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_NONE                 (2)
  64 #define RB_EMPTY(node)          ((node)->rb_node == NULL)
  65 #define RB_CLEAR_COLOR(node)    (node)->rb_color = RB_NONE
  66 #define RB_CLEAR(node)          do {    \
  67         (node)->rb_parent = NULL;       \
  68         RB_CLEAR_COLOR((node));         \
  69         (node)->rb_right = NULL;        \
  70         (node)->rb_left = NULL;         \
  71 } while (0)
  72 #define RB_CLEAR_ROOT(root)     ((root)->rb_node = NULL)
  73 #define rb_entry_crq(node)      rb_entry((node), struct cfq_rq, rb_node)
  74 #define rq_rb_key(rq)           (rq)->sector
  75
  76 static kmem_cache_t *crq_pool;
  77 static kmem_cache_t *cfq_pool;
  78 static kmem_cache_t *cfq_ioc_pool;
  79
  80 static atomic_t ioc_count = ATOMIC_INIT(0);
  81 static struct completion *ioc_gone;
  82
  83 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  84 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  85 #define cfq_class_be(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_BE)
  86 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  87
  88 #define ASYNC                   (0)
  89 #define SYNC                    (1)
  90
  91 #define cfq_cfqq_dispatched(cfqq)       \
  92         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
  93
  94 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
  95
  96 #define cfq_cfqq_sync(cfqq)             \
  97         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
  98
  99 #define sample_valid(samples)   ((samples) > 80)
 100
 101 /*
 102  * Per block device queue structure
 103  */
 104 struct cfq_data {
 105         request_queue_t *queue;
 106
 107         /*
 108          * rr list of queues with requests and the count of them
 109          */
 110         struct list_head rr_list[CFQ_PRIO_LISTS];
 111         struct list_head busy_rr;
 112         struct list_head cur_rr;
 113         struct list_head idle_rr;
 114         unsigned int busy_queues;
 115
 116         /*
 117          * non-ordered list of empty cfqq's
 118          */
 119         struct list_head empty_list;
 120
 121         /*
 122          * cfqq lookup hash
 123          */
 124         struct hlist_head *cfq_hash;
 125
 126         /*
 127          * global crq hash for all queues
 128          */
 129         struct hlist_head *crq_hash;
 130
 131         unsigned int max_queued;
 132
 133         mempool_t *crq_pool;
 134
 135         int rq_in_driver;
 136
 137         /*
 138          * schedule slice state info
 139          */
 140         /*
 141          * idle window management
 142          */
 143         struct timer_list idle_slice_timer;
 144         struct work_struct unplug_work;
 145
 146         struct cfq_queue *active_queue;
 147         struct cfq_io_context *active_cic;
 148         int cur_prio, cur_end_prio;
 149         unsigned int dispatch_slice;
 150
 151         struct timer_list idle_class_timer;
 152
 153         sector_t last_sector;
 154         unsigned long last_end_request;
 155
 156         unsigned int rq_starved;
 157
 158         /*
 159          * tunables, see top of file
 160          */
 161         unsigned int cfq_quantum;
 162         unsigned int cfq_queued;
 163         unsigned int cfq_fifo_expire[2];
 164         unsigned int cfq_back_penalty;
 165         unsigned int cfq_back_max;
 166         unsigned int cfq_slice[2];
 167         unsigned int cfq_slice_async_rq;
 168         unsigned int cfq_slice_idle;
 169
 170         struct list_head cic_list;
 171 };
 172
 173 /*
 174  * Per process-grouping structure
 175  */
 176 struct cfq_queue {
 177         /* reference count */
 178         atomic_t ref;
 179         /* parent cfq_data */
 180         struct cfq_data *cfqd;
 181         /* cfqq lookup hash */
 182         struct hlist_node cfq_hash;
 183         /* hash key */
 184         unsigned int key;
 185         /* on either rr or empty list of cfqd */
 186         struct list_head cfq_list;
 187         /* sorted list of pending requests */
 188         struct rb_root sort_list;
 189         /* if fifo isn't expired, next request to serve */
 190         struct cfq_rq *next_crq;
 191         /* requests queued in sort_list */
 192         int queued[2];
 193         /* currently allocated requests */
 194         int allocated[2];
 195         /* fifo list of requests in sort_list */
 196         struct list_head fifo;
 197
 198         unsigned long slice_start;
 199         unsigned long slice_end;
 200         unsigned long slice_left;
 201         unsigned long service_last;
 202
 203         /* number of requests that are on the dispatch list */
 204         int on_dispatch[2];
 205
 206         /* io prio of this group */
 207         unsigned short ioprio, org_ioprio;
 208         unsigned short ioprio_class, org_ioprio_class;
 209
 210         /* various state flags, see below */
 211         unsigned int flags;
 212 };
 213
 214 struct cfq_rq {
 215         struct rb_node rb_node;
 216         sector_t rb_key;
 217         struct request *request;
 218         struct hlist_node hash;
 219
 220         struct cfq_queue *cfq_queue;
 221         struct cfq_io_context *io_context;
 222
 223         unsigned int crq_flags;
 224 };
 225
 226 enum cfqq_state_flags {
 227         CFQ_CFQQ_FLAG_on_rr = 0,
 228         CFQ_CFQQ_FLAG_wait_request,
 229         CFQ_CFQQ_FLAG_must_alloc,
 230         CFQ_CFQQ_FLAG_must_alloc_slice,
 231         CFQ_CFQQ_FLAG_must_dispatch,
 232         CFQ_CFQQ_FLAG_fifo_expire,
 233         CFQ_CFQQ_FLAG_idle_window,
 234         CFQ_CFQQ_FLAG_prio_changed,
 235 };
 236
 237 #define CFQ_CFQQ_FNS(name)                                              \
 238 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 239 {                                                                       \
 240         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
 241 }                                                                       \
 242 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 243 {                                                                       \
 244         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
 245 }                                                                       \
 246 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 247 {                                                                       \
 248         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
 249 }
 250
 251 CFQ_CFQQ_FNS(on_rr);
 252 CFQ_CFQQ_FNS(wait_request);
 253 CFQ_CFQQ_FNS(must_alloc);
 254 CFQ_CFQQ_FNS(must_alloc_slice);
 255 CFQ_CFQQ_FNS(must_dispatch);
 256 CFQ_CFQQ_FNS(fifo_expire);
 257 CFQ_CFQQ_FNS(idle_window);
 258 CFQ_CFQQ_FNS(prio_changed);
 259 #undef CFQ_CFQQ_FNS
 260
 261 enum cfq_rq_state_flags {
 262         CFQ_CRQ_FLAG_is_sync = 0,
 263 };
 264
 265 #define CFQ_CRQ_FNS(name)                                               \
 266 static inline void cfq_mark_crq_##name(struct cfq_rq *crq)              \
 267 {                                                                       \
 268         crq->crq_flags |= (1 << CFQ_CRQ_FLAG_##name);                   \
 269 }                                                                       \
 270 static inline void cfq_clear_crq_##name(struct cfq_rq *crq)             \
 271 {                                                                       \
 272         crq->crq_flags &= ~(1 << CFQ_CRQ_FLAG_##name);                  \
 273 }                                                                       \
 274 static inline int cfq_crq_##name(const struct cfq_rq *crq)              \
 275 {                                                                       \
 276         return (crq->crq_flags & (1 << CFQ_CRQ_FLAG_##name)) != 0;      \
 277 }
 278
 279 CFQ_CRQ_FNS(is_sync);
 280 #undef CFQ_CRQ_FNS
 281
 282 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
 283 static void cfq_dispatch_insert(request_queue_t *, struct cfq_rq *);
 284 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
 285
 286 #define process_sync(tsk)       ((tsk)->flags & PF_SYNCWRITE)
 287
 288 /*
 289  * lots of deadline iosched dupes, can be abstracted later...
 290  */
 291 static inline void cfq_del_crq_hash(struct cfq_rq *crq)
 292 {
 293         hlist_del_init(&crq->hash);
 294 }
 295
 296 static inline void cfq_add_crq_hash(struct cfq_data *cfqd, struct cfq_rq *crq)
 297 {
 298         const int hash_idx = CFQ_MHASH_FN(rq_hash_key(crq->request));
 299
 300         hlist_add_head(&crq->hash, &cfqd->crq_hash[hash_idx]);
 301 }
 302
 303 static struct request *cfq_find_rq_hash(struct cfq_data *cfqd, sector_t offset)
 304 {
 305         struct hlist_head *hash_list = &cfqd->crq_hash[CFQ_MHASH_FN(offset)];
 306         struct hlist_node *entry, *next;
 307
 308         hlist_for_each_safe(entry, next, hash_list) {
 309                 struct cfq_rq *crq = list_entry_hash(entry);
 310                 struct request *__rq = crq->request;
 311
 312                 if (!rq_mergeable(__rq)) {
 313                         cfq_del_crq_hash(crq);
 314                         continue;
 315                 }
 316
 317                 if (rq_hash_key(__rq) == offset)
 318                         return __rq;
 319         }
 320
 321         return NULL;
 322 }
 323
 324 /*
 325  * scheduler run of queue, if there are requests pending and no one in the
 326  * driver that will restart queueing
 327  */
 328 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 329 {
 330         if (cfqd->busy_queues)
 331                 kblockd_schedule_work(&cfqd->unplug_work);
 332 }
 333
 334 static int cfq_queue_empty(request_queue_t *q)
 335 {
 336         struct cfq_data *cfqd = q->elevator->elevator_data;
 337
 338         return !cfqd->busy_queues;
 339 }
 340
 341 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw)
 342 {
 343         if (rw == READ || process_sync(task))
 344                 return task->pid;
 345
 346         return CFQ_KEY_ASYNC;
 347 }
 348
 349 /*
 350  * Lifted from AS - choose which of crq1 and crq2 that is best served now.
 351  * We choose the request that is closest to the head right now. Distance
 352  * behind the head is penalized and only allowed to a certain extent.
