block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/rbtree.h>
  13 #include <linux/ioprio.h>
  14 #include <linux/blktrace_api.h>
  15
  16 /*
  17  * tunables
  18  */
  19 /* max queue in one round of service */
  20 static const int cfq_quantum = 4;
  21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  22 /* maximum backwards seek, in KiB */
  23 static const int cfq_back_max = 16 * 1024;
  24 /* penalty of a backwards seek */
  25 static const int cfq_back_penalty = 2;
  26 static const int cfq_slice_sync = HZ / 10;
  27 static int cfq_slice_async = HZ / 25;
  28 static const int cfq_slice_async_rq = 2;
  29 static int cfq_slice_idle = HZ / 125;
  30
  31 /*
  32  * offset from end of service tree
  33  */
  34 #define CFQ_IDLE_DELAY          (HZ / 5)
  35
  36 /*
  37  * below this threshold, we consider thinktime immediate
  38  */
  39 #define CFQ_MIN_TT              (2)
  40
  41 #define CFQ_SLICE_SCALE         (5)
  42 #define CFQ_HW_QUEUE_MIN        (5)
  43
  44 #define RQ_CIC(rq)              \
  45         ((struct cfq_io_context *) (rq)->elevator_private)
  46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
  47
  48 static struct kmem_cache *cfq_pool;
  49 static struct kmem_cache *cfq_ioc_pool;
  50
  51 static DEFINE_PER_CPU(unsigned long, ioc_count);
  52 static struct completion *ioc_gone;
  53 static DEFINE_SPINLOCK(ioc_gone_lock);
  54
  55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  58
  59 #define ASYNC                   (0)
  60 #define SYNC                    (1)
  61
  62 #define sample_valid(samples)   ((samples) > 80)
  63
  64 /*
  65  * Most of our rbtree usage is for sorting with min extraction, so
  66  * if we cache the leftmost node we don't have to walk down the tree
  67  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  68  * move this into the elevator for the rq sorting as well.
  69  */
  70 struct cfq_rb_root {
  71         struct rb_root rb;
  72         struct rb_node *left;
  73 };
  74 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
  75
  76 /*
  77  * Per block device queue structure
  78  */
  79 struct cfq_data {
  80         struct request_queue *queue;
  81
  82         /*
  83          * rr list of queues with requests and the count of them
  84          */
  85         struct cfq_rb_root service_tree;
  86         unsigned int busy_queues;
  87
  88         int rq_in_driver;
  89         int sync_flight;
  90
  91         /*
  92          * queue-depth detection
  93          */
  94         int rq_queued;
  95         int hw_tag;
  96         int hw_tag_samples;
  97         int rq_in_driver_peak;
  98
  99         /*
 100          * idle window management
 101          */
 102         struct timer_list idle_slice_timer;
 103         struct work_struct unplug_work;
 104
 105         struct cfq_queue *active_queue;
 106         struct cfq_io_context *active_cic;
 107
 108         /*
 109          * async queue for each priority case
 110          */
 111         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 112         struct cfq_queue *async_idle_cfqq;
 113
 114         sector_t last_position;
 115         unsigned long last_end_request;
 116
 117         /*
 118          * tunables, see top of file
 119          */
 120         unsigned int cfq_quantum;
 121         unsigned int cfq_fifo_expire[2];
 122         unsigned int cfq_back_penalty;
 123         unsigned int cfq_back_max;
 124         unsigned int cfq_slice[2];
 125         unsigned int cfq_slice_async_rq;
 126         unsigned int cfq_slice_idle;
 127
 128         struct list_head cic_list;
 129 };
 130
 131 /*
 132  * Per process-grouping structure
 133  */
 134 struct cfq_queue {
 135         /* reference count */
 136         atomic_t ref;
 137         /* various state flags, see below */
 138         unsigned int flags;
 139         /* parent cfq_data */
 140         struct cfq_data *cfqd;
 141         /* service_tree member */
 142         struct rb_node rb_node;
 143         /* service_tree key */
 144         unsigned long rb_key;
 145         /* sorted list of pending requests */
 146         struct rb_root sort_list;
 147         /* if fifo isn't expired, next request to serve */
 148         struct request *next_rq;
 149         /* requests queued in sort_list */
 150         int queued[2];
 151         /* currently allocated requests */
 152         int allocated[2];
 153         /* fifo list of requests in sort_list */
 154         struct list_head fifo;
 155
 156         unsigned long slice_end;
 157         long slice_resid;
 158
 159         /* pending metadata requests */
 160         int meta_pending;
 161         /* number of requests that are on the dispatch list or inside driver */
 162         int dispatched;
 163
 164         /* io prio of this group */
 165         unsigned short ioprio, org_ioprio;
 166         unsigned short ioprio_class, org_ioprio_class;
 167
 168         pid_t pid;
 169 };
 170
 171 enum cfqq_state_flags {
 172         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 173         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 174         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 175         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 176         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
 177         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 178         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 179         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 180         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
 181         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 182         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 183 };
 184
 185 #define CFQ_CFQQ_FNS(name)                                              \
 186 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 187 {                                                                       \
 188         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 189 }                                                                       \
 190 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 191 {                                                                       \
 192         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 193 }                                                                       \
 194 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 195 {                                                                       \
 196         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 197 }
 198
 199 CFQ_CFQQ_FNS(on_rr);
 200 CFQ_CFQQ_FNS(wait_request);
 201 CFQ_CFQQ_FNS(must_alloc);
 202 CFQ_CFQQ_FNS(must_alloc_slice);
 203 CFQ_CFQQ_FNS(must_dispatch);
 204 CFQ_CFQQ_FNS(fifo_expire);
 205 CFQ_CFQQ_FNS(idle_window);
 206 CFQ_CFQQ_FNS(prio_changed);
 207 CFQ_CFQQ_FNS(queue_new);
 208 CFQ_CFQQ_FNS(slice_new);
 209 CFQ_CFQQ_FNS(sync);
 210 #undef CFQ_CFQQ_FNS
 211
 212 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 213         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 214 #define cfq_log(cfqd, fmt, args...)     \
 215         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 216
 217 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 218 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
 219                                        struct io_context *, gfp_t);
 220 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 221                                                 struct io_context *);
 222
 223 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 224                                             int is_sync)
 225 {
 226         return cic->cfqq[!!is_sync];
 227 }
 228
 229 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 230                                 struct cfq_queue *cfqq, int is_sync)
 231 {
 232         cic->cfqq[!!is_sync] = cfqq;
 233 }
 234
 235 /*
 236  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 237  * set (in which case it could also be direct WRITE).
 238  */
 239 static inline int cfq_bio_sync(struct bio *bio)
 240 {
 241         if (bio_data_dir(bio) == READ || bio_sync(bio))
 242                 return 1;
 243
 244         return 0;
 245 }
 246
 247 /*
 248  * scheduler run of queue, if there are requests pending and no one in the
 249  * driver that will restart queueing
 250  */
 251 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 252 {
 253         if (cfqd->busy_queues) {
 254                 cfq_log(cfqd, "schedule dispatch");
 255                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 256         }
 257 }
 258
 259 static int cfq_queue_empty(struct request_queue *q)
 260 {
 261         struct cfq_data *cfqd = q->elevator->elevator_data;
 262
 263         return !cfqd->busy_queues;
 264 }
 265
 266 /*
 267  * Scale schedule slice based on io priority. Use the sync time slice only
 268  * if a queue is marked sync and has sync io queued. A sync queue with async
 269  * io only, should not get full sync slice length.
