io_context;
1753
1754         /*
1755          * we never wait for an async request and we don't allow preemption
1756          * of an async request. so just return early
1757          */
1758         if (!cfq_crq_is_sync(crq)) {
1759                 /*
1760                  * sync process issued an async request, if it's waiting
1761                  * then expire it and kick rq handling.
1762                  */
1763                 if (cic == cfqd->active_cic &&
1764                     del_timer(&cfqd->idle_slice_timer)) {
1765                         cfq_slice_expired(cfqd, 0);
1766                         cfq_start_queueing(cfqd, cfqq);
1767                 }
1768                 return;
1769         }
1770
1771         cfq_update_io_thinktime(cfqd, cic);
1772         cfq_update_io_seektime(cfqd, cic, crq);
1773         cfq_update_idle_window(cfqd, cfqq, cic);
1774
1775         cic->last_queue = jiffies;
1776         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1777
1778         if (cfqq == cfqd->active_queue) {
1779                 /*
1780                  * if we are waiting for a request for this queue, let it rip
1781                  * immediately and flag that we must not expire this queue
1782                  * just now
1783                  */
1784                 if (cfq_cfqq_wait_request(cfqq)) {
1785                         cfq_mark_cfqq_must_dispatch(cfqq);
1786                         del_timer(&cfqd->idle_slice_timer);
1787                         cfq_start_queueing(cfqd, cfqq);
1788                 }
1789         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1790                 /*
1791                  * not the active queue - expire current slice if it is
1792                  * idle and has expired it's mean thinktime or this new queue
1793                  * has some old slice time left and is of higher priority
1794                  */
1795                 cfq_preempt_queue(cfqd, cfqq);
1796                 cfq_mark_cfqq_must_dispatch(cfqq);
1797                 cfq_start_queueing(cfqd, cfqq);
1798         }
1799 }
1800
1801 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1802 {
1803         struct cfq_data *cfqd = q->elevator->elevator_data;
1804         struct cfq_rq *crq = RQ_DATA(rq);
1805         struct cfq_queue *cfqq = crq->cfq_queue;
1806
1807         cfq_init_prio_data(cfqq);
1808
1809         cfq_add_crq_rb(crq);
1810
1811         list_add_tail(&rq->queuelist, &cfqq->fifo);
1812
1813         if (rq_mergeable(rq))
1814                 cfq_add_crq_hash(cfqd, crq);
1815
1816         cfq_crq_enqueued(cfqd, cfqq, crq);
1817 }
1818
1819 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1820 {
1821         struct cfq_rq *crq = RQ_DATA(rq);
1822         struct cfq_queue *cfqq = crq->cfq_queue;
1823         struct cfq_data *cfqd = cfqq->cfqd;
1824         const int sync = cfq_crq_is_sync(crq);
1825         unsigned long now;
1826
1827         now = jiffies;
1828
1829         WARN_ON(!cfqd->rq_in_driver);
1830         WARN_ON(!cfqq->on_dispatch[sync]);
1831         cfqd->rq_in_driver--;
1832         cfqq->on_dispatch[sync]--;
1833
1834         if (!cfq_class_idle(cfqq))
1835                 cfqd->last_end_request = now;
1836
1837         if (!cfq_cfqq_dispatched(cfqq)) {
1838                 if (cfq_cfqq_on_rr(cfqq)) {
1839                         cfqq->service_last = now;
1840                         cfq_resort_rr_list(cfqq, 0);
1841                 }
1842                 cfq_schedule_dispatch(cfqd);
1843         }
1844
1845         if (cfq_crq_is_sync(crq))
1846                 crq->io_context->last_end_request = now;
1847 }
1848
1849 static struct request *
1850 cfq_former_request(request_queue_t *q, struct request *rq)
1851 {
1852         struct cfq_rq *crq = RQ_DATA(rq);
1853         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1854
1855         if (rbprev)
1856                 return rb_entry_crq(rbprev)->request;
1857
1858         return NULL;
1859 }
1860
1861 static struct request *
1862 cfq_latter_request(request_queue_t *q, struct request *rq)
1863 {
1864         struct cfq_rq *crq = RQ_DATA(rq);
1865         struct rb_node *rbnext = rb_next(&crq->rb_node);
1866
1867         if (rbnext)
1868                 return rb_entry_crq(rbnext)->request;
1869
1870         return NULL;
1871 }
1872
1873 /*
1874  * we temporarily boost lower priority queues if they are holding fs exclusive
1875  * resources. they are boosted to normal prio (CLASS_BE/4)
1876  */