 353  */
 354 static struct cfq_rq *
 355 cfq_choose_req(struct cfq_data *cfqd, struct cfq_rq *crq1, struct cfq_rq *crq2)
 356 {
 357         sector_t last, s1, s2, d1 = 0, d2 = 0;
 358         unsigned long back_max;
 359 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 360 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 361         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 362
 363         if (crq1 == NULL || crq1 == crq2)
 364                 return crq2;
 365         if (crq2 == NULL)
 366                 return crq1;
 367
 368         if (cfq_crq_is_sync(crq1) && !cfq_crq_is_sync(crq2))
 369                 return crq1;
 370         else if (cfq_crq_is_sync(crq2) && !cfq_crq_is_sync(crq1))
 371                 return crq2;
 372
 373         s1 = crq1->request->sector;
 374         s2 = crq2->request->sector;
 375
 376         last = cfqd->last_sector;
 377
 378         /*
 379          * by definition, 1KiB is 2 sectors
 380          */
 381         back_max = cfqd->cfq_back_max * 2;
 382
 383         /*
 384          * Strict one way elevator _except_ in the case where we allow
 385          * short backward seeks which are biased as twice the cost of a
 386          * similar forward seek.
 387          */
 388         if (s1 >= last)
 389                 d1 = s1 - last;
 390         else if (s1 + back_max >= last)
 391                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 392         else
 393                 wrap |= CFQ_RQ1_WRAP;
 394
 395         if (s2 >= last)
 396                 d2 = s2 - last;
 397         else if (s2 + back_max >= last)
 398                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 399         else
 400                 wrap |= CFQ_RQ2_WRAP;
 401
 402         /* Found required data */
 403
 404         /*
 405          * By doing switch() on the bit mask "wrap" we avoid having to
 406          * check two variables for all permutations: --> faster!
 407          */
 408         switch (wrap) {
 409         case 0: /* common case for CFQ: crq1 and crq2 not wrapped */
 410                 if (d1 < d2)
 411                         return crq1;
 412                 else if (d2 < d1)
 413                         return crq2;
 414                 else {
 415                         if (s1 >= s2)
 416                                 return crq1;
 417                         else
 418                                 return crq2;
 419                 }
 420
 421         case CFQ_RQ2_WRAP:
 422                 return crq1;
 423         case CFQ_RQ1_WRAP:
 424                 return crq2;
 425         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both crqs wrapped */
 426         default:
 427                 /*
 428                  * Since both rqs are wrapped,
 429                  * start with the one that's further behind head
 430                  * (--> only *one* back seek required),
 431                  * since back seek takes more time than forward.
 432                  */
 433                 if (s1 <= s2)
 434                         return crq1;
 435                 else
 436                         return crq2;
 437         }
 438 }
 439
 440 /*
 441  * would be nice to take fifo expire time into account as well
 442  */
 443 static struct cfq_rq *
 444 cfq_find_next_crq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 445                   struct cfq_rq *last)
 446 {
 447         struct cfq_rq *crq_next = NULL, *crq_prev = NULL;
 448         struct rb_node *rbnext, *rbprev;
 449
 450         if (!(rbnext = rb_next(&last->rb_node))) {
 451                 rbnext = rb_first(&cfqq->sort_list);
 452                 if (rbnext == &last->rb_node)
 453                         rbnext = NULL;
 454         }
 455
 456         rbprev = rb_prev(&last->rb_node);
 457
 458         if (rbprev)
 459                 crq_prev = rb_entry_crq(rbprev);
 460         if (rbnext)
 461                 crq_next = rb_entry_crq(rbnext);
 462
 463         return cfq_choose_req(cfqd, crq_next, crq_prev);
 464 }
 465
 466 static void cfq_update_next_crq(struct cfq_rq *crq)
 467 {
 468         struct cfq_queue *cfqq = crq->cfq_queue;
 469
 470         if (cfqq->next_crq == crq)
 471                 cfqq->next_crq = cfq_find_next_crq(cfqq->cfqd, cfqq, crq);
 472 }
 473
 474 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
 475 {
 476         struct cfq_data *cfqd = cfqq->cfqd;
 477         struct list_head *list, *entry;
 478
 479         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 480
 481         list_del(&cfqq->cfq_list);
 482
 483         if (cfq_class_rt(cfqq))
 484                 list = &cfqd->cur_rr;
 485         else if (cfq_class_idle(cfqq))
 486                 list = &cfqd->idle_rr;
 487         else {
 488                 /*
 489                  * if cfqq has requests in flight, don't allow it to be
 490                  * found in cfq_set_active_queue before it has finished them.
 491                  * this is done to increase fairness between a process that
 492                  * has lots of io pending vs one that only generates one
 493                  * sporadically or synchronously
 494                  */
 495                 if (cfq_cfqq_dispatched(cfqq))
 496                         list = &cfqd->busy_rr;
 497                 else
 498                         list = &cfqd->rr_list[cfqq->ioprio];
 499         }
 500
 501         /*
 502          * if queue was preempted, just add to front to be fair. busy_rr
 503          * isn't sorted.
 504          */
 505         if (preempted || list == &cfqd->busy_rr) {
 506                 list_add(&cfqq->cfq_list, list);
 507                 return;
 508         }
 509
 510         /*
 511          * sort by when queue was last serviced
 512          */
 513         entry = list;
 514         while ((entry = entry->prev) != list) {
 515                 struct cfq_queue *__cfqq = list_entry_cfqq(entry);
 516
 517                 if (!__cfqq->service_last)
 518                         break;
 519                 if (time_before(__cfqq->service_last, cfqq->service_last))
 520                         break;
 521         }
 522
 523         list_add(&cfqq->cfq_list, entry);
 524 }
 525
 526 /*
 527  * add to busy list of queues for service, trying to be fair in ordering
 528  * the pending list according to last request service
 529  */
 530 static inline void
 531 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 532 {
 533         BUG_ON(cfq_cfqq_on_rr(cfqq));
 534         cfq_mark_cfqq_on_rr(cfqq);
 535         cfqd->busy_queues++;
 536
 537         cfq_resort_rr_list(cfqq, 0);
 538 }
 539
 540 static inline void
 541 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 542 {
 543         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 544         cfq_clear_cfqq_on_rr(cfqq);
 545         list_move(&cfqq->cfq_list, &cfqd->empty_list);
 546
 547         BUG_ON(!cfqd->busy_queues);
 548         cfqd->busy_queues--;
 549 }
 550
 551 /*
 552  * rb tree support functions
 553  */
 554 static inline void cfq_del_crq_rb(struct cfq_rq *crq)
 555 {
 556         struct cfq_queue *cfqq = crq->cfq_queue;
 557         struct cfq_data *cfqd = cfqq->cfqd;
 558         const int sync = cfq_crq_is_sync(crq);
 559
 560         BUG_ON(!cfqq->queued[sync]);
 561         cfqq->queued[sync]--;
 562
 563         cfq_update_next_crq(crq);
 564
 565         rb_erase(&crq->rb_node, &cfqq->sort_list);
 566         RB_CLEAR_COLOR(&crq->rb_node);
 567
 568         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY(&cfqq->sort_list))
 569                 cfq_del_cfqq_rr(cfqd, cfqq);
 570 }
 571
 572 static struct cfq_rq *
 573 __cfq_add_crq_rb(struct cfq_rq *crq)
 574 {
 575         struct rb_node **p = &crq->cfq_queue->sort_list.rb_node;
 576         struct rb_node *parent = NULL;
 577         struct cfq_rq *__crq;
 578
 579         while (*p) {
 580                 parent = *p;
 581                 __crq = rb_entry_crq(parent);
 582
 583                 if (crq->rb_key < __crq->rb_key)
 584                         p = &(*p)->rb_left;
 585                 else if (crq->rb_key > __crq->rb_key)
 586                         p = &(*p)->rb_right;
 587                 else
 588                         return __crq;
 589         }
 590
 591         rb_link_node(&crq->rb_node, parent, p);
 592         return NULL;
 593 }
 594
 595 static void cfq_add_crq_rb(struct cfq_rq *crq)
 596 {
 597         struct cfq_queue *cfqq = crq->cfq_queue;
 598         struct cfq_data *cfqd = cfqq->cfqd;
 599         struct request *rq = crq->request;
 600         struct cfq_rq *__alias;
 601
 602         crq->rb_key = rq_rb_key(rq);
 603         cfqq->queued[cfq_crq_is_sync(crq)]++;
 604
 605         /*
 606          * looks a little odd, but the first insert might return an alias.