 270  */
 271 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
 272                                  unsigned short prio)
 273 {
 274         const int base_slice = cfqd->cfq_slice[sync];
 275
 276         WARN_ON(prio >= IOPRIO_BE_NR);
 277
 278         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 279 }
 280
 281 static inline int
 282 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 283 {
 284         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 285 }
 286
 287 static inline void
 288 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 289 {
 290         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 291         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 292 }
 293
 294 /*
 295  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 296  * isn't valid until the first request from the dispatch is activated
 297  * and the slice time set.
 298  */
 299 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 300 {
 301         if (cfq_cfqq_slice_new(cfqq))
 302                 return 0;
 303         if (time_before(jiffies, cfqq->slice_end))
 304                 return 0;
 305
 306         return 1;
 307 }
 308
 309 /*
 310  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 311  * We choose the request that is closest to the head right now. Distance
 312  * behind the head is penalized and only allowed to a certain extent.
 313  */
 314 static struct request *
 315 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 316 {
 317         sector_t last, s1, s2, d1 = 0, d2 = 0;
 318         unsigned long back_max;
 319 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 320 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 321         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 322
 323         if (rq1 == NULL || rq1 == rq2)
 324                 return rq2;
 325         if (rq2 == NULL)
 326                 return rq1;
 327
 328         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 329                 return rq1;
 330         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 331                 return rq2;
 332         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 333                 return rq1;
 334         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 335                 return rq2;
 336
 337         s1 = rq1->sector;
 338         s2 = rq2->sector;
 339
 340         last = cfqd->last_position;
 341
 342         /*
 343          * by definition, 1KiB is 2 sectors
 344          */
 345         back_max = cfqd->cfq_back_max * 2;
 346
 347         /*
 348          * Strict one way elevator _except_ in the case where we allow
 349          * short backward seeks which are biased as twice the cost of a
 350          * similar forward seek.
 351          */
 352         if (s1 >= last)
 353                 d1 = s1 - last;
 354         else if (s1 + back_max >= last)
 355                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 356         else
 357                 wrap |= CFQ_RQ1_WRAP;
 358
 359         if (s2 >= last)
 360                 d2 = s2 - last;
 361         else if (s2 + back_max >= last)
 362                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 363         else
 364                 wrap |= CFQ_RQ2_WRAP;
 365
 366         /* Found required data */
 367
 368         /*
 369          * By doing switch() on the bit mask "wrap" we avoid having to
 370          * check two variables for all permutations: --> faster!
 371          */
 372         switch (wrap) {
 373         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 374                 if (d1 < d2)
 375                         return rq1;
 376                 else if (d2 < d1)
 377                         return rq2;
 378                 else {
 379                         if (s1 >= s2)
 380                                 return rq1;
 381                         else
 382                                 return rq2;
 383                 }
 384
 385         case CFQ_RQ2_WRAP:
 386                 return rq1;
 387         case CFQ_RQ1_WRAP:
 388                 return rq2;
 389         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 390         default:
 391                 /*
 392                  * Since both rqs are wrapped,
 393                  * start with the one that's further behind head
 394                  * (--> only *one* back seek required),
 395                  * since back seek takes more time than forward.
 396                  */
 397                 if (s1 <= s2)
 398                         return rq1;
 399                 else
 400                         return rq2;
 401         }
 402 }
 403
 404 /*
 405  * The below is leftmost cache rbtree addon
 406  */
 407 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 408 {
 409         if (!root->left)
 410                 root->left = rb_first(&root->rb);
 411
 412         if (root->left)
 413                 return rb_entry(root->left, struct cfq_queue, rb_node);
 414
 415         return NULL;
 416 }
 417
 418 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 419 {
 420         if (root->left == n)
 421                 root->left = NULL;
 422
 423         rb_erase(n, &root->rb);
 424         RB_CLEAR_NODE(n);
 425 }
 426
 427 /*
 428  * would be nice to take fifo expire time into account as well
 429  */
 430 static struct request *
 431 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 432                   struct request *last)
 433 {
 434         struct rb_node *rbnext = rb_next(&last->rb_node);
 435         struct rb_node *rbprev = rb_prev(&last->rb_node);
 436         struct request *next = NULL, *prev = NULL;
 437
 438         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 439
 440         if (rbprev)
 441                 prev = rb_entry_rq(rbprev);
 442
 443         if (rbnext)
 444                 next = rb_entry_rq(rbnext);
 445         else {
 446                 rbnext = rb_first(&cfqq->sort_list);
 447                 if (rbnext && rbnext != &last->rb_node)
 448                         next = rb_entry_rq(rbnext);
 449         }
 450
 451         return cfq_choose_req(cfqd, next, prev);
 452 }
 453
 454 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 455                                       struct cfq_queue *cfqq)
 456 {
 457         /*
 458          * just an approximation, should be ok.
 459          */
 460         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 461                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 462 }
 463
 464 /*
 465  * The cfqd->service_tree holds all pending cfq_queue's that have
 466  * requests waiting to be processed. It is sorted in the order that
 467  * we will service the queues.
 468  */
 469 static void cfq_service_tree_add(struct cfq_data *cfqd,
 470                                     struct cfq_queue *cfqq, int add_front)
 471 {
 472         struct rb_node **p, *parent;
 473         struct cfq_queue *__cfqq;
 474         unsigned long rb_key;
 475         int left;
 476
 477         if (cfq_class_idle(cfqq)) {
 478                 rb_key = CFQ_IDLE_DELAY;
 479                 parent = rb_last(&cfqd->service_tree.rb);
 480                 if (parent && parent != &cfqq->rb_node) {
 481                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 482                         rb_key += __cfqq->rb_key;
 483                 } else
 484                         rb_key += jiffies;
 485         } else if (!add_front) {
 486                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 487                 rb_key += cfqq->slice_resid;
 488                 cfqq->slice_resid = 0;
 489         } else
 490                 rb_key = 0;
 491
 492         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 493                 /*
 494                  * same position, nothing more to do
 495                  */
 496                 if (rb_key == cfqq->rb_key)
 497                         return;
 498
 499                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 500         }
 501
 502         left = 1;
 503         parent = NULL;
 504         p = &cfqd->service_tree.rb.rb_node;
 505         while (*p) {
 506                 struct rb_node **n;
 507
 508                 parent = *p;
 509                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 510
 511                 /*
 512                  * sort RT queues first, we always want to give
 513                  * preference to them. IDLE queues goes to the back.
 514                  * after that, sort on the next service time.
 515                  */
 516                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 517                         n = &(*p)->rb_left;
 518                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 519                         n = &(*p)->rb_right;
 520                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 521                         n = &(*p)->rb_left;
 522                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 523                         n = &(*p)->rb_right;
 524                 else if (rb_key < __cfqq->rb_key)
 525                         n = &(*p)->rb_left;
 526                 else
 527                         n = &(*p)->rb_right;
 528
 529                 if (n == &(*p)->rb_right)
 530                         left = 0;
 531
 532                 p = n;
 533         }
 534
 535         if (left)
 536                 cfqd->service_tree.left = &cfqq->rb_node;
 537
 538         cfqq->rb_key = rb_key;
 539         rb_link_node(&cfqq->rb_node, parent, p);
 540         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 541 }
 542
 543 /*
 544  * Update cfqq's position in the service tree.