1877 static void cfq_prio_boost(struct cfq_queue *cfqq)
1878 {
1879         const int ioprio_class = cfqq->ioprio_class;
1880         const int ioprio = cfqq->ioprio;
1881
1882         if (has_fs_excl()) {
1883                 /*
1884                  * boost idle prio on transactions that would lock out other
1885                  * users of the filesystem
1886                  */
1887                 if (cfq_class_idle(cfqq))
1888                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1889                 if (cfqq->ioprio > IOPRIO_NORM)
1890                         cfqq->ioprio = IOPRIO_NORM;
1891         } else {
1892                 /*
1893                  * check if we need to unboost the queue
1894                  */
1895                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1896                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1897                 if (cfqq->ioprio != cfqq->org_ioprio)
1898                         cfqq->ioprio = cfqq->org_ioprio;
1899         }
1900
1901         /*
1902          * refile between round-robin lists if we moved the priority class
1903          */
1904         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1905             cfq_cfqq_on_rr(cfqq))
1906                 cfq_resort_rr_list(cfqq, 0);
1907 }
1908
1909 static inline int
1910 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1911                 struct task_struct *task, int rw)
1912 {
1913 #if 1
1914         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1915             !cfq_cfqq_must_alloc_slice(cfqq)) {
1916                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1917                 return ELV_MQUEUE_MUST;
1918         }
1919
1920         return ELV_MQUEUE_MAY;
1921 #else
1922         if (!cfqq || task->flags & PF_MEMALLOC)
1923                 return ELV_MQUEUE_MAY;
1924         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1925                 if (cfq_cfqq_wait_request(cfqq))
1926                         return ELV_MQUEUE_MUST;
1927
1928                 /*
1929                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1930                  * can quickly flood the queue with writes from a single task
1931                  */
1932                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1933                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1934                         return ELV_MQUEUE_MUST;
1935                 }
1936
1937                 return ELV_MQUEUE_MAY;
1938         }
1939         if (cfq_class_idle(cfqq))
1940                 return ELV_MQUEUE_NO;
1941         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1942                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1943                 int ret = ELV_MQUEUE_NO;
1944
1945                 if (ioc && ioc->nr_batch_requests)
1946                         ret = ELV_MQUEUE_MAY;
1947
1948                 put_io_context(ioc);
1949                 return ret;
1950         }
1951
1952         return ELV_MQUEUE_MAY;
1953 #endif
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         struct request_list *rl = &q->rq;
1983
1984         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1985                 smp_mb();
1986                 if (waitqueue_active(&rl->wait[READ]))
1987                         wake_up(&rl->wait[READ]);
1988         }
1989
1990         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1991                 smp_mb();
1992                 if (waitqueue_active(&rl->wait[WRITE]))
1993                         wake_up(&rl->wait[WRITE]);
1994         }
1995 }
1996
1997 /*
1998  * queue lock held here
1999  */
2000 static void cfq_put_request(request_queue_t *q, struct request *rq)
2001 {
2002         struct cfq_data *cfqd = q->elevator->elevator_data;
2003         struct cfq_rq *crq = RQ_DATA(rq);
2004
2005         if (crq) {
2006                 struct cfq_queue *cfqq = crq->cfq_queue;
2007                 const int rw = rq_data_dir(rq);
2008
2009                 BUG_ON(!cfqq->allocated[rw]);
2010                 cfqq->allocated[rw]--;
2011
2012                 put_io_context(crq->io_context->ioc);
2013
2014                 mempool_free(crq, cfqd->crq_pool);
2015                 rq->elevator_private = NULL;
2016
2017                 cfq_check_waiters(q, cfqq);
2018                 cfq_put_queue(cfqq);
2019         }
2020 }
2021
2022 /*
2023  * Allocate cfq data structures associated with this request.