 607          * if that happens, put the alias on the dispatch list
 608          */
 609         while ((__alias = __cfq_add_crq_rb(crq)) != NULL)
 610                 cfq_dispatch_insert(cfqd->queue, __alias);
 611
 612         rb_insert_color(&crq->rb_node, &cfqq->sort_list);
 613
 614         if (!cfq_cfqq_on_rr(cfqq))
 615                 cfq_add_cfqq_rr(cfqd, cfqq);
 616
 617         /*
 618          * check if this request is a better next-serve candidate
 619          */
 620         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
 621 }
 622
 623 static inline void
 624 cfq_reposition_crq_rb(struct cfq_queue *cfqq, struct cfq_rq *crq)
 625 {
 626         rb_erase(&crq->rb_node, &cfqq->sort_list);
 627         cfqq->queued[cfq_crq_is_sync(crq)]--;
 628
 629         cfq_add_crq_rb(crq);
 630 }
 631
 632 static struct request *
 633 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 634 {
 635         struct task_struct *tsk = current;
 636         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio));
 637         struct cfq_queue *cfqq;
 638         struct rb_node *n;
 639         sector_t sector;
 640
 641         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
 642         if (!cfqq)
 643                 goto out;
 644
 645         sector = bio->bi_sector + bio_sectors(bio);
 646         n = cfqq->sort_list.rb_node;
 647         while (n) {
 648                 struct cfq_rq *crq = rb_entry_crq(n);
 649
 650                 if (sector < crq->rb_key)
 651                         n = n->rb_left;
 652                 else if (sector > crq->rb_key)
 653                         n = n->rb_right;
 654                 else
 655                         return crq->request;
 656         }
 657
 658 out:
 659         return NULL;
 660 }
 661
 662 static void cfq_activate_request(request_queue_t *q, struct request *rq)
 663 {
 664         struct cfq_data *cfqd = q->elevator->elevator_data;
 665
 666         cfqd->rq_in_driver++;
 667 }
 668
 669 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
 670 {
 671         struct cfq_data *cfqd = q->elevator->elevator_data;
 672
 673         WARN_ON(!cfqd->rq_in_driver);
 674         cfqd->rq_in_driver--;
 675 }
 676
 677 static void cfq_remove_request(struct request *rq)
 678 {
 679         struct cfq_rq *crq = RQ_DATA(rq);
 680
 681         list_del_init(&rq->queuelist);
 682         cfq_del_crq_rb(crq);
 683         cfq_del_crq_hash(crq);
 684 }
 685
 686 static int
 687 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
 688 {
 689         struct cfq_data *cfqd = q->elevator->elevator_data;
 690         struct request *__rq;
 691         int ret;
 692
 693         __rq = cfq_find_rq_hash(cfqd, bio->bi_sector);
 694         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 695                 ret = ELEVATOR_BACK_MERGE;
 696                 goto out;
 697         }
 698
 699         __rq = cfq_find_rq_fmerge(cfqd, bio);
 700         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 701                 ret = ELEVATOR_FRONT_MERGE;
 702                 goto out;
 703         }
 704
 705         return ELEVATOR_NO_MERGE;
 706 out:
 707         *req = __rq;
 708         return ret;
 709 }
 710
 711 static void cfq_merged_request(request_queue_t *q, struct request *req)
 712 {
 713         struct cfq_data *cfqd = q->elevator->elevator_data;
 714         struct cfq_rq *crq = RQ_DATA(req);
 715
 716         cfq_del_crq_hash(crq);
 717         cfq_add_crq_hash(cfqd, crq);
 718
 719         if (rq_rb_key(req) != crq->rb_key) {
 720                 struct cfq_queue *cfqq = crq->cfq_queue;
 721
 722                 cfq_update_next_crq(crq);
 723                 cfq_reposition_crq_rb(cfqq, crq);
 724         }
 725 }
 726
 727 static void
 728 cfq_merged_requests(request_queue_t *q, struct request *rq,
 729                     struct request *next)
 730 {
 731         cfq_merged_request(q, rq);
 732
 733         /*
 734          * reposition in fifo if next is older than rq
 735          */
 736         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 737             time_before(next->start_time, rq->start_time))
 738                 list_move(&rq->queuelist, &next->queuelist);
 739
 740         cfq_remove_request(next);
 741 }
 742
 743 static inline void
 744 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 745 {
 746         if (cfqq) {
 747                 /*
 748                  * stop potential idle class queues waiting service
 749                  */
 750                 del_timer(&cfqd->idle_class_timer);
 751
 752                 cfqq->slice_start = jiffies;
 753                 cfqq->slice_end = 0;
 754                 cfqq->slice_left = 0;
 755                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 756                 cfq_clear_cfqq_fifo_expire(cfqq);
 757         }
 758
 759         cfqd->active_queue = cfqq;
 760 }
 761
 762 /*
 763  * current cfqq expired its slice (or was too idle), select new one
 764  */
 765 static void
 766 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 767                     int preempted)
 768 {
 769         unsigned long now = jiffies;
 770
 771         if (cfq_cfqq_wait_request(cfqq))
 772                 del_timer(&cfqd->idle_slice_timer);
 773
 774         if (!preempted && !cfq_cfqq_dispatched(cfqq)) {
 775                 cfqq->service_last = now;
 776                 cfq_schedule_dispatch(cfqd);
 777         }
 778
 779         cfq_clear_cfqq_must_dispatch(cfqq);
 780         cfq_clear_cfqq_wait_request(cfqq);
 781
 782         /*
 783          * store what was left of this slice, if the queue idled out
 784          * or was preempted
 785          */
 786         if (time_after(cfqq->slice_end, now))
 787                 cfqq->slice_left = cfqq->slice_end - now;
 788         else
 789                 cfqq->slice_left = 0;
 790
 791         if (cfq_cfqq_on_rr(cfqq))
 792                 cfq_resort_rr_list(cfqq, preempted);
 793
 794         if (cfqq == cfqd->active_queue)
 795                 cfqd->active_queue = NULL;
 796
 797         if (cfqd->active_cic) {
 798                 put_io_context(cfqd->active_cic->ioc);
 799                 cfqd->active_cic = NULL;
 800         }
 801
 802         cfqd->dispatch_slice = 0;
 803 }
 804
 805 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
 806 {
 807         struct cfq_queue *cfqq = cfqd->active_queue;
 808
 809         if (cfqq)
 810                 __cfq_slice_expired(cfqd, cfqq, preempted);
 811 }
 812
 813 /*
 814  * 0
 815  * 0,1
 816  * 0,1,2
 817  * 0,1,2,3
 818  * 0,1,2,3,4
 819  * 0,1,2,3,4,5
 820  * 0,1,2,3,4,5,6
 821  * 0,1,2,3,4,5,6,7
 822  */
 823 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
 824 {
 825         int prio, wrap;
 826
 827         prio = -1;
 828         wrap = 0;
 829         do {
 830                 int p;
 831
 832                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
 833                         if (!list_empty(&cfqd->rr_list[p])) {
 834                                 prio = p;
 835                                 break;
 836                         }
 837                 }
 838
 839                 if (prio != -1)
 840                         break;
 841                 cfqd->cur_prio = 0;
 842                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 843                         cfqd->cur_end_prio = 0;
 844                         if (wrap)
 845                                 break;
 846                         wrap = 1;
 847                 }
 848         } while (1);
 849
 850         if (unlikely(prio == -1))
 851                 return -1;
 852
 853         BUG_ON(prio >= CFQ_PRIO_LISTS);
 854
 855         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
 856
 857         cfqd->cur_prio = prio + 1;
 858         if (cfqd->cur_prio > cfqd->cur_end_prio) {
 859                 cfqd->cur_end_prio = cfqd->cur_prio;
 860                 cfqd->cur_prio = 0;
 861         }
 862         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
 863                 cfqd->cur_prio = 0;
 864                 cfqd->cur_end_prio = 0;
 865         }
 866
 867         return prio;
 868 }
 869
 870 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 871 {
 872         struct cfq_queue *cfqq = NULL;
 873
 874         /*
 875          * if current list is non-empty, grab first entry. if it is empty,
 876          * get next prio level and grab first entry then if any are spliced
 877          */
 878         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1)
 879                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
 880
 881         /*
 882          * If no new queues are available, check if the busy list has some
 883          * before falling back to idle io.
 884          */
 885         if (!cfqq && !list_empty(&cfqd->busy_rr))
 886                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
 887
 888         /*
 889          * if we have idle queues and no rt or be queues had pending
 890          * requests, either allow immediate service if the grace period
 891          * has passed or arm the idle grace timer
 892          */
 893         if (!cfqq && !list_empty(&cfqd->idle_rr)) {
 894                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
 895
 896                 if (time_after_eq(jiffies, end))
 897                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
 898                 else
 899                         mod_timer(&cfqd->idle_class_timer, end);
 900         }
 901
 902         __cfq_set_active_queue(cfqd, cfqq);
 903         return cfqq;
 904 }
 905
 906 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 907
 908 {
 909         struct cfq_io_context *cic;
 910         unsigned long sl;
 911
 912         WARN_ON(!RB_EMPTY(&cfqq->sort_list));
 913         WARN_ON(cfqq != cfqd->active_queue);
 914
 915         /*
 916          * idle is disabled, either manually or by past process history
 917          */
 918         if (!cfqd->cfq_slice_idle)
 919                 return 0;
 920         if (!cfq_cfqq_idle_window(cfqq))
 921                 return 0;
 922         /*
 923          * task has exited, don't wait
 924          */
 925         cic = cfqd->active_cic;
 926         if (!cic || !cic->ioc->task)
 927                 return 0;
 928
 929         cfq_mark_cfqq_must_dispatch(cfqq);
 930         cfq_mark_cfqq_wait_request(cfqq);
 931
 932         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
 933
 934         /*
 935          * we don't want to idle for seeks, but we do want to allow
 936          * fair distribution of slice time for a process doing back-to-back
 937          * seeks. so allow a little bit of time for him to submit a new rq
 938          */
 939         if (sample_valid(cic->seek_samples) && cic->seek_mean > 131072)
 940                 sl = 2;
 941
 942         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 943         return 1;
 944 }
 945
 946 static void cfq_dispatch_insert(request_queue_t *q, struct cfq_rq *crq)
 947 {
 948         struct cfq_data *cfqd = q->elevator->elevator_data;
 949         struct cfq_queue *cfqq = crq->cfq_queue;
 950
 951         cfqq->next_crq = cfq_find_next_crq(cfqd, cfqq, crq);
 952         cfq_remove_request(crq->request);
 953         cfqq->on_dispatch[cfq_crq_is_sync(crq)]++;
 954         elv_dispatch_sort(q, crq->request);
 955 }
 956
 957 /*
 958  * return expired entry, or NULL to just start from scratch in rbtree
 959  */
 960 static inline struct cfq_rq *cfq_check_fifo(struct cfq_queue *cfqq)
 961 {
 962         struct cfq_data *cfqd = cfqq->cfqd;
 963         struct request *rq;
 964         struct cfq_rq *crq;
 965
 966         if (cfq_cfqq_fifo_expire(cfqq))
 967                 return NULL;
 968
 969         if (!list_empty(&cfqq->fifo)) {
 970                 int fifo = cfq_cfqq_class_sync(cfqq);
 971
 972                 crq = RQ_DATA(list_entry_fifo(cfqq->fifo.next));
 973                 rq = crq->request;
 974                 if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
 975                         cfq_mark_cfqq_fifo_expire(cfqq);
 976                         return crq;
 977                 }
 978         }
 979
 980         return NULL;
 981 }
 982
 983 /*
 984  * Scale schedule slice based on io priority. Use the sync time slice only
 985  * if a queue is marked sync and has sync io queued. A sync queue with async
 986  * io only, should not get full sync slice length.