 545  */
 546 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 547 {
 548         /*
 549          * Resorting requires the cfqq to be on the RR list already.
 550          */
 551         if (cfq_cfqq_on_rr(cfqq))
 552                 cfq_service_tree_add(cfqd, cfqq, 0);
 553 }
 554
 555 /*
 556  * add to busy list of queues for service, trying to be fair in ordering
 557  * the pending list according to last request service
 558  */
 559 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 560 {
 561         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
 562         BUG_ON(cfq_cfqq_on_rr(cfqq));
 563         cfq_mark_cfqq_on_rr(cfqq);
 564         cfqd->busy_queues++;
 565
 566         cfq_resort_rr_list(cfqd, cfqq);
 567 }
 568
 569 /*
 570  * Called when the cfqq no longer has requests pending, remove it from
 571  * the service tree.
 572  */
 573 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 574 {
 575         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
 576         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 577         cfq_clear_cfqq_on_rr(cfqq);
 578
 579         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 580                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 581
 582         BUG_ON(!cfqd->busy_queues);
 583         cfqd->busy_queues--;
 584 }
 585
 586 /*
 587  * rb tree support functions
 588  */
 589 static void cfq_del_rq_rb(struct request *rq)
 590 {
 591         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 592         struct cfq_data *cfqd = cfqq->cfqd;
 593         const int sync = rq_is_sync(rq);
 594
 595         BUG_ON(!cfqq->queued[sync]);
 596         cfqq->queued[sync]--;
 597
 598         elv_rb_del(&cfqq->sort_list, rq);
 599
 600         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 601                 cfq_del_cfqq_rr(cfqd, cfqq);
 602 }
 603
 604 static void cfq_add_rq_rb(struct request *rq)
 605 {
 606         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 607         struct cfq_data *cfqd = cfqq->cfqd;
 608         struct request *__alias;
 609
 610         cfqq->queued[rq_is_sync(rq)]++;
 611
 612         /*
 613          * looks a little odd, but the first insert might return an alias.
 614          * if that happens, put the alias on the dispatch list
 615          */
 616         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 617                 cfq_dispatch_insert(cfqd->queue, __alias);
 618
 619         if (!cfq_cfqq_on_rr(cfqq))
 620                 cfq_add_cfqq_rr(cfqd, cfqq);
 621
 622         /*
 623          * check if this request is a better next-serve candidate
 624          */
 625         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 626         BUG_ON(!cfqq->next_rq);
 627 }
 628
 629 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 630 {
 631         elv_rb_del(&cfqq->sort_list, rq);
 632         cfqq->queued[rq_is_sync(rq)]--;
 633         cfq_add_rq_rb(rq);
 634 }
 635
 636 static struct request *
 637 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 638 {
 639         struct task_struct *tsk = current;
 640         struct cfq_io_context *cic;
 641         struct cfq_queue *cfqq;
 642
 643         cic = cfq_cic_lookup(cfqd, tsk->io_context);
 644         if (!cic)
 645                 return NULL;
 646
 647         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 648         if (cfqq) {
 649                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 650
 651                 return elv_rb_find(&cfqq->sort_list, sector);
 652         }
 653
 654         return NULL;
 655 }
 656
 657 static void cfq_activate_request(struct request_queue *q, struct request *rq)
 658 {
 659         struct cfq_data *cfqd = q->elevator->elevator_data;
 660
 661         cfqd->rq_in_driver++;
 662         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
 663                                                 cfqd->rq_in_driver);
 664
 665         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 666 }
 667
 668 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 669 {
 670         struct cfq_data *cfqd = q->elevator->elevator_data;
 671
 672         WARN_ON(!cfqd->rq_in_driver);
 673         cfqd->rq_in_driver--;
 674         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
 675                                                 cfqd->rq_in_driver);
 676 }
 677
 678 static void cfq_remove_request(struct request *rq)
 679 {
 680         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 681
 682         if (cfqq->next_rq == rq)
 683                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 684
 685         list_del_init(&rq->queuelist);
 686         cfq_del_rq_rb(rq);
 687
 688         cfqq->cfqd->rq_queued--;
 689         if (rq_is_meta(rq)) {
 690                 WARN_ON(!cfqq->meta_pending);
 691                 cfqq->meta_pending--;
 692         }
 693 }
 694
 695 static int cfq_merge(struct request_queue *q, struct request **req,
 696                      struct bio *bio)
 697 {
 698         struct cfq_data *cfqd = q->elevator->elevator_data;
 699         struct request *__rq;
 700
 701         __rq = cfq_find_rq_fmerge(cfqd, bio);
 702         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 703                 *req = __rq;
 704                 return ELEVATOR_FRONT_MERGE;
 705         }
 706
 707         return ELEVATOR_NO_MERGE;
 708 }
 709
 710 static void cfq_merged_request(struct request_queue *q, struct request *req,
 711                                int type)
 712 {
 713         if (type == ELEVATOR_FRONT_MERGE) {
 714                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 715
 716                 cfq_reposition_rq_rb(cfqq, req);
 717         }
 718 }
 719
 720 static void
 721 cfq_merged_requests(struct request_queue *q, struct request *rq,
 722                     struct request *next)
 723 {
 724         /*
 725          * reposition in fifo if next is older than rq
 726          */
 727         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
 728             time_before(next->start_time, rq->start_time))
 729                 list_move(&rq->queuelist, &next->queuelist);
 730
 731         cfq_remove_request(next);
 732 }
 733
 734 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
 735                            struct bio *bio)
 736 {
 737         struct cfq_data *cfqd = q->elevator->elevator_data;
 738         struct cfq_io_context *cic;
 739         struct cfq_queue *cfqq;
 740
 741         /*
 742          * Disallow merge of a sync bio into an async request.
 743          */
 744         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 745                 return 0;
 746
 747         /*
 748          * Lookup the cfqq that this bio will be queued with. Allow
 749          * merge only if rq is queued there.
 750          */
 751         cic = cfq_cic_lookup(cfqd, current->io_context);
 752         if (!cic)
 753                 return 0;
 754
 755         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 756         if (cfqq == RQ_CFQQ(rq))
 757                 return 1;
 758
 759         return 0;
 760 }
 761
 762 static void __cfq_set_active_queue(struct cfq_data *cfqd,
 763                                    struct cfq_queue *cfqq)
 764 {
 765         if (cfqq) {
 766                 cfq_log_cfqq(cfqd, cfqq, "set_active");
 767                 cfqq->slice_end = 0;
 768                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 769                 cfq_clear_cfqq_fifo_expire(cfqq);
 770                 cfq_mark_cfqq_slice_new(cfqq);
 771                 cfq_clear_cfqq_queue_new(cfqq);
 772         }
 773
 774         cfqd->active_queue = cfqq;
 775 }
 776
 777 /*
 778  * current cfqq expired its slice (or was too idle), select new one
 779  */
 780 static void
 781 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 782                     int timed_out)
 783 {
 784         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
 785
 786         if (cfq_cfqq_wait_request(cfqq))
 787                 del_timer(&cfqd->idle_slice_timer);
 788
 789         cfq_clear_cfqq_must_dispatch(cfqq);
 790         cfq_clear_cfqq_wait_request(cfqq);
 791
 792         /*
 793          * store what was left of this slice, if the queue idled/timed out
 794          */
 795         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
 796                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 797                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
 798         }
 799
 800         cfq_resort_rr_list(cfqd, cfqq);
 801
 802         if (cfqq == cfqd->active_queue)
 803                 cfqd->active_queue = NULL;
 804
 805         if (cfqd->active_cic) {
 806                 put_io_context(cfqd->active_cic->ioc);
 807                 cfqd->active_cic = NULL;
 808         }
 809 }
 810
 811 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
 812 {
 813         struct cfq_queue *cfqq = cfqd->active_queue;
 814
 815         if (cfqq)
 816                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 817 }
 818
 819 /*
 820  * Get next queue for service. Unless we have a queue preemption,
 821  * we'll simply select the first cfqq in the service tree.