2024  */
2025 static int
2026 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2027                 gfp_t gfp_mask)
2028 {
2029         struct cfq_data *cfqd = q->elevator->elevator_data;
2030         struct task_struct *tsk = current;
2031         struct cfq_io_context *cic;
2032         const int rw = rq_data_dir(rq);
2033         pid_t key = cfq_queue_pid(tsk, rw);
2034         struct cfq_queue *cfqq;
2035         struct cfq_rq *crq;
2036         unsigned long flags;
2037         int is_sync = key != CFQ_KEY_ASYNC;
2038
2039         might_sleep_if(gfp_mask & __GFP_WAIT);
2040
2041         cic = cfq_get_io_context(cfqd, gfp_mask);
2042
2043         spin_lock_irqsave(q->queue_lock, flags);
2044
2045         if (!cic)
2046                 goto queue_fail;
2047
2048         if (!cic->cfqq[is_sync]) {
2049                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2050                 if (!cfqq)
2051                         goto queue_fail;
2052
2053                 cic->cfqq[is_sync] = cfqq;
2054         } else
2055                 cfqq = cic->cfqq[is_sync];
2056
2057         cfqq->allocated[rw]++;
2058         cfq_clear_cfqq_must_alloc(cfqq);
2059         cfqd->rq_starved = 0;
2060         atomic_inc(&cfqq->ref);
2061         spin_unlock_irqrestore(q->queue_lock, flags);
2062
2063         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2064         if (crq) {
2065                 RB_CLEAR(&crq->rb_node);
2066                 crq->rb_key = 0;
2067                 crq->request = rq;
2068                 INIT_HLIST_NODE(&crq->hash);
2069                 crq->cfq_queue = cfqq;
2070                 crq->io_context = cic;
2071
2072                 if (is_sync)
2073                         cfq_mark_crq_is_sync(crq);
2074                 else
2075                         cfq_clear_crq_is_sync(crq);
2076
2077                 rq->elevator_private = crq;
2078                 return 0;
2079         }
2080
2081         spin_lock_irqsave(q->queue_lock, flags);
2082         cfqq->allocated[rw]--;
2083         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2084                 cfq_mark_cfqq_must_alloc(cfqq);
2085         cfq_put_queue(cfqq);
2086 queue_fail:
2087         if (cic)
2088                 put_io_context(cic->ioc);
2089         /*
2090          * mark us rq allocation starved. we need to kickstart the process
2091          * ourselves if there are no pending requests that can do it for us.