 987  */
 988 static inline int
 989 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 990 {
 991         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
 992
 993         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 994
 995         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
 996 }
 997
 998 static inline void
 999 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1000 {
1001         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
1002 }
1003
1004 static inline int
1005 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1006 {
1007         const int base_rq = cfqd->cfq_slice_async_rq;
1008
1009         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1010
1011         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1012 }
1013
1014 /*
1015  * get next queue for service
1016  */
1017 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1018 {
1019         unsigned long now = jiffies;
1020         struct cfq_queue *cfqq;
1021
1022         cfqq = cfqd->active_queue;
1023         if (!cfqq)
1024                 goto new_queue;
1025
1026         /*
1027          * slice has expired
1028          */
1029         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
1030                 goto expire;
1031
1032         /*
1033          * if queue has requests, dispatch one. if not, check if
1034          * enough slice is left to wait for one
1035          */
1036         if (!RB_EMPTY(&cfqq->sort_list))
1037                 goto keep_queue;
1038         else if (cfq_cfqq_class_sync(cfqq) &&
1039                  time_before(now, cfqq->slice_end)) {
1040                 if (cfq_arm_slice_timer(cfqd, cfqq))
1041                         return NULL;
1042         }
1043
1044 expire:
1045         cfq_slice_expired(cfqd, 0);
1046 new_queue:
1047         cfqq = cfq_set_active_queue(cfqd);
1048 keep_queue:
1049         return cfqq;
1050 }
1051
1052 static int
1053 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1054                         int max_dispatch)
1055 {
1056         int dispatched = 0;
1057
1058         BUG_ON(RB_EMPTY(&cfqq->sort_list));
1059
1060         do {
1061                 struct cfq_rq *crq;
1062
1063                 /*
1064                  * follow expired path, else get first next available
1065                  */
1066                 if ((crq = cfq_check_fifo(cfqq)) == NULL)
1067                         crq = cfqq->next_crq;
1068
1069                 /*
1070                  * finally, insert request into driver dispatch list
1071                  */
1072                 cfq_dispatch_insert(cfqd->queue, crq);
1073
1074                 cfqd->dispatch_slice++;
1075                 dispatched++;
1076
1077                 if (!cfqd->active_cic) {
1078                         atomic_inc(&crq->io_context->ioc->refcount);
1079                         cfqd->active_cic = crq->io_context;
1080                 }
1081
1082                 if (RB_EMPTY(&cfqq->sort_list))
1083                         break;
1084
1085         } while (dispatched < max_dispatch);
1086
1087         /*
1088          * if slice end isn't set yet, set it. if at least one request was
1089          * sync, use the sync time slice value
1090          */
1091         if (!cfqq->slice_end)
1092                 cfq_set_prio_slice(cfqd, cfqq);
1093
1094         /*
1095          * expire an async queue immediately if it has used up its slice. idle
1096          * queue always expire after 1 dispatch round.
1097          */
1098         if ((!cfq_cfqq_sync(cfqq) &&
1099             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1100             cfq_class_idle(cfqq))
1101                 cfq_slice_expired(cfqd, 0);
1102
1103         return dispatched;
1104 }
1105
1106 static int
1107 cfq_forced_dispatch_cfqqs(struct list_head *list)
1108 {
1109         int dispatched = 0;
1110         struct cfq_queue *cfqq, *next;
1111         struct cfq_rq *crq;
1112
1113         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
1114                 while ((crq = cfqq->next_crq)) {
1115                         cfq_dispatch_insert(cfqq->cfqd->queue, crq);
1116                         dispatched++;
1117                 }
1118                 BUG_ON(!list_empty(&cfqq->fifo));
1119         }
1120         return dispatched;
1121 }
1122
1123 static int
1124 cfq_forced_dispatch(struct cfq_data *cfqd)
1125 {
1126         int i, dispatched = 0;
1127
1128         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1129                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
1130
1131         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
1132         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
1133         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
1134
1135         cfq_slice_expired(cfqd, 0);
1136
1137         BUG_ON(cfqd->busy_queues);
1138
1139         return dispatched;
1140 }
1141
1142 static int
1143 cfq_dispatch_requests(request_queue_t *q, int force)
1144 {
1145         struct cfq_data *cfqd = q->elevator->elevator_data;
1146         struct cfq_queue *cfqq;
1147
1148         if (!cfqd->busy_queues)
1149                 return 0;
1150
1151         if (unlikely(force))
1152                 return cfq_forced_dispatch(cfqd);
1153
1154         cfqq = cfq_select_queue(cfqd);
1155         if (cfqq) {
1156                 int max_dispatch;
1157
1158                 cfq_clear_cfqq_must_dispatch(cfqq);
1159                 cfq_clear_cfqq_wait_request(cfqq);
1160                 del_timer(&cfqd->idle_slice_timer);
1161
1162                 max_dispatch = cfqd->cfq_quantum;
1163                 if (cfq_class_idle(cfqq))
1164                         max_dispatch = 1;
1165
1166                 return __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1167         }
1168
1169         return 0;
1170 }
1171
1172 /*
1173  * task holds one reference to the queue, dropped when task exits. each crq
1174  * in-flight on this queue also holds a reference, dropped when crq is freed.
1175  *
1176  * queue lock must be held here.
1177  */
1178 static void cfq_put_queue(struct cfq_queue *cfqq)
1179 {
1180         struct cfq_data *cfqd = cfqq->cfqd;
1181
1182         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1183
1184         if (!atomic_dec_and_test(&cfqq->ref))
1185                 return;
1186
1187         BUG_ON(rb_first(&cfqq->sort_list));
1188         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1189         BUG_ON(cfq_cfqq_on_rr(cfqq));
1190
1191         if (unlikely(cfqd->active_queue == cfqq))
1192                 __cfq_slice_expired(cfqd, cfqq, 0);
1193
1194         /*
1195          * it's on the empty list and still hashed
1196          */
1197         list_del(&cfqq->cfq_list);
1198         hlist_del(&cfqq->cfq_hash);
1199         kmem_cache_free(cfq_pool, cfqq);
1200 }
1201
1202 static inline struct cfq_queue *
1203 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1204                     const int hashval)
1205 {
1206         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1207         struct hlist_node *entry;
1208         struct cfq_queue *__cfqq;
1209
1210         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1211                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1212
1213                 if (__cfqq->key == key && (__p == prio || !prio))
1214                         return __cfqq;
1215         }
1216
1217         return NULL;
1218 }
1219
1220 static struct cfq_queue *
1221 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1222 {
1223         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1224 }
1225
1226 static void cfq_free_io_context(struct io_context *ioc)
1227 {
1228         struct cfq_io_context *__cic;
1229         struct rb_node *n;
1230         int freed = 0;
1231
1232         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1233                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1234                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1235                 kmem_cache_free(cfq_ioc_pool, __cic);
1236                 freed++;
1237         }
1238
1239         if (atomic_sub_and_test(freed, &ioc_count) && ioc_gone)
1240                 complete(ioc_gone);
1241 }
1242
1243 static void cfq_trim(struct io_context *ioc)
1244 {
1245         ioc->set_ioprio = NULL;
1246         cfq_free_io_context(ioc);
1247 }
1248
1249 /*
1250  * Called with interrupts disabled
1251  */
1252 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1253 {
1254         struct cfq_data *cfqd = cic->key;
1255         request_queue_t *q;
1256
1257         if (!cfqd)
1258                 return;
1259
1260         q = cfqd->queue;
1261
1262         WARN_ON(!irqs_disabled());
1263
1264         spin_lock(q->queue_lock);
1265
1266         if (cic->cfqq[ASYNC]) {
1267                 if (unlikely(cic->cfqq[ASYNC] == cfqd->active_queue))
1268                         __cfq_slice_expired(cfqd, cic->cfqq[ASYNC], 0);
1269                 cfq_put_queue(cic->cfqq[ASYNC]);
1270                 cic->cfqq[ASYNC] = NULL;
1271         }
1272
1273         if (cic->cfqq[SYNC]) {
1274                 if (unlikely(cic->cfqq[SYNC] == cfqd->active_queue))
1275                         __cfq_slice_expired(cfqd, cic->cfqq[SYNC], 0);
1276                 cfq_put_queue(cic->cfqq[SYNC]);
1277                 cic->cfqq[SYNC] = NULL;
1278         }
1279
1280         cic->key = NULL;
1281         list_del_init(&cic->queue_list);
1282         spin_unlock(q->queue_lock);
1283 }
1284
1285 static void cfq_exit_io_context(struct io_context *ioc)
1286 {
1287         struct cfq_io_context *__cic;
1288         unsigned long flags;
1289         struct rb_node *n;
1290
1291         /*
1292          * put the reference this task is holding to the various queues
1293          */
1294         spin_lock_irqsave(&cfq_exit_lock, flags);
1295
1296         n = rb_first(&ioc->cic_root);
1297         while (n != NULL) {
1298                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1299
1300                 cfq_exit_single_io_context(__cic);
1301                 n = rb_next(n);
1302         }
1303
1304         spin_unlock_irqrestore(&cfq_exit_lock, flags);
1305 }
1306
1307 static struct cfq_io_context *
1308 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1309 {
1310         struct cfq_io_context *cic = kmem_cache_alloc(cfq_ioc_pool, gfp_mask);
1311
1312         if (cic) {
1313                 RB_CLEAR(&cic->rb_node);
1314                 cic->key = NULL;
1315                 cic->cfqq[ASYNC] = NULL;
1316                 cic->cfqq[SYNC] = NULL;
1317                 cic->last_end_request = jiffies;
1318                 cic->ttime_total = 0;
1319                 cic->ttime_samples = 0;
1320                 cic->ttime_mean = 0;
1321                 cic->dtor = cfq_free_io_context;
1322                 cic->exit = cfq_exit_io_context;
1323                 INIT_LIST_HEAD(&cic->queue_list);
1324                 atomic_inc(&ioc_count);
1325         }
1326
1327         return cic;
1328 }
1329
1330 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1331 {
1332         struct task_struct *tsk = current;
1333         int ioprio_class;
1334
1335         if (!cfq_cfqq_prio_changed(cfqq))
1336                 return;
1337
1338         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1339         switch (ioprio_class) {
1340                 default:
1341                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1342                 case IOPRIO_CLASS_NONE:
1343                         /*
1344                          * no prio set, place us in the middle of the BE classes
1345                          */
1346                         cfqq->ioprio = task_nice_ioprio(tsk);