 822  */
 823 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 824 {
 825         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 826                 return NULL;
 827
 828         return cfq_rb_first(&cfqd->service_tree);
 829 }
 830
 831 /*
 832  * Get and set a new active queue for service.
 833  */
 834 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 835 {
 836         struct cfq_queue *cfqq;
 837
 838         cfqq = cfq_get_next_queue(cfqd);
 839         __cfq_set_active_queue(cfqd, cfqq);
 840         return cfqq;
 841 }
 842
 843 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 844                                           struct request *rq)
 845 {
 846         if (rq->sector >= cfqd->last_position)
 847                 return rq->sector - cfqd->last_position;
 848         else
 849                 return cfqd->last_position - rq->sector;
 850 }
 851
 852 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 853 {
 854         struct cfq_io_context *cic = cfqd->active_cic;
 855
 856         if (!sample_valid(cic->seek_samples))
 857                 return 0;
 858
 859         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
 860 }
 861
 862 static int cfq_close_cooperator(struct cfq_data *cfq_data,
 863                                 struct cfq_queue *cfqq)
 864 {
 865         /*
 866          * We should notice if some of the queues are cooperating, eg
 867          * working closely on the same area of the disk. In that case,
 868          * we can group them together and don't waste time idling.
 869          */
 870         return 0;
 871 }
 872
 873 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
 874
 875 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 876 {
 877         struct cfq_queue *cfqq = cfqd->active_queue;
 878         struct cfq_io_context *cic;
 879         unsigned long sl;
 880
 881         /*
 882          * SSD device without seek penalty, disable idling. But only do so
 883          * for devices that support queuing, otherwise we still have a problem
 884          * with sync vs async workloads.
 885          */
 886         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
 887                 return;
 888
 889         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 890         WARN_ON(cfq_cfqq_slice_new(cfqq));
 891
 892         /*
 893          * idle is disabled, either manually or by past process history
 894          */
 895         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
 896                 return;
 897
 898         /*
 899          * still requests with the driver, don't idle
 900          */
 901         if (cfqd->rq_in_driver)
 902                 return;
 903
 904         /*
 905          * task has exited, don't wait
 906          */
 907         cic = cfqd->active_cic;
 908         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
 909                 return;
 910
 911         /*
 912          * See if this prio level has a good candidate
 913          */
 914         if (cfq_close_cooperator(cfqd, cfqq) &&
 915             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
 916                 return;
 917
 918         cfq_mark_cfqq_must_dispatch(cfqq);
 919         cfq_mark_cfqq_wait_request(cfqq);
 920
 921         /*
 922          * we don't want to idle for seeks, but we do want to allow
 923          * fair distribution of slice time for a process doing back-to-back
 924          * seeks. so allow a little bit of time for him to submit a new rq
 925          */
 926         sl = cfqd->cfq_slice_idle;
 927         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 928                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
 929
 930         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 931         cfq_log(cfqd, "arm_idle: %lu", sl);
 932 }
 933
 934 /*
 935  * Move request from internal lists to the request queue dispatch list.
 936  */
 937 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
 938 {
 939         struct cfq_data *cfqd = q->elevator->elevator_data;
 940         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 941
 942         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
 943
 944         cfq_remove_request(rq);
 945         cfqq->dispatched++;
 946         elv_dispatch_sort(q, rq);
 947
 948         if (cfq_cfqq_sync(cfqq))
 949                 cfqd->sync_flight++;
 950 }
 951
 952 /*
 953  * return expired entry, or NULL to just start from scratch in rbtree
 954  */
 955 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 956 {
 957         struct cfq_data *cfqd = cfqq->cfqd;
 958         struct request *rq;
 959         int fifo;
 960
 961         if (cfq_cfqq_fifo_expire(cfqq))
 962                 return NULL;
 963
 964         cfq_mark_cfqq_fifo_expire(cfqq);
 965
 966         if (list_empty(&cfqq->fifo))
 967                 return NULL;
 968
 969         fifo = cfq_cfqq_sync(cfqq);
 970         rq = rq_entry_fifo(cfqq->fifo.next);
 971
 972         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
 973                 rq = NULL;
 974
 975         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
 976         return rq;
 977 }
 978
 979 static inline int
 980 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 981 {
 982         const int base_rq = cfqd->cfq_slice_async_rq;
 983
 984         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 985
 986         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 987 }
 988
 989 /*
 990  * Select a queue for service. If we have a current active queue,
 991  * check whether to continue servicing it, or retrieve and set a new one.
 992  */
 993 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
 994 {
 995         struct cfq_queue *cfqq;
 996
 997         cfqq = cfqd->active_queue;
 998         if (!cfqq)
 999                 goto new_queue;
1000
1001         /*
1002          * The active queue has run out of time, expire it and select new.
1003          */
1004         if (cfq_slice_used(cfqq))
1005                 goto expire;
1006
1007         /*
1008          * The active queue has requests and isn't expired, allow it to
1009          * dispatch.
1010          */
1011         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1012                 goto keep_queue;
1013
1014         /*
1015          * No requests pending. If the active queue still has requests in
1016          * flight or is idling for a new request, allow either of these
1017          * conditions to happen (or time out) before selecting a new queue.
1018          */
1019         if (timer_pending(&cfqd->idle_slice_timer) ||
1020             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1021                 cfqq = NULL;
1022                 goto keep_queue;
1023         }
1024
1025 expire:
1026         cfq_slice_expired(cfqd, 0);
1027 new_queue:
1028         cfqq = cfq_set_active_queue(cfqd);
1029 keep_queue:
1030         return cfqq;
1031 }
1032
1033 /*
1034  * Dispatch some requests from cfqq, moving them to the request queue
1035  * dispatch list.
1036  */
1037 static int
1038 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1039                         int max_dispatch)
1040 {
1041         int dispatched = 0;
1042
1043         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1044
1045         do {
1046                 struct request *rq;
1047
1048                 /*
1049                  * follow expired path, else get first next available
1050                  */
1051                 rq = cfq_check_fifo(cfqq);
1052                 if (rq == NULL)
1053                         rq = cfqq->next_rq;
1054
1055                 /*
1056                  * finally, insert request into driver dispatch list
1057                  */
1058                 cfq_dispatch_insert(cfqd->queue, rq);
1059
1060                 dispatched++;
1061
1062                 if (!cfqd->active_cic) {
1063                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1064                         cfqd->active_cic = RQ_CIC(rq);
1065                 }
1066
1067                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1068                         break;
1069
1070         } while (dispatched < max_dispatch);
1071
1072         /*
1073          * expire an async queue immediately if it has used up its slice. idle
1074          * queue always expire after 1 dispatch round.