2092          * that would be an extremely rare OOM situation
2093          */
2094         cfqd->rq_starved = 1;
2095         cfq_schedule_dispatch(cfqd);
2096         spin_unlock_irqrestore(q->queue_lock, flags);
2097         return 1;
2098 }
2099
2100 static void cfq_kick_queue(void *data)
2101 {
2102         request_queue_t *q = data;
2103         struct cfq_data *cfqd = q->elevator->elevator_data;
2104         unsigned long flags;
2105
2106         spin_lock_irqsave(q->queue_lock, flags);
2107
2108         if (cfqd->rq_starved) {
2109                 struct request_list *rl = &q->rq;
2110
2111                 /*
2112                  * we aren't guaranteed to get a request after this, but we
2113                  * have to be opportunistic
2114                  */
2115                 smp_mb();
2116                 if (waitqueue_active(&rl->wait[READ]))
2117                         wake_up(&rl->wait[READ]);
2118                 if (waitqueue_active(&rl->wait[WRITE]))
2119                         wake_up(&rl->wait[WRITE]);
2120         }
2121
2122         blk_remove_plug(q);
2123         q->request_fn(q);
2124         spin_unlock_irqrestore(q->queue_lock, flags);
2125 }
2126
2127 /*
2128  * Timer running if the active_queue is currently idling inside its time slice
2129  */
2130 static void cfq_idle_slice_timer(unsigned long data)
2131 {
2132         struct cfq_data *cfqd = (struct cfq_data *) data;
2133         struct cfq_queue *cfqq;
2134         unsigned long flags;
2135
2136         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2137
2138         if ((cfqq = cfqd->active_queue) != NULL) {
2139                 unsigned long now = jiffies;
2140
2141                 /*
2142                  * expired
2143                  */
2144                 if (time_after(now, cfqq->slice_end))
2145                         goto expire;
2146
2147                 /*
2148                  * only expire and reinvoke request handler, if there are
2149                  * other queues with pending requests
2150                  */
2151                 if (!cfqd->busy_queues) {
2152                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2153                         add_timer(&cfqd->idle_slice_timer);
2154                         goto out_cont;
2155                 }
2156
2157                 /*
2158                  * not expired and it has a request pending, let it dispatch
2159                  */
2160                 if (!RB_EMPTY(&cfqq->sort_list)) {
2161                         cfq_mark_cfqq_must_dispatch(cfqq);
2162                         goto out_kick;
2163                 }
2164         }
2165 expire:
2166         cfq_slice_expired(cfqd, 0);
2167 out_kick:
2168         cfq_schedule_dispatch(cfqd);
2169 out_cont:
2170         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2171 }
2172
2173 /*
2174  * Timer running if an idle class queue is waiting for service
2175  */
2176 static void cfq_idle_class_timer(unsigned long data)
2177 {
2178         struct cfq_data *cfqd = (struct cfq_data *) data;
2179         unsigned long flags, end;
2180
2181         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2182
2183         /*
2184          * race with a non-idle queue, reset timer
2185          */
2186         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2187         if (!time_after_eq(jiffies, end))
2188                 mod_timer(&cfqd->idle_class_timer, end);
2189         else
2190                 cfq_schedule_dispatch(cfqd);
2191
2192         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2193 }
2194
2195 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2196 {
2197         del_timer_sync(&cfqd->idle_slice_timer);
2198         del_timer_sync(&cfqd->idle_class_timer);
2199         blk_sync_queue(cfqd->queue);
2200 }
2201
2202 static void cfq_exit_queue(elevator_t *e)
2203 {
2204         struct cfq_data *cfqd = e->elevator_data;
2205         request_queue_t *q = cfqd->queue;
2206
2207         cfq_shutdown_timer_wq(cfqd);
2208
2209         spin_lock(&cfq_exit_lock);
2210         spin_lock_irq(q->queue_lock);
2211
2212         if (cfqd->active_queue)
2213                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2214
2215         while (!list_empty(&cfqd->cic_list)) {
2216                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2217                                                         struct cfq_io_context,
2218                                                         queue_list);
2219                 if (cic->cfqq[ASYNC]) {
2220                         cfq_put_queue(cic->cfqq[ASYNC]);
2221                         cic->cfqq[ASYNC] = NULL;
2222                 }
2223                 if (cic->cfqq[SYNC]) {
2224                         cfq_put_queue(cic->cfqq[SYNC]);
2225                         cic->cfqq[SYNC] = NULL;
2226                 }
2227                 cic->key = NULL;
2228                 list_del_init(&cic->queue_list);
2229         }
2230
2231         spin_unlock_irq(q->queue_lock);
2232         spin_unlock(&cfq_exit_lock);
2233
2234         cfq_shutdown_timer_wq(cfqd);
2235
2236         mempool_destroy(cfqd->crq_pool);
2237         kfree(cfqd->crq_hash);