1347                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1348                         break;
1349                 case IOPRIO_CLASS_RT:
1350                         cfqq->ioprio = task_ioprio(tsk);
1351                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1352                         break;
1353                 case IOPRIO_CLASS_BE:
1354                         cfqq->ioprio = task_ioprio(tsk);
1355                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1356                         break;
1357                 case IOPRIO_CLASS_IDLE:
1358                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1359                         cfqq->ioprio = 7;
1360                         cfq_clear_cfqq_idle_window(cfqq);
1361                         break;
1362         }
1363
1364         /*
1365          * keep track of original prio settings in case we have to temporarily
1366          * elevate the priority of this queue
1367          */
1368         cfqq->org_ioprio = cfqq->ioprio;
1369         cfqq->org_ioprio_class = cfqq->ioprio_class;
1370
1371         if (cfq_cfqq_on_rr(cfqq))
1372                 cfq_resort_rr_list(cfqq, 0);
1373
1374         cfq_clear_cfqq_prio_changed(cfqq);
1375 }
1376
1377 static inline void changed_ioprio(struct cfq_io_context *cic)
1378 {
1379         struct cfq_data *cfqd = cic->key;
1380         struct cfq_queue *cfqq;
1381         if (cfqd) {
1382                 spin_lock(cfqd->queue->queue_lock);
1383                 cfqq = cic->cfqq[ASYNC];
1384                 if (cfqq) {
1385                         struct cfq_queue *new_cfqq;
1386                         new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC,
1387                                                 cic->ioc->task, GFP_ATOMIC);
1388                         if (new_cfqq) {
1389                                 cic->cfqq[ASYNC] = new_cfqq;
1390                                 cfq_put_queue(cfqq);
1391                         }
1392                 }
1393                 cfqq = cic->cfqq[SYNC];
1394                 if (cfqq) {
1395                         cfq_mark_cfqq_prio_changed(cfqq);
1396                         cfq_init_prio_data(cfqq);
1397                 }
1398                 spin_unlock(cfqd->queue->queue_lock);
1399         }
1400 }
1401
1402 /*
1403  * callback from sys_ioprio_set, irqs are disabled
1404  */
1405 static int cfq_ioc_set_ioprio(struct io_context *ioc, unsigned int ioprio)
1406 {
1407         struct cfq_io_context *cic;
1408         struct rb_node *n;
1409
1410         spin_lock(&cfq_exit_lock);
1411
1412         n = rb_first(&ioc->cic_root);
1413         while (n != NULL) {
1414                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1415
1416                 changed_ioprio(cic);
1417                 n = rb_next(n);
1418         }
1419
1420         spin_unlock(&cfq_exit_lock);
1421
1422         return 0;
1423 }
1424
1425 static struct cfq_queue *
1426 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1427               gfp_t gfp_mask)
1428 {
1429         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1430         struct cfq_queue *cfqq, *new_cfqq = NULL;
1431         unsigned short ioprio;
1432
1433 retry:
1434         ioprio = tsk->ioprio;
1435         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1436
1437         if (!cfqq) {
1438                 if (new_cfqq) {
1439                         cfqq = new_cfqq;
1440                         new_cfqq = NULL;
1441                 } else if (gfp_mask & __GFP_WAIT) {
1442                         spin_unlock_irq(cfqd->queue->queue_lock);
1443                         new_cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1444                         spin_lock_irq(cfqd->queue->queue_lock);
1445                         goto retry;
1446                 } else {
1447                         cfqq = kmem_cache_alloc(cfq_pool, gfp_mask);
1448                         if (!cfqq)
1449                                 goto out;
1450                 }
1451
1452                 memset(cfqq, 0, sizeof(*cfqq));
1453
1454                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1455                 INIT_LIST_HEAD(&cfqq->cfq_list);
1456                 RB_CLEAR_ROOT(&cfqq->sort_list);
1457                 INIT_LIST_HEAD(&cfqq->fifo);
1458
1459                 cfqq->key = key;
1460                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1461                 atomic_set(&cfqq->ref, 0);
1462                 cfqq->cfqd = cfqd;
1463                 cfqq->service_last = 0;
1464                 /*
1465                  * set ->slice_left to allow preemption for a new process
1466                  */
1467                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1468                 cfq_mark_cfqq_idle_window(cfqq);
1469                 cfq_mark_cfqq_prio_changed(cfqq);
1470                 cfq_init_prio_data(cfqq);
1471         }
1472
1473         if (new_cfqq)
1474                 kmem_cache_free(cfq_pool, new_cfqq);
1475
1476         atomic_inc(&cfqq->ref);
1477 out:
1478         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1479         return cfqq;
1480 }
1481
1482 static void
1483 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1484 {
1485         spin_lock(&cfq_exit_lock);
1486         rb_erase(&cic->rb_node, &ioc->cic_root);
1487         list_del_init(&cic->queue_list);
1488         spin_unlock(&cfq_exit_lock);
1489         kmem_cache_free(cfq_ioc_pool, cic);
1490         atomic_dec(&ioc_count);
1491 }
1492
1493 static struct cfq_io_context *
1494 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1495 {
1496         struct rb_node *n;
1497         struct cfq_io_context *cic;
1498         void *k, *key = cfqd;
1499
1500 restart:
1501         n = ioc->cic_root.rb_node;
1502         while (n) {
1503                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1504                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1505                 k = cic->key;
1506                 if (unlikely(!k)) {
1507                         cfq_drop_dead_cic(ioc, cic);
1508                         goto restart;
1509                 }
1510
1511                 if (key < k)
1512                         n = n->rb_left;
1513                 else if (key > k)
1514                         n = n->rb_right;
1515                 else
1516                         return cic;
1517         }
1518
1519         return NULL;
1520 }
1521
1522 static inline void
1523 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1524              struct cfq_io_context *cic)
1525 {
1526         struct rb_node **p;
1527         struct rb_node *parent;
1528         struct cfq_io_context *__cic;
1529         void *k;
1530
1531         cic->ioc = ioc;
1532         cic->key = cfqd;
1533
1534         ioc->set_ioprio = cfq_ioc_set_ioprio;
1535 restart:
1536         parent = NULL;
1537         p = &ioc->cic_root.rb_node;
1538         while (*p) {
1539                 parent = *p;
1540                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1541                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1542                 k = __cic->key;
1543                 if (unlikely(!k)) {
1544                         cfq_drop_dead_cic(ioc, cic);
1545                         goto restart;
1546                 }
1547
1548                 if (cic->key < k)
1549                         p = &(*p)->rb_left;
1550                 else if (cic->key > k)
1551                         p = &(*p)->rb_right;
1552                 else
1553                         BUG();
1554         }
1555
1556         spin_lock(&cfq_exit_lock);
1557         rb_link_node(&cic->rb_node, parent, p);
1558         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1559         list_add(&cic->queue_list, &cfqd->cic_list);
1560         spin_unlock(&cfq_exit_lock);
1561 }
1562
1563 /*
1564  * Setup general io context and cfq io context. There can be several cfq
1565  * io contexts per general io context, if this process is doing io to more
1566  * than one device managed by cfq.
1567  */
1568 static struct cfq_io_context *
1569 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1570 {
1571         struct io_context *ioc = NULL;
1572         struct cfq_io_context *cic;
1573
1574         might_sleep_if(gfp_mask & __GFP_WAIT);
1575
1576         ioc = get_io_context(gfp_mask);
1577         if (!ioc)
1578                 return NULL;
1579
1580         cic = cfq_cic_rb_lookup(cfqd, ioc);
1581         if (cic)
1582                 goto out;
1583
1584         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1585         if (cic == NULL)
1586                 goto err;
1587
1588         cfq_cic_link(cfqd, ioc, cic);
1589 out:
1590         return cic;
1591 err:
1592         put_io_context(ioc);
1593         return NULL;
1594 }
1595
1596 static void
1597 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1598 {
1599         unsigned long elapsed, ttime;
1600
1601         /*
1602          * if this context already has stuff queued, thinktime is from
1603          * last queue not last end
1604          */
1605 #if 0
1606         if (time_after(cic->last_end_request, cic->last_queue))
1607                 elapsed = jiffies - cic->last_end_request;
1608         else
1609                 elapsed = jiffies - cic->last_queue;
1610 #else
1611                 elapsed = jiffies - cic->last_end_request;
1612 #endif
1613
1614         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1615
1616         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1617         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1618         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1619 }
1620
1621 static void
1622 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1623                        struct cfq_rq *crq)
1624 {
1625         sector_t sdist;
1626         u64 total;
1627
1628         if (cic->last_request_pos < crq->request->sector)
1629                 sdist = crq->request->sector - cic->last_request_pos;
1630         else
1631                 sdist = cic->last_request_pos - crq->request->sector;
1632
1633         /*
1634          * Don't allow the seek distance to get too large from the
1635          * odd fragment, pagein, etc
1636          */
1637         if (cic->seek_samples <= 60) /* second&third seek */
1638                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1639         else
1640                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1641
1642         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1643         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1644         total = cic->seek_total + (cic->seek_samples/2);
1645         do_div(total, cic->seek_samples);
1646         cic->seek_mean = (sector_t)total;
1647 }
1648
1649 /*
1650  * Disable idle window if the process thinks too long or seeks so much that
1651  * it doesn't matter
1652  */
1653 static void
1654 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1655                        struct cfq_io_context *cic)
1656 {
1657         int enable_idle = cfq_cfqq_idle_window(cfqq);
1658
1659         if (!cic->ioc->task || !cfqd->cfq_slice_idle)
1660                 enable_idle = 0;
1661         else if (sample_valid(cic->ttime_samples)) {
1662                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1663                         enable_idle = 0;
1664                 else
1665                         enable_idle = 1;
1666         }
1667
1668         if (enable_idle)
1669                 cfq_mark_cfqq_idle_window(cfqq);
1670         else
1671                 cfq_clear_cfqq_idle_window(cfqq);
1672 }
1673
1674
1675 /*
1676  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1677  * no or if we aren't sure, a 1 will cause a preempt.