1075          */
1076         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1077             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1078             cfq_class_idle(cfqq))) {
1079                 cfqq->slice_end = jiffies + 1;
1080                 cfq_slice_expired(cfqd, 0);
1081         }
1082
1083         return dispatched;
1084 }
1085
1086 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1087 {
1088         int dispatched = 0;
1089
1090         while (cfqq->next_rq) {
1091                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1092                 dispatched++;
1093         }
1094
1095         BUG_ON(!list_empty(&cfqq->fifo));
1096         return dispatched;
1097 }
1098
1099 /*
1100  * Drain our current requests. Used for barriers and when switching
1101  * io schedulers on-the-fly.
1102  */
1103 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1104 {
1105         struct cfq_queue *cfqq;
1106         int dispatched = 0;
1107
1108         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1109                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1110
1111         cfq_slice_expired(cfqd, 0);
1112
1113         BUG_ON(cfqd->busy_queues);
1114
1115         cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1116         return dispatched;
1117 }
1118
1119 static int cfq_dispatch_requests(struct request_queue *q, int force)
1120 {
1121         struct cfq_data *cfqd = q->elevator->elevator_data;
1122         struct cfq_queue *cfqq;
1123         int dispatched;
1124
1125         if (!cfqd->busy_queues)
1126                 return 0;
1127
1128         if (unlikely(force))
1129                 return cfq_forced_dispatch(cfqd);
1130
1131         dispatched = 0;
1132         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1133                 int max_dispatch;
1134
1135                 max_dispatch = cfqd->cfq_quantum;
1136                 if (cfq_class_idle(cfqq))
1137                         max_dispatch = 1;
1138
1139                 if (cfqq->dispatched >= max_dispatch) {
1140                         if (cfqd->busy_queues > 1)
1141                                 break;
1142                         if (cfqq->dispatched >= 4 * max_dispatch)
1143                                 break;
1144                 }
1145
1146                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1147                         break;
1148
1149                 cfq_clear_cfqq_must_dispatch(cfqq);
1150                 cfq_clear_cfqq_wait_request(cfqq);
1151                 del_timer(&cfqd->idle_slice_timer);
1152
1153                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1154         }
1155
1156         cfq_log(cfqd, "dispatched=%d", dispatched);
1157         return dispatched;
1158 }
1159
1160 /*
1161  * task holds one reference to the queue, dropped when task exits. each rq
1162  * in-flight on this queue also holds a reference, dropped when rq is freed.
1163  *
1164  * queue lock must be held here.
1165  */
1166 static void cfq_put_queue(struct cfq_queue *cfqq)
1167 {
1168         struct cfq_data *cfqd = cfqq->cfqd;
1169
1170         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1171
1172         if (!atomic_dec_and_test(&cfqq->ref))
1173                 return;
1174
1175         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1176         BUG_ON(rb_first(&cfqq->sort_list));
1177         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1178         BUG_ON(cfq_cfqq_on_rr(cfqq));
1179
1180         if (unlikely(cfqd->active_queue == cfqq)) {
1181                 __cfq_slice_expired(cfqd, cfqq, 0);
1182                 cfq_schedule_dispatch(cfqd);
1183         }
1184
1185         kmem_cache_free(cfq_pool, cfqq);
1186 }
1187
1188 /*
1189  * Must always be called with the rcu_read_lock() held
1190  */
1191 static void
1192 __call_for_each_cic(struct io_context *ioc,
1193                     void (*func)(struct io_context *, struct cfq_io_context *))
1194 {
1195         struct cfq_io_context *cic;
1196         struct hlist_node *n;
1197
1198         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1199                 func(ioc, cic);
1200 }
1201
1202 /*
1203  * Call func for each cic attached to this ioc.
1204  */
1205 static void
1206 call_for_each_cic(struct io_context *ioc,
1207                   void (*func)(struct io_context *, struct cfq_io_context *))
1208 {
1209         rcu_read_lock();
1210         __call_for_each_cic(ioc, func);
1211         rcu_read_unlock();
1212 }
1213
1214 static void cfq_cic_free_rcu(struct rcu_head *head)
1215 {
1216         struct cfq_io_context *cic;
1217
1218         cic = container_of(head, struct cfq_io_context, rcu_head);
1219
1220         kmem_cache_free(cfq_ioc_pool, cic);
1221         elv_ioc_count_dec(ioc_count);
1222
1223         if (ioc_gone) {
1224                 /*
1225                  * CFQ scheduler is exiting, grab exit lock and check
1226                  * the pending io context count. If it hits zero,
1227                  * complete ioc_gone and set it back to NULL
1228                  */
1229                 spin_lock(&ioc_gone_lock);
1230                 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1231                         complete(ioc_gone);
1232                         ioc_gone = NULL;
1233                 }
1234                 spin_unlock(&ioc_gone_lock);
1235         }
1236 }
1237
1238 static void cfq_cic_free(struct cfq_io_context *cic)
1239 {
1240         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1241 }
1242
1243 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1244 {
1245         unsigned long flags;
1246
1247         BUG_ON(!cic->dead_key);
1248
1249         spin_lock_irqsave(&ioc->lock, flags);
1250         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1251         hlist_del_rcu(&cic->cic_list);
1252         spin_unlock_irqrestore(&ioc->lock, flags);
1253
1254         cfq_cic_free(cic);
1255 }
1256
1257 /*
1258  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1259  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1260  * and ->trim() which is called with the task lock held
1261  */
1262 static void cfq_free_io_context(struct io_context *ioc)
1263 {
1264         /*
1265          * ioc->refcount is zero here, or we are called from elv_unregister(),
1266          * so no more cic's are allowed to be linked into this ioc.  So it
1267          * should be ok to iterate over the known list, we will see all cic's
1268          * since no new ones are added.
1269          */
1270         __call_for_each_cic(ioc, cic_free_func);
1271 }
1272
1273 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1274 {
1275         if (unlikely(cfqq == cfqd->active_queue)) {
1276                 __cfq_slice_expired(cfqd, cfqq, 0);
1277                 cfq_schedule_dispatch(cfqd);
1278         }
1279
1280         cfq_put_queue(cfqq);
1281 }
1282
1283 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1284                                          struct cfq_io_context *cic)
1285 {
1286         struct io_context *ioc = cic->ioc;
1287
1288         list_del_init(&cic->queue_list);
1289
1290         /*
1291          * Make sure key == NULL is seen for dead queues
1292          */
1293         smp_wmb();
1294         cic->dead_key = (unsigned long) cic->key;
1295         cic->key = NULL;
1296
1297         if (ioc->ioc_data == cic)
1298                 rcu_assign_pointer(ioc->ioc_data, NULL);
1299
1300         if (cic->cfqq[ASYNC]) {
1301                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1302                 cic->cfqq[ASYNC] = NULL;
1303         }
1304
1305         if (cic->cfqq[SYNC]) {
1306                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1307                 cic->cfqq[SYNC] = NULL;
1308         }
1309 }
1310
1311 static void cfq_exit_single_io_context(struct io_context *ioc,
1312                                        struct cfq_io_context *cic)
1313 {
1314         struct cfq_data *cfqd = cic->key;
1315
1316         if (cfqd) {
1317                 struct request_queue *q = cfqd->queue;
1318                 unsigned long flags;
1319
1320                 spin_lock_irqsave(q->queue_lock, flags);
1321                 __cfq_exit_single_io_context(cfqd, cic);
1322                 spin_unlock_irqrestore(q->queue_lock, flags);
1323         }
1324 }
1325
1326 /*
1327  * The process that ioc belongs to has exited, we need to clean up
1328  * and put the internal structures we have that belongs to that process.