2238         kfree(cfqd->cfq_hash);
2239         kfree(cfqd);
2240 }
2241
2242 static void *cfq_init_queue(request_queue_t *q, elevator_t *e)
2243 {
2244         struct cfq_data *cfqd;
2245         int i;
2246
2247         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2248         if (!cfqd)
2249                 return NULL;
2250
2251         memset(cfqd, 0, sizeof(*cfqd));
2252
2253         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2254                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2255
2256         INIT_LIST_HEAD(&cfqd->busy_rr);
2257         INIT_LIST_HEAD(&cfqd->cur_rr);
2258         INIT_LIST_HEAD(&cfqd->idle_rr);
2259         INIT_LIST_HEAD(&cfqd->empty_list);
2260         INIT_LIST_HEAD(&cfqd->cic_list);
2261
2262         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2263         if (!cfqd->crq_hash)
2264                 goto out_crqhash;
2265
2266         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2267         if (!cfqd->cfq_hash)
2268                 goto out_cfqhash;
2269
2270         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2271         if (!cfqd->crq_pool)
2272                 goto out_crqpool;
2273
2274         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2275                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2276         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2277                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2278
2279         cfqd->queue = q;
2280
2281         cfqd->max_queued = q->nr_requests / 4;
2282         q->nr_batching = cfq_queued;
2283
2284         init_timer(&cfqd->idle_slice_timer);
2285         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2286         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2287
2288         init_timer(&cfqd->idle_class_timer);
2289         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2290         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2291
2292         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2293
2294         cfqd->cfq_queued = cfq_queued;
2295         cfqd->cfq_quantum = cfq_quantum;
2296         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2297         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2298         cfqd->cfq_back_max = cfq_back_max;
2299         cfqd->cfq_back_penalty = cfq_back_penalty;
2300         cfqd->cfq_slice[0] = cfq_slice_async;
2301         cfqd->cfq_slice[1] = cfq_slice_sync;
2302         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2303         cfqd->cfq_slice_idle = cfq_slice_idle;
2304
2305         return cfqd;
2306 out_crqpool:
2307         kfree(cfqd->cfq_hash);
2308 out_cfqhash:
2309         kfree(cfqd->crq_hash);
2310 out_crqhash:
2311         kfree(cfqd);
2312         return NULL;
2313 }
2314
2315 static void cfq_slab_kill(void)
2316 {
2317         if (crq_pool)
2318                 kmem_cache_destroy(crq_pool);
2319         if (cfq_pool)
2320                 kmem_cache_destroy(cfq_pool);
2321         if (cfq_ioc_pool)
2322                 kmem_cache_destroy(cfq_ioc_pool);
2323 }
2324
2325 static int __init cfq_slab_setup(void)
2326 {
2327         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2328                                         NULL, NULL);
2329         if (!crq_pool)
2330                 goto fail;
2331
2332         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2333                                         NULL, NULL);
2334         if (!cfq_pool)
2335                 goto fail;
2336
2337         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2338                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2339         if (!cfq_ioc_pool)
2340                 goto fail;
2341
2342         return 0;
2343 fail:
2344         cfq_slab_kill();
2345         return -ENOMEM;
2346 }
2347
2348 /*
2349  * sysfs parts below -->
2350  */
2351
2352 static ssize_t
2353 cfq_var_show(unsigned int var, char *page)
2354 {
2355         return sprintf(page, "%d\n", var);
2356 }
2357
2358 static ssize_t
2359 cfq_var_store(unsigned int *var, const char *page, size_t count)
2360 {
2361         char *p = (char *) page;
2362
2363         *var = simple_strtoul(p, &p, 10);
2364         return count;
2365 }
2366
2367 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2368 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2369 {                                                                       \
2370         struct cfq_data *cfqd = e->elevator_data;                       \
2371         unsigned int __data = __VAR;                                    \
2372         if (__CONV)                                                     \
2373                 __data = jiffies_to_msecs(__data);                      \
2374         return cfq_var_show(__data, (page));                            \
2375 }
2376 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2377 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2378 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2379 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2380 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2381 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2382 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2383 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2384 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2385 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2386 #undef SHOW_FUNCTION