1678  */
1679 static int
1680 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1681                    struct cfq_rq *crq)
1682 {
1683         struct cfq_queue *cfqq = cfqd->active_queue;
1684
1685         if (cfq_class_idle(new_cfqq))
1686                 return 0;
1687
1688         if (!cfqq)
1689                 return 1;
1690
1691         if (cfq_class_idle(cfqq))
1692                 return 1;
1693         if (!cfq_cfqq_wait_request(new_cfqq))
1694                 return 0;
1695         /*
1696          * if it doesn't have slice left, forget it
1697          */
1698         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1699                 return 0;
1700         if (cfq_crq_is_sync(crq) && !cfq_cfqq_sync(cfqq))
1701                 return 1;
1702
1703         return 0;
1704 }
1705
1706 /*
1707  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1708  * let it have half of its nominal slice.
1709  */
1710 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1711 {
1712         struct cfq_queue *__cfqq, *next;
1713
1714         list_for_each_entry_safe(__cfqq, next, &cfqd->cur_rr, cfq_list)
1715                 cfq_resort_rr_list(__cfqq, 1);
1716
1717         if (!cfqq->slice_left)
1718                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1719
1720         cfqq->slice_end = cfqq->slice_left + jiffies;
1721         __cfq_slice_expired(cfqd, cfqq, 1);
1722         __cfq_set_active_queue(cfqd, cfqq);
1723 }
1724
1725 /*
1726  * should really be a ll_rw_blk.c helper
1727  */
1728 static void cfq_start_queueing(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1729 {
1730         request_queue_t *q = cfqd->queue;
1731
1732         if (!blk_queue_plugged(q))
1733                 q->request_fn(q);
1734         else
1735                 __generic_unplug_device(q);
1736 }
1737
1738 /*
1739  * Called when a new fs request (crq) is added (to cfqq). Check if there's
1740  * something we should do about it
1741  */
1742 static void
1743 cfq_crq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1744                  struct cfq_rq *crq)
1745 {
1746         struct cfq_io_context *cic;
1747
1748         cfqq->next_crq = cfq_choose_req(cfqd, cfqq->next_crq, crq);
1749
1750         cic = crq->io_context;
1751
1752         /*
1753          * we never wait for an async request and we don't allow preemption
1754          * of an async request. so just return early
1755          */
1756         if (!cfq_crq_is_sync(crq)) {
1757                 /*
1758                  * sync process issued an async request, if it's waiting
1759                  * then expire it and kick rq handling.
1760                  */
1761                 if (cic == cfqd->active_cic &&
1762                     del_timer(&cfqd->idle_slice_timer)) {
1763                         cfq_slice_expired(cfqd, 0);
1764                         cfq_start_queueing(cfqd, cfqq);
1765                 }
1766                 return;
1767         }
1768
1769         cfq_update_io_thinktime(cfqd, cic);
1770         cfq_update_io_seektime(cfqd, cic, crq);
1771         cfq_update_idle_window(cfqd, cfqq, cic);
1772
1773         cic->last_queue = jiffies;
1774         cic->last_request_pos = crq->request->sector + crq->request->nr_sectors;
1775
1776         if (cfqq == cfqd->active_queue) {
1777                 /*
1778                  * if we are waiting for a request for this queue, let it rip
1779                  * immediately and flag that we must not expire this queue
1780                  * just now
1781                  */
1782                 if (cfq_cfqq_wait_request(cfqq)) {
1783                         cfq_mark_cfqq_must_dispatch(cfqq);
1784                         del_timer(&cfqd->idle_slice_timer);
1785                         cfq_start_queueing(cfqd, cfqq);
1786                 }
1787         } else if (cfq_should_preempt(cfqd, cfqq, crq)) {
1788                 /*
1789                  * not the active queue - expire current slice if it is
1790                  * idle and has expired it's mean thinktime or this new queue
1791                  * has some old slice time left and is of higher priority
1792                  */
1793                 cfq_preempt_queue(cfqd, cfqq);
1794                 cfq_mark_cfqq_must_dispatch(cfqq);
1795                 cfq_start_queueing(cfqd, cfqq);
1796         }
1797 }
1798
1799 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1800 {
1801         struct cfq_data *cfqd = q->elevator->elevator_data;
1802         struct cfq_rq *crq = RQ_DATA(rq);
1803         struct cfq_queue *cfqq = crq->cfq_queue;
1804
1805         cfq_init_prio_data(cfqq);
1806
1807         cfq_add_crq_rb(crq);
1808
1809         list_add_tail(&rq->queuelist, &cfqq->fifo);
1810
1811         if (rq_mergeable(rq))
1812                 cfq_add_crq_hash(cfqd, crq);
1813
1814         cfq_crq_enqueued(cfqd, cfqq, crq);
1815 }
1816
1817 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1818 {
1819         struct cfq_rq *crq = RQ_DATA(rq);
1820         struct cfq_queue *cfqq = crq->cfq_queue;
1821         struct cfq_data *cfqd = cfqq->cfqd;
1822         const int sync = cfq_crq_is_sync(crq);
1823         unsigned long now;
1824
1825         now = jiffies;
1826
1827         WARN_ON(!cfqd->rq_in_driver);
1828         WARN_ON(!cfqq->on_dispatch[sync]);
1829         cfqd->rq_in_driver--;
1830         cfqq->on_dispatch[sync]--;
1831
1832         if (!cfq_class_idle(cfqq))
1833                 cfqd->last_end_request = now;
1834
1835         if (!cfq_cfqq_dispatched(cfqq)) {
1836                 if (cfq_cfqq_on_rr(cfqq)) {
1837                         cfqq->service_last = now;
1838                         cfq_resort_rr_list(cfqq, 0);
1839                 }
1840                 cfq_schedule_dispatch(cfqd);
1841         }
1842
1843         if (cfq_crq_is_sync(crq))
1844                 crq->io_context->last_end_request = now;
1845 }
1846
1847 static struct request *
1848 cfq_former_request(request_queue_t *q, struct request *rq)
1849 {
1850         struct cfq_rq *crq = RQ_DATA(rq);
1851         struct rb_node *rbprev = rb_prev(&crq->rb_node);
1852
1853         if (rbprev)
1854                 return rb_entry_crq(rbprev)->request;
1855
1856         return NULL;
1857 }
1858
1859 static struct request *
1860 cfq_latter_request(request_queue_t *q, struct request *rq)
1861 {
1862         struct cfq_rq *crq = RQ_DATA(rq);
1863         struct rb_node *rbnext = rb_next(&crq->rb_node);
1864
1865         if (rbnext)
1866                 return rb_entry_crq(rbnext)->request;
1867
1868         return NULL;
1869 }
1870
1871 /*
1872  * we temporarily boost lower priority queues if they are holding fs exclusive
1873  * resources. they are boosted to normal prio (CLASS_BE/4)
1874  */
1875 static void cfq_prio_boost(struct cfq_queue *cfqq)
1876 {
1877         const int ioprio_class = cfqq->ioprio_class;
1878         const int ioprio = cfqq->ioprio;
1879
1880         if (has_fs_excl()) {
1881                 /*
1882                  * boost idle prio on transactions that would lock out other
1883                  * users of the filesystem
1884                  */
1885                 if (cfq_class_idle(cfqq))
1886                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1887                 if (cfqq->ioprio > IOPRIO_NORM)
1888                         cfqq->ioprio = IOPRIO_NORM;
1889         } else {
1890                 /*
1891                  * check if we need to unboost the queue
1892                  */
1893                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1894                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1895                 if (cfqq->ioprio != cfqq->org_ioprio)
1896                         cfqq->ioprio = cfqq->org_ioprio;
1897         }
1898
1899         /*
1900          * refile between round-robin lists if we moved the priority class
1901          */
1902         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1903             cfq_cfqq_on_rr(cfqq))
1904                 cfq_resort_rr_list(cfqq, 0);
1905 }
1906
1907 static inline int
1908 __cfq_may_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1909                 struct task_struct *task, int rw)
1910 {
1911 #if 1
1912         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1913             !cfq_cfqq_must_alloc_slice(cfqq)) {
1914                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1915                 return ELV_MQUEUE_MUST;
1916         }
1917
1918         return ELV_MQUEUE_MAY;
1919 #else
1920         if (!cfqq || task->flags & PF_MEMALLOC)
1921                 return ELV_MQUEUE_MAY;
1922         if (!cfqq->allocated[rw] || cfq_cfqq_must_alloc(cfqq)) {
1923                 if (cfq_cfqq_wait_request(cfqq))
1924                         return ELV_MQUEUE_MUST;
1925
1926                 /*
1927                  * only allow 1 ELV_MQUEUE_MUST per slice, otherwise we
1928                  * can quickly flood the queue with writes from a single task
1929                  */
1930                 if (rw == READ || !cfq_cfqq_must_alloc_slice(cfqq)) {
1931                         cfq_mark_cfqq_must_alloc_slice(cfqq);
1932                         return ELV_MQUEUE_MUST;
1933                 }
1934
1935                 return ELV_MQUEUE_MAY;
1936         }
1937         if (cfq_class_idle(cfqq))
1938                 return ELV_MQUEUE_NO;
1939         if (cfqq->allocated[rw] >= cfqd->max_queued) {
1940                 struct io_context *ioc = get_io_context(GFP_ATOMIC);
1941                 int ret = ELV_MQUEUE_NO;
1942
1943                 if (ioc && ioc->nr_batch_requests)
1944                         ret = ELV_MQUEUE_MAY;
1945
1946                 put_io_context(ioc);
1947                 return ret;
1948         }
1949
1950         return ELV_MQUEUE_MAY;
1951 #endif
1952 }
1953
1954 static int cfq_may_queue(request_queue_t *q, int rw, struct bio *bio)
1955 {
1956         struct cfq_data *cfqd = q->elevator->elevator_data;
1957         struct task_struct *tsk = current;
1958         struct cfq_queue *cfqq;
1959
1960         /*
1961          * don't force setup of a queue from here, as a call to may_queue
1962          * does not necessarily imply that a request actually will be queued.