1329  */
1330 static void cfq_exit_io_context(struct io_context *ioc)
1331 {
1332         call_for_each_cic(ioc, cfq_exit_single_io_context);
1333 }
1334
1335 static struct cfq_io_context *
1336 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1337 {
1338         struct cfq_io_context *cic;
1339
1340         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1341                                                         cfqd->queue->node);
1342         if (cic) {
1343                 cic->last_end_request = jiffies;
1344                 INIT_LIST_HEAD(&cic->queue_list);
1345                 INIT_HLIST_NODE(&cic->cic_list);
1346                 cic->dtor = cfq_free_io_context;
1347                 cic->exit = cfq_exit_io_context;
1348                 elv_ioc_count_inc(ioc_count);
1349         }
1350
1351         return cic;
1352 }
1353
1354 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1355 {
1356         struct task_struct *tsk = current;
1357         int ioprio_class;
1358
1359         if (!cfq_cfqq_prio_changed(cfqq))
1360                 return;
1361
1362         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1363         switch (ioprio_class) {
1364         default:
1365                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1366         case IOPRIO_CLASS_NONE:
1367                 /*
1368                  * no prio set, inherit CPU scheduling settings
1369                  */
1370                 cfqq->ioprio = task_nice_ioprio(tsk);
1371                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1372                 break;
1373         case IOPRIO_CLASS_RT:
1374                 cfqq->ioprio = task_ioprio(ioc);
1375                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1376                 break;
1377         case IOPRIO_CLASS_BE:
1378                 cfqq->ioprio = task_ioprio(ioc);
1379                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1380                 break;
1381         case IOPRIO_CLASS_IDLE:
1382                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1383                 cfqq->ioprio = 7;
1384                 cfq_clear_cfqq_idle_window(cfqq);
1385                 break;
1386         }
1387
1388         /*
1389          * keep track of original prio settings in case we have to temporarily
1390          * elevate the priority of this queue
1391          */
1392         cfqq->org_ioprio = cfqq->ioprio;
1393         cfqq->org_ioprio_class = cfqq->ioprio_class;
1394         cfq_clear_cfqq_prio_changed(cfqq);
1395 }
1396
1397 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1398 {
1399         struct cfq_data *cfqd = cic->key;
1400         struct cfq_queue *cfqq;
1401         unsigned long flags;
1402
1403         if (unlikely(!cfqd))
1404                 return;
1405
1406         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1407
1408         cfqq = cic->cfqq[ASYNC];
1409         if (cfqq) {
1410                 struct cfq_queue *new_cfqq;
1411                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1412                 if (new_cfqq) {
1413                         cic->cfqq[ASYNC] = new_cfqq;
1414                         cfq_put_queue(cfqq);
1415                 }
1416         }
1417
1418         cfqq = cic->cfqq[SYNC];
1419         if (cfqq)
1420                 cfq_mark_cfqq_prio_changed(cfqq);
1421
1422         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1423 }
1424
1425 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1426 {
1427         call_for_each_cic(ioc, changed_ioprio);
1428         ioc->ioprio_changed = 0;
1429 }
1430
1431 static struct cfq_queue *
1432 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1433                      struct io_context *ioc, gfp_t gfp_mask)
1434 {
1435         struct cfq_queue *cfqq, *new_cfqq = NULL;
1436         struct cfq_io_context *cic;
1437
1438 retry:
1439         cic = cfq_cic_lookup(cfqd, ioc);
1440         /* cic always exists here */
1441         cfqq = cic_to_cfqq(cic, is_sync);
1442
1443         if (!cfqq) {
1444                 if (new_cfqq) {
1445                         cfqq = new_cfqq;
1446                         new_cfqq = NULL;
1447                 } else if (gfp_mask & __GFP_WAIT) {
1448                         /*
1449                          * Inform the allocator of the fact that we will
1450                          * just repeat this allocation if it fails, to allow
1451                          * the allocator to do whatever it needs to attempt to
1452                          * free memory.
1453                          */
1454                         spin_unlock_irq(cfqd->queue->queue_lock);
1455                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1456                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1457                                         cfqd->queue->node);
1458                         spin_lock_irq(cfqd->queue->queue_lock);
1459                         goto retry;
1460                 } else {
1461                         cfqq = kmem_cache_alloc_node(cfq_pool,
1462                                         gfp_mask | __GFP_ZERO,
1463                                         cfqd->queue->node);
1464                         if (!cfqq)
1465                                 goto out;
1466                 }
1467
1468                 RB_CLEAR_NODE(&cfqq->rb_node);
1469                 INIT_LIST_HEAD(&cfqq->fifo);
1470
1471                 atomic_set(&cfqq->ref, 0);
1472                 cfqq->cfqd = cfqd;
1473
1474                 cfq_mark_cfqq_prio_changed(cfqq);
1475                 cfq_mark_cfqq_queue_new(cfqq);
1476
1477                 cfq_init_prio_data(cfqq, ioc);
1478
1479                 if (is_sync) {
1480                         if (!cfq_class_idle(cfqq))
1481                                 cfq_mark_cfqq_idle_window(cfqq);
1482                         cfq_mark_cfqq_sync(cfqq);
1483                 }
1484                 cfqq->pid = current->pid;
1485                 cfq_log_cfqq(cfqd, cfqq, "alloced");
1486         }
1487
1488         if (new_cfqq)
1489                 kmem_cache_free(cfq_pool, new_cfqq);
1490
1491 out:
1492         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1493         return cfqq;
1494 }
1495
1496 static struct cfq_queue **
1497 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1498 {
1499         switch (ioprio_class) {
1500         case IOPRIO_CLASS_RT:
1501                 return &cfqd->async_cfqq[0][ioprio];
1502         case IOPRIO_CLASS_BE:
1503                 return &cfqd->async_cfqq[1][ioprio];
1504         case IOPRIO_CLASS_IDLE:
1505                 return &cfqd->async_idle_cfqq;
1506         default:
1507                 BUG();
1508         }
1509 }
1510
1511 static struct cfq_queue *
1512 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1513               gfp_t gfp_mask)
1514 {
1515         const int ioprio = task_ioprio(ioc);
1516         const int ioprio_class = task_ioprio_class(ioc);
1517         struct cfq_queue **async_cfqq = NULL;
1518         struct cfq_queue *cfqq = NULL;
1519
1520         if (!is_sync) {
1521                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1522                 cfqq = *async_cfqq;
1523         }
1524
1525         if (!cfqq) {
1526                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1527                 if (!cfqq)
1528                         return NULL;
1529         }
1530
1531         /*
1532          * pin the queue now that it's allocated, scheduler exit will prune it
1533          */
1534         if (!is_sync && !(*async_cfqq)) {
1535                 atomic_inc(&cfqq->ref);
1536                 *async_cfqq = cfqq;
1537         }
1538
1539         atomic_inc(&cfqq->ref);
1540         return cfqq;
1541 }
1542
1543 /*
1544  * We drop cfq io contexts lazily, so we may find a dead one.