2387
2388 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2389 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2390 {                                                                       \
2391         struct cfq_data *cfqd = e->elevator_data;                       \
2392         unsigned int __data;                                            \
2393         int ret = cfq_var_store(&__data, (page), count);                \
2394         if (__data < (MIN))                                             \
2395                 __data = (MIN);                                         \
2396         else if (__data > (MAX))                                        \
2397                 __data = (MAX);                                         \
2398         if (__CONV)                                                     \
2399                 *(__PTR) = msecs_to_jiffies(__data);                    \
2400         else                                                            \
2401                 *(__PTR) = __data;                                      \
2402         return ret;                                                     \
2403 }
2404 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2405 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2406 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2407 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2408 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2409 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2410 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2411 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2412 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2413 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2414 #undef STORE_FUNCTION
2415
2416 #define CFQ_ATTR(name) \
2417         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2418
2419 static struct elv_fs_entry cfq_attrs[] = {
2420         CFQ_ATTR(quantum),
2421         CFQ_ATTR(queued),
2422         CFQ_ATTR(fifo_expire_sync),
2423         CFQ_ATTR(fifo_expire_async),
2424         CFQ_ATTR(back_seek_max),
2425         CFQ_ATTR(back_seek_penalty),
2426         CFQ_ATTR(slice_sync),
2427         CFQ_ATTR(slice_async),
2428         CFQ_ATTR(slice_async_rq),
2429         CFQ_ATTR(slice_idle),
2430         __ATTR_NULL
2431 };
2432
2433 static struct elevator_type iosched_cfq = {
2434         .ops = {
2435                 .elevator_merge_fn =            cfq_merge,
2436                 .elevator_merged_fn =           cfq_merged_request,
2437                 .elevator_merge_req_fn =        cfq_merged_requests,
2438                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2439                 .elevator_add_req_fn =          cfq_insert_request,
2440                 .elevator_activate_req_fn =     cfq_activate_request,
2441                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2442                 .elevator_queue_empty_fn =      cfq_queue_empty,
2443                 .elevator_completed_req_fn =    cfq_completed_request,
2444                 .elevator_former_req_fn =       cfq_former_request,
2445                 .elevator_latter_req_fn =       cfq_latter_request,
2446                 .elevator_set_req_fn =          cfq_set_request,
2447                 .elevator_put_req_fn =          cfq_put_request,
2448                 .elevator_may_queue_fn =        cfq_may_queue,
2449                 .elevator_init_fn =             cfq_init_queue,
2450                 .elevator_exit_fn =             cfq_exit_queue,
2451                 .trim =                         cfq_trim,
2452         },
2453         .elevator_attrs =       cfq_attrs,
2454         .elevator_name =        "cfq",
2455         .elevator_owner =       THIS_MODULE,
2456 };
2457
2458 static int __init cfq_init(void)
2459 {
2460         int ret;
2461
2462         /*
2463          * could be 0 on HZ < 1000 setups
2464          */
2465         if (!cfq_slice_async)
2466                 cfq_slice_async = 1;
2467         if (!cfq_slice_idle)
2468                 cfq_slice_idle = 1;
2469
2470         if (cfq_slab_setup())
2471                 return -ENOMEM;
2472
2473         ret = elv_register(&iosched_cfq);
2474         if (ret)
2475                 cfq_slab_kill();
2476
2477         return ret;
2478 }
2479
2480 static void __exit cfq_exit(void)
2481 {
2482         DECLARE_COMPLETION(all_gone);
2483         elv_unregister(&iosched_cfq);
2484         ioc_gone = &all_gone;
2485         /* ioc_gone's update must be visible before reading ioc_count */
2486         smp_wmb();
2487         if (atomic_read(&ioc_count))
2488                 wait_for_completion(ioc_gone);
2489         synchronize_rcu();
2490         cfq_slab_kill();
2491 }
2492
2493 module_init(cfq_init);
2494 module_exit(cfq_exit);
2495
2496 MODULE_AUTHOR("Jens Axboe");
2497 MODULE_LICENSE("GPL");
2498 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");