1963          * so just lookup a possibly existing queue, or return 'may queue'
1964          * if that fails
1965          */
1966         cfqq = cfq_find_cfq_hash(cfqd, cfq_queue_pid(tsk, rw), tsk->ioprio);
1967         if (cfqq) {
1968                 cfq_init_prio_data(cfqq);
1969                 cfq_prio_boost(cfqq);
1970
1971                 return __cfq_may_queue(cfqd, cfqq, tsk, rw);
1972         }
1973
1974         return ELV_MQUEUE_MAY;
1975 }
1976
1977 static void cfq_check_waiters(request_queue_t *q, struct cfq_queue *cfqq)
1978 {
1979         struct cfq_data *cfqd = q->elevator->elevator_data;
1980         struct request_list *rl = &q->rq;
1981
1982         if (cfqq->allocated[READ] <= cfqd->max_queued || cfqd->rq_starved) {
1983                 smp_mb();
1984                 if (waitqueue_active(&rl->wait[READ]))
1985                         wake_up(&rl->wait[READ]);
1986         }
1987
1988         if (cfqq->allocated[WRITE] <= cfqd->max_queued || cfqd->rq_starved) {
1989                 smp_mb();
1990                 if (waitqueue_active(&rl->wait[WRITE]))
1991                         wake_up(&rl->wait[WRITE]);
1992         }
1993 }
1994
1995 /*
1996  * queue lock held here
1997  */
1998 static void cfq_put_request(request_queue_t *q, struct request *rq)
1999 {
2000         struct cfq_data *cfqd = q->elevator->elevator_data;
2001         struct cfq_rq *crq = RQ_DATA(rq);
2002
2003         if (crq) {
2004                 struct cfq_queue *cfqq = crq->cfq_queue;
2005                 const int rw = rq_data_dir(rq);
2006
2007                 BUG_ON(!cfqq->allocated[rw]);
2008                 cfqq->allocated[rw]--;
2009
2010                 put_io_context(crq->io_context->ioc);
2011
2012                 mempool_free(crq, cfqd->crq_pool);
2013                 rq->elevator_private = NULL;
2014
2015                 cfq_check_waiters(q, cfqq);
2016                 cfq_put_queue(cfqq);
2017         }
2018 }
2019
2020 /*
2021  * Allocate cfq data structures associated with this request.
2022  */
2023 static int
2024 cfq_set_request(request_queue_t *q, struct request *rq, struct bio *bio,
2025                 gfp_t gfp_mask)
2026 {
2027         struct cfq_data *cfqd = q->elevator->elevator_data;
2028         struct task_struct *tsk = current;
2029         struct cfq_io_context *cic;
2030         const int rw = rq_data_dir(rq);
2031         pid_t key = cfq_queue_pid(tsk, rw);
2032         struct cfq_queue *cfqq;
2033         struct cfq_rq *crq;
2034         unsigned long flags;
2035         int is_sync = key != CFQ_KEY_ASYNC;
2036
2037         might_sleep_if(gfp_mask & __GFP_WAIT);
2038
2039         cic = cfq_get_io_context(cfqd, gfp_mask);
2040
2041         spin_lock_irqsave(q->queue_lock, flags);
2042
2043         if (!cic)
2044                 goto queue_fail;
2045
2046         if (!cic->cfqq[is_sync]) {
2047                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
2048                 if (!cfqq)
2049                         goto queue_fail;
2050
2051                 cic->cfqq[is_sync] = cfqq;
2052         } else
2053                 cfqq = cic->cfqq[is_sync];
2054
2055         cfqq->allocated[rw]++;
2056         cfq_clear_cfqq_must_alloc(cfqq);
2057         cfqd->rq_starved = 0;
2058         atomic_inc(&cfqq->ref);
2059         spin_unlock_irqrestore(q->queue_lock, flags);
2060
2061         crq = mempool_alloc(cfqd->crq_pool, gfp_mask);
2062         if (crq) {
2063                 RB_CLEAR(&crq->rb_node);
2064                 crq->rb_key = 0;
2065                 crq->request = rq;
2066                 INIT_HLIST_NODE(&crq->hash);
2067                 crq->cfq_queue = cfqq;
2068                 crq->io_context = cic;
2069
2070                 if (is_sync)
2071                         cfq_mark_crq_is_sync(crq);
2072                 else
2073                         cfq_clear_crq_is_sync(crq);
2074
2075                 rq->elevator_private = crq;
2076                 return 0;
2077         }
2078
2079         spin_lock_irqsave(q->queue_lock, flags);
2080         cfqq->allocated[rw]--;
2081         if (!(cfqq->allocated[0] + cfqq->allocated[1]))
2082                 cfq_mark_cfqq_must_alloc(cfqq);
2083         cfq_put_queue(cfqq);
2084 queue_fail:
2085         if (cic)
2086                 put_io_context(cic->ioc);
2087         /*
2088          * mark us rq allocation starved. we need to kickstart the process
2089          * ourselves if there are no pending requests that can do it for us.
2090          * that would be an extremely rare OOM situation
2091          */
2092         cfqd->rq_starved = 1;
2093         cfq_schedule_dispatch(cfqd);
2094         spin_unlock_irqrestore(q->queue_lock, flags);
2095         return 1;
2096 }
2097
2098 static void cfq_kick_queue(void *data)
2099 {
2100         request_queue_t *q = data;
2101         struct cfq_data *cfqd = q->elevator->elevator_data;
2102         unsigned long flags;
2103
2104         spin_lock_irqsave(q->queue_lock, flags);
2105
2106         if (cfqd->rq_starved) {
2107                 struct request_list *rl = &q->rq;
2108
2109                 /*
2110                  * we aren't guaranteed to get a request after this, but we
2111                  * have to be opportunistic
2112                  */
2113                 smp_mb();
2114                 if (waitqueue_active(&rl->wait[READ]))
2115                         wake_up(&rl->wait[READ]);
2116                 if (waitqueue_active(&rl->wait[WRITE]))
2117                         wake_up(&rl->wait[WRITE]);
2118         }
2119
2120         blk_remove_plug(q);
2121         q->request_fn(q);
2122         spin_unlock_irqrestore(q->queue_lock, flags);
2123 }
2124
2125 /*
2126  * Timer running if the active_queue is currently idling inside its time slice
2127  */
2128 static void cfq_idle_slice_timer(unsigned long data)
2129 {
2130         struct cfq_data *cfqd = (struct cfq_data *) data;
2131         struct cfq_queue *cfqq;
2132         unsigned long flags;
2133
2134         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2135
2136         if ((cfqq = cfqd->active_queue) != NULL) {
2137                 unsigned long now = jiffies;
2138
2139                 /*
2140                  * expired
2141                  */
2142                 if (time_after(now, cfqq->slice_end))
2143                         goto expire;
2144
2145                 /*
2146                  * only expire and reinvoke request handler, if there are
2147                  * other queues with pending requests
2148                  */
2149                 if (!cfqd->busy_queues) {
2150                         cfqd->idle_slice_timer.expires = min(now + cfqd->cfq_slice_idle, cfqq->slice_end);
2151                         add_timer(&cfqd->idle_slice_timer);
2152                         goto out_cont;
2153                 }
2154
2155                 /*
2156                  * not expired and it has a request pending, let it dispatch
2157                  */
2158                 if (!RB_EMPTY(&cfqq->sort_list)) {
2159                         cfq_mark_cfqq_must_dispatch(cfqq);
2160                         goto out_kick;
2161                 }
2162         }
2163 expire:
2164         cfq_slice_expired(cfqd, 0);
2165 out_kick:
2166         cfq_schedule_dispatch(cfqd);
2167 out_cont:
2168         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2169 }
2170
2171 /*
2172  * Timer running if an idle class queue is waiting for service
2173  */
2174 static void cfq_idle_class_timer(unsigned long data)
2175 {
2176         struct cfq_data *cfqd = (struct cfq_data *) data;
2177         unsigned long flags, end;
2178
2179         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2180
2181         /*
2182          * race with a non-idle queue, reset timer
2183          */
2184         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
2185         if (!time_after_eq(jiffies, end)) {
2186                 cfqd->idle_class_timer.expires = end;
2187                 add_timer(&cfqd->idle_class_timer);
2188         } else
2189                 cfq_schedule_dispatch(cfqd);
2190
2191         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2192 }
2193
2194 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2195 {
2196         del_timer_sync(&cfqd->idle_slice_timer);
2197         del_timer_sync(&cfqd->idle_class_timer);
2198         blk_sync_queue(cfqd->queue);
2199 }
2200
2201 static void cfq_exit_queue(elevator_t *e)
2202 {
2203         struct cfq_data *cfqd = e->elevator_data;
2204         request_queue_t *q = cfqd->queue;
2205
2206         cfq_shutdown_timer_wq(cfqd);
2207
2208         spin_lock(&cfq_exit_lock);
2209         spin_lock_irq(q->queue_lock);
2210
2211         if (cfqd->active_queue)
2212                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2213
2214         while (!list_empty(&cfqd->cic_list)) {
2215                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2216                                                         struct cfq_io_context,
2217                                                         queue_list);
2218                 if (cic->cfqq[ASYNC]) {
2219                         cfq_put_queue(cic->cfqq[ASYNC]);
2220                         cic->cfqq[ASYNC] = NULL;
2221                 }
2222                 if (cic->cfqq[SYNC]) {
2223                         cfq_put_queue(cic->cfqq[SYNC]);
2224                         cic->cfqq[SYNC] = NULL;
2225                 }
2226                 cic->key = NULL;
2227                 list_del_init(&cic->queue_list);
2228         }
2229
2230         spin_unlock_irq(q->queue_lock);
2231         spin_unlock(&cfq_exit_lock);
2232
2233         cfq_shutdown_timer_wq(cfqd);
2234
2235         mempool_destroy(cfqd->crq_pool);
2236         kfree(cfqd->crq_hash);
2237         kfree(cfqd->cfq_hash);
2238         kfree(cfqd);
2239 }
2240
2241 static int cfq_init_queue(request_queue_t *q, elevator_t *e)
2242 {
2243         struct cfq_data *cfqd;
2244         int i;
2245
2246         cfqd = kmalloc(sizeof(*cfqd), GFP_KERNEL);
2247         if (!cfqd)
2248                 return -ENOMEM;
2249
2250         memset(cfqd, 0, sizeof(*cfqd));