1545  */
1546 static void
1547 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1548                   struct cfq_io_context *cic)
1549 {
1550         unsigned long flags;
1551
1552         WARN_ON(!list_empty(&cic->queue_list));
1553
1554         spin_lock_irqsave(&ioc->lock, flags);
1555
1556         BUG_ON(ioc->ioc_data == cic);
1557
1558         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1559         hlist_del_rcu(&cic->cic_list);
1560         spin_unlock_irqrestore(&ioc->lock, flags);
1561
1562         cfq_cic_free(cic);
1563 }
1564
1565 static struct cfq_io_context *
1566 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1567 {
1568         struct cfq_io_context *cic;
1569         unsigned long flags;
1570         void *k;
1571
1572         if (unlikely(!ioc))
1573                 return NULL;
1574
1575         rcu_read_lock();
1576
1577         /*
1578          * we maintain a last-hit cache, to avoid browsing over the tree
1579          */
1580         cic = rcu_dereference(ioc->ioc_data);
1581         if (cic && cic->key == cfqd) {
1582                 rcu_read_unlock();
1583                 return cic;
1584         }
1585
1586         do {
1587                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1588                 rcu_read_unlock();
1589                 if (!cic)
1590                         break;
1591                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1592                 k = cic->key;
1593                 if (unlikely(!k)) {
1594                         cfq_drop_dead_cic(cfqd, ioc, cic);
1595                         rcu_read_lock();
1596                         continue;
1597                 }
1598
1599                 spin_lock_irqsave(&ioc->lock, flags);
1600                 rcu_assign_pointer(ioc->ioc_data, cic);
1601                 spin_unlock_irqrestore(&ioc->lock, flags);
1602                 break;
1603         } while (1);
1604
1605         return cic;
1606 }
1607
1608 /*
1609  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1610  * the process specific cfq io context when entered from the block layer.
1611  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1612  */
1613 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1614                         struct cfq_io_context *cic, gfp_t gfp_mask)
1615 {
1616         unsigned long flags;
1617         int ret;
1618
1619         ret = radix_tree_preload(gfp_mask);
1620         if (!ret) {
1621                 cic->ioc = ioc;
1622                 cic->key = cfqd;
1623
1624                 spin_lock_irqsave(&ioc->lock, flags);
1625                 ret = radix_tree_insert(&ioc->radix_root,
1626                                                 (unsigned long) cfqd, cic);
1627                 if (!ret)
1628                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1629                 spin_unlock_irqrestore(&ioc->lock, flags);
1630
1631                 radix_tree_preload_end();
1632
1633                 if (!ret) {
1634                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1635                         list_add(&cic->queue_list, &cfqd->cic_list);
1636                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1637                 }
1638         }
1639
1640         if (ret)
1641                 printk(KERN_ERR "cfq: cic link failed!\n");
1642
1643         return ret;
1644 }
1645
1646 /*
1647  * Setup general io context and cfq io context. There can be several cfq
1648  * io contexts per general io context, if this process is doing io to more
1649  * than one device managed by cfq.
1650  */
1651 static struct cfq_io_context *
1652 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1653 {
1654         struct io_context *ioc = NULL;
1655         struct cfq_io_context *cic;
1656
1657         might_sleep_if(gfp_mask & __GFP_WAIT);
1658
1659         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1660         if (!ioc)
1661                 return NULL;
1662
1663         cic = cfq_cic_lookup(cfqd, ioc);
1664         if (cic)
1665                 goto out;
1666
1667         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1668         if (cic == NULL)
1669                 goto err;
1670
1671         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1672                 goto err_free;
1673
1674 out:
1675         smp_read_barrier_depends();
1676         if (unlikely(ioc->ioprio_changed))
1677                 cfq_ioc_set_ioprio(ioc);
1678
1679         return cic;
1680 err_free:
1681         cfq_cic_free(cic);
1682 err:
1683         put_io_context(ioc);
1684         return NULL;
1685 }
1686
1687 static void
1688 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1689 {
1690         unsigned long elapsed = jiffies - cic->last_end_request;
1691         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1692
1693         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1694         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1695         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1696 }
1697
1698 static void
1699 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1700                        struct request *rq)
1701 {
1702         sector_t sdist;
1703         u64 total;
1704
1705         if (cic->last_request_pos < rq->sector)
1706                 sdist = rq->sector - cic->last_request_pos;
1707         else
1708                 sdist = cic->last_request_pos - rq->sector;
1709
1710         /*
1711          * Don't allow the seek distance to get too large from the
1712          * odd fragment, pagein, etc
1713          */
1714         if (cic->seek_samples <= 60) /* second&third seek */
1715                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1716         else
1717                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1718
1719         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1720         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1721         total = cic->seek_total + (cic->seek_samples/2);
1722         do_div(total, cic->seek_samples);
1723         cic->seek_mean = (sector_t)total;
1724 }
1725
1726 /*
1727  * Disable idle window if the process thinks too long or seeks so much that
1728  * it doesn't matter
1729  */
1730 static void
1731 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1732                        struct cfq_io_context *cic)
1733 {
1734         int old_idle, enable_idle;
1735
1736         /*
1737          * Don't idle for async or idle io prio class
1738          */
1739         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1740                 return;
1741
1742         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1743
1744         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1745             (cfqd->hw_tag && CIC_SEEKY(cic)))
1746                 enable_idle = 0;
1747         else if (sample_valid(cic->ttime_samples)) {
1748                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1749                         enable_idle = 0;
1750                 else
1751                         enable_idle = 1;
1752         }
1753
1754         if (old_idle != enable_idle) {
1755                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1756                 if (enable_idle)
1757                         cfq_mark_cfqq_idle_window(cfqq);
1758                 else
1759                         cfq_clear_cfqq_idle_window(cfqq);
1760         }
1761 }
1762
1763 /*
1764  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1765  * no or if we aren't sure, a 1 will cause a preempt.
1766  */
1767 static int
1768 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1769                    struct request *rq)
1770 {
1771         struct cfq_queue *cfqq;
1772
1773         cfqq = cfqd->active_queue;
1774         if (!cfqq)
1775                 return 0;
1776
1777         if (cfq_slice_used(cfqq))
1778                 return 1;
1779
1780         if (cfq_class_idle(new_cfqq))
1781                 return 0;
1782
1783         if (cfq_class_idle(cfqq))
1784                 return 1;
1785
1786         /*
1787          * if the new request is sync, but the currently running queue is
1788          * not, let the sync request have priority.
1789          */
1790         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1791                 return 1;
1792
1793         /*
1794          * So both queues are sync. Let the new request get disk time if
1795          * it's a metadata request and the current queue is doing regular IO.
1796          */
1797         if (rq_is_meta(rq) && !cfqq->meta_pending)
1798                 return 1;
1799
1800         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1801                 return 0;
1802
1803         /*
1804          * if this request is as-good as one we would expect from the
1805          * current cfqq, let it preempt
1806          */
1807         if (cfq_rq_close(cfqd, rq))
1808                 return 1;
1809
1810         return 0;
1811 }
1812
1813 /*
1814  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1815  * let it have half of its nominal slice.
1816  */
1817 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1818 {
1819         cfq_log_cfqq(cfqd, cfqq, "preempt");
1820         cfq_slice_expired(cfqd, 1);
1821
1822         /*
1823          * Put the new queue at the front of the of the current list,
1824          * so we know that it will be selected next.