2251
2252         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2253                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
2254
2255         INIT_LIST_HEAD(&cfqd->busy_rr);
2256         INIT_LIST_HEAD(&cfqd->cur_rr);
2257         INIT_LIST_HEAD(&cfqd->idle_rr);
2258         INIT_LIST_HEAD(&cfqd->empty_list);
2259         INIT_LIST_HEAD(&cfqd->cic_list);
2260
2261         cfqd->crq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_MHASH_ENTRIES, GFP_KERNEL);
2262         if (!cfqd->crq_hash)
2263                 goto out_crqhash;
2264
2265         cfqd->cfq_hash = kmalloc(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL);
2266         if (!cfqd->cfq_hash)
2267                 goto out_cfqhash;
2268
2269         cfqd->crq_pool = mempool_create_slab_pool(BLKDEV_MIN_RQ, crq_pool);
2270         if (!cfqd->crq_pool)
2271                 goto out_crqpool;
2272
2273         for (i = 0; i < CFQ_MHASH_ENTRIES; i++)
2274                 INIT_HLIST_HEAD(&cfqd->crq_hash[i]);
2275         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
2276                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
2277
2278         e->elevator_data = cfqd;
2279
2280         cfqd->queue = q;
2281
2282         cfqd->max_queued = q->nr_requests / 4;
2283         q->nr_batching = cfq_queued;
2284
2285         init_timer(&cfqd->idle_slice_timer);
2286         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2287         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2288
2289         init_timer(&cfqd->idle_class_timer);
2290         cfqd->idle_class_timer.function = cfq_idle_class_timer;
2291         cfqd->idle_class_timer.data = (unsigned long) cfqd;
2292
2293         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue, q);
2294
2295         cfqd->cfq_queued = cfq_queued;
2296         cfqd->cfq_quantum = cfq_quantum;
2297         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2298         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2299         cfqd->cfq_back_max = cfq_back_max;
2300         cfqd->cfq_back_penalty = cfq_back_penalty;
2301         cfqd->cfq_slice[0] = cfq_slice_async;
2302         cfqd->cfq_slice[1] = cfq_slice_sync;
2303         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2304         cfqd->cfq_slice_idle = cfq_slice_idle;
2305
2306         return 0;
2307 out_crqpool:
2308         kfree(cfqd->cfq_hash);
2309 out_cfqhash:
2310         kfree(cfqd->crq_hash);
2311 out_crqhash:
2312         kfree(cfqd);
2313         return -ENOMEM;
2314 }
2315
2316 static void cfq_slab_kill(void)
2317 {
2318         if (crq_pool)
2319                 kmem_cache_destroy(crq_pool);
2320         if (cfq_pool)
2321                 kmem_cache_destroy(cfq_pool);
2322         if (cfq_ioc_pool)
2323                 kmem_cache_destroy(cfq_ioc_pool);
2324 }
2325
2326 static int __init cfq_slab_setup(void)
2327 {
2328         crq_pool = kmem_cache_create("crq_pool", sizeof(struct cfq_rq), 0, 0,
2329                                         NULL, NULL);
2330         if (!crq_pool)
2331                 goto fail;
2332
2333         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2334                                         NULL, NULL);
2335         if (!cfq_pool)
2336                 goto fail;
2337
2338         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2339                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2340         if (!cfq_ioc_pool)
2341                 goto fail;
2342
2343         return 0;
2344 fail:
2345         cfq_slab_kill();
2346         return -ENOMEM;
2347 }
2348
2349 /*
2350  * sysfs parts below -->
2351  */
2352
2353 static ssize_t
2354 cfq_var_show(unsigned int var, char *page)
2355 {
2356         return sprintf(page, "%d\n", var);
2357 }
2358
2359 static ssize_t
2360 cfq_var_store(unsigned int *var, const char *page, size_t count)
2361 {
2362         char *p = (char *) page;
2363
2364         *var = simple_strtoul(p, &p, 10);
2365         return count;
2366 }
2367
2368 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2369 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2370 {                                                                       \
2371         struct cfq_data *cfqd = e->elevator_data;                       \
2372         unsigned int __data = __VAR;                                    \
2373         if (__CONV)                                                     \
2374                 __data = jiffies_to_msecs(__data);                      \
2375         return cfq_var_show(__data, (page));                            \
2376 }
2377 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2378 SHOW_FUNCTION(cfq_queued_show, cfqd->cfq_queued, 0);
2379 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2380 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2381 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2382 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2383 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2384 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2385 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2386 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2387 #undef SHOW_FUNCTION
2388
2389 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2390 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2391 {                                                                       \
2392         struct cfq_data *cfqd = e->elevator_data;                       \
2393         unsigned int __data;                                            \
2394         int ret = cfq_var_store(&__data, (page), count);                \
2395         if (__data < (MIN))                                             \
2396                 __data = (MIN);                                         \
2397         else if (__data > (MAX))                                        \
2398                 __data = (MAX);                                         \
2399         if (__CONV)                                                     \
2400                 *(__PTR) = msecs_to_jiffies(__data);                    \
2401         else                                                            \
2402                 *(__PTR) = __data;                                      \
2403         return ret;                                                     \
2404 }
2405 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2406 STORE_FUNCTION(cfq_queued_store, &cfqd->cfq_queued, 1, UINT_MAX, 0);
2407 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2408 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2409 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2410 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2411 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2412 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2413 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2414 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2415 #undef STORE_FUNCTION
2416
2417 #define CFQ_ATTR(name) \
2418         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2419
2420 static struct elv_fs_entry cfq_attrs[] = {
2421         CFQ_ATTR(quantum),
2422         CFQ_ATTR(queued),
2423         CFQ_ATTR(fifo_expire_sync),
2424         CFQ_ATTR(fifo_expire_async),
2425         CFQ_ATTR(back_seek_max),
2426         CFQ_ATTR(back_seek_penalty),
2427         CFQ_ATTR(slice_sync),
2428         CFQ_ATTR(slice_async),
2429         CFQ_ATTR(slice_async_rq),
2430         CFQ_ATTR(slice_idle),
2431         __ATTR_NULL
2432 };
2433
2434 static struct elevator_type iosched_cfq = {
2435         .ops = {
2436                 .elevator_merge_fn =            cfq_merge,
2437                 .elevator_merged_fn =           cfq_merged_request,
2438                 .elevator_merge_req_fn =        cfq_merged_requests,
2439                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2440                 .elevator_add_req_fn =          cfq_insert_request,
2441                 .elevator_activate_req_fn =     cfq_activate_request,
2442                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2443                 .elevator_queue_empty_fn =      cfq_queue_empty,
2444                 .elevator_completed_req_fn =    cfq_completed_request,
2445                 .elevator_former_req_fn =       cfq_former_request,
2446                 .elevator_latter_req_fn =       cfq_latter_request,
2447                 .elevator_set_req_fn =          cfq_set_request,
2448                 .elevator_put_req_fn =          cfq_put_request,
2449                 .elevator_may_queue_fn =        cfq_may_queue,
2450                 .elevator_init_fn =             cfq_init_queue,
2451                 .elevator_exit_fn =             cfq_exit_queue,
2452                 .trim =                         cfq_trim,
2453         },
2454         .elevator_attrs =       cfq_attrs,
2455         .elevator_name =        "cfq",
2456         .elevator_owner =       THIS_MODULE,
2457 };
2458
2459 static int __init cfq_init(void)
2460 {
2461         int ret;
2462
2463         /*
2464          * could be 0 on HZ < 1000 setups
2465          */
2466         if (!cfq_slice_async)
2467                 cfq_slice_async = 1;
2468         if (!cfq_slice_idle)
2469                 cfq_slice_idle = 1;
2470
2471         if (cfq_slab_setup())
2472                 return -ENOMEM;
2473
2474         ret = elv_register(&iosched_cfq);
2475         if (ret)
2476                 cfq_slab_kill();
2477
2478         return ret;
2479 }
2480
2481 static void __exit cfq_exit(void)
2482 {
2483         DECLARE_COMPLETION(all_gone);
2484         elv_unregister(&iosched_cfq);
2485         ioc_gone = &all_gone;
2486         /* ioc_gone's update must be visible before reading ioc_count */
2487         smp_wmb();
2488         if (atomic_read(&ioc_count))
2489                 wait_for_completion(ioc_gone);
2490         synchronize_rcu();
2491         cfq_slab_kill();
2492 }
2493
2494 module_init(cfq_init);
2495 module_exit(cfq_exit);
2496
2497 MODULE_AUTHOR("Jens Axboe");
2498 MODULE_LICENSE("GPL");
2499 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");