1825          */
1826         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1827
1828         cfq_service_tree_add(cfqd, cfqq, 1);
1829
1830         cfqq->slice_end = 0;
1831         cfq_mark_cfqq_slice_new(cfqq);
1832 }
1833
1834 /*
1835  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1836  * something we should do about it
1837  */
1838 static void
1839 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1840                 struct request *rq)
1841 {
1842         struct cfq_io_context *cic = RQ_CIC(rq);
1843
1844         cfqd->rq_queued++;
1845         if (rq_is_meta(rq))
1846                 cfqq->meta_pending++;
1847
1848         cfq_update_io_thinktime(cfqd, cic);
1849         cfq_update_io_seektime(cfqd, cic, rq);
1850         cfq_update_idle_window(cfqd, cfqq, cic);
1851
1852         cic->last_request_pos = rq->sector + rq->nr_sectors;
1853
1854         if (cfqq == cfqd->active_queue) {
1855                 /*
1856                  * if we are waiting for a request for this queue, let it rip
1857                  * immediately and flag that we must not expire this queue
1858                  * just now
1859                  */
1860                 if (cfq_cfqq_wait_request(cfqq)) {
1861                         cfq_mark_cfqq_must_dispatch(cfqq);
1862                         del_timer(&cfqd->idle_slice_timer);
1863                         blk_start_queueing(cfqd->queue);
1864                 }
1865         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1866                 /*
1867                  * not the active queue - expire current slice if it is
1868                  * idle and has expired it's mean thinktime or this new queue
1869                  * has some old slice time left and is of higher priority
1870                  */
1871                 cfq_preempt_queue(cfqd, cfqq);
1872                 cfq_mark_cfqq_must_dispatch(cfqq);
1873                 blk_start_queueing(cfqd->queue);
1874         }
1875 }
1876
1877 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1878 {
1879         struct cfq_data *cfqd = q->elevator->elevator_data;
1880         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1881
1882         cfq_log_cfqq(cfqd, cfqq, "insert_request");
1883         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1884
1885         cfq_add_rq_rb(rq);
1886
1887         list_add_tail(&rq->queuelist, &cfqq->fifo);
1888
1889         cfq_rq_enqueued(cfqd, cfqq, rq);
1890 }
1891
1892 /*
1893  * Update hw_tag based on peak queue depth over 50 samples under
1894  * sufficient load.
1895  */
1896 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1897 {
1898         if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1899                 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1900
1901         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1902             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1903                 return;
1904
1905         if (cfqd->hw_tag_samples++ < 50)
1906                 return;
1907
1908         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1909                 cfqd->hw_tag = 1;
1910         else
1911                 cfqd->hw_tag = 0;
1912
1913         cfqd->hw_tag_samples = 0;
1914         cfqd->rq_in_driver_peak = 0;
1915 }
1916
1917 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1918 {
1919         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1920         struct cfq_data *cfqd = cfqq->cfqd;
1921         const int sync = rq_is_sync(rq);
1922         unsigned long now;
1923
1924         now = jiffies;
1925         cfq_log_cfqq(cfqd, cfqq, "complete");
1926
1927         cfq_update_hw_tag(cfqd);
1928
1929         WARN_ON(!cfqd->rq_in_driver);
1930         WARN_ON(!cfqq->dispatched);
1931         cfqd->rq_in_driver--;
1932         cfqq->dispatched--;
1933
1934         if (cfq_cfqq_sync(cfqq))
1935                 cfqd->sync_flight--;
1936
1937         if (!cfq_class_idle(cfqq))
1938                 cfqd->last_end_request = now;
1939
1940         if (sync)
1941                 RQ_CIC(rq)->last_end_request = now;
1942
1943         /*
1944          * If this is the active queue, check if it needs to be expired,
1945          * or if we want to idle in case it has no pending requests.
1946          */
1947         if (cfqd->active_queue == cfqq) {
1948                 if (cfq_cfqq_slice_new(cfqq)) {
1949                         cfq_set_prio_slice(cfqd, cfqq);
1950                         cfq_clear_cfqq_slice_new(cfqq);
1951                 }
1952                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1953                         cfq_slice_expired(cfqd, 1);
1954                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list))
1955                         cfq_arm_slice_timer(cfqd);
1956         }
1957
1958         if (!cfqd->rq_in_driver)
1959                 cfq_schedule_dispatch(cfqd);
1960 }
1961
1962 /*
1963  * we temporarily boost lower priority queues if they are holding fs exclusive
1964  * resources. they are boosted to normal prio (CLASS_BE/4)
1965  */
1966 static void cfq_prio_boost(struct cfq_queue *cfqq)
1967 {
1968         if (has_fs_excl()) {
1969                 /*
1970                  * boost idle prio on transactions that would lock out other
1971                  * users of the filesystem
1972                  */
1973                 if (cfq_class_idle(cfqq))
1974                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1975                 if (cfqq->ioprio > IOPRIO_NORM)
1976                         cfqq->ioprio = IOPRIO_NORM;
1977         } else {
1978                 /*
1979                  * check if we need to unboost the queue
1980                  */
1981                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1982                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1983                 if (cfqq->ioprio != cfqq->org_ioprio)
1984                         cfqq->ioprio = cfqq->org_ioprio;
1985         }
1986 }
1987
1988 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1989 {
1990         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1991             !cfq_cfqq_must_alloc_slice(cfqq)) {
1992                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1993                 return ELV_MQUEUE_MUST;
1994         }
1995
1996         return ELV_MQUEUE_MAY;
1997 }
1998
1999 static int cfq_may_queue(struct request_queue *q, int rw)
2000 {
2001         struct cfq_data *cfqd = q->elevator->elevator_data;
2002         struct task_struct *tsk = current;
2003         struct cfq_io_context *cic;
2004         struct cfq_queue *cfqq;
2005
2006         /*
2007          * don't force setup of a queue from here, as a call to may_queue
2008          * does not necessarily imply that a request actually will be queued.
2009          * so just lookup a possibly existing queue, or return 'may queue'
2010          * if that fails
2011          */
2012         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2013         if (!cic)
2014                 return ELV_MQUEUE_MAY;
2015
2016         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
2017         if (cfqq) {
2018                 cfq_init_prio_data(cfqq, cic->ioc);
2019                 cfq_prio_boost(cfqq);
2020
2021                 return __cfq_may_queue(cfqq);
2022         }
2023
2024         return ELV_MQUEUE_MAY;
2025 }
2026
2027 /*
2028  * queue lock held here
2029  */
2030 static void cfq_put_request(struct request *rq)
2031 {
2032         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2033
2034         if (cfqq) {
2035                 const int rw = rq_data_dir(rq);
2036
2037                 BUG_ON(!cfqq->allocated[rw]);
2038                 cfqq->allocated[rw]--;
2039
2040                 put_io_context(RQ_CIC(rq)->ioc);
2041
2042                 rq->elevator_private = NULL;
2043                 rq->elevator_private2 = NULL;
2044
2045                 cfq_put_queue(cfqq);
2046         }
2047 }
2048
2049 /*
2050  * Allocate cfq data structures associated with this request.
2051  */
2052 static int
2053 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2054 {
2055         struct cfq_data *cfqd = q->elevator->elevator_data;
2056         struct cfq_io_context *cic;
2057         const int rw = rq_data_dir(rq);
2058         const int is_sync = rq_is_sync(rq);
2059         struct cfq_queue *cfqq;
2060         unsigned long flags;
2061
2062         might_sleep_if(gfp_mask & __GFP_WAIT);
2063
2064         cic = cfq_get_io_context(cfqd, gfp_mask);
2065
2066         spin_lock_irqsave(q->queue_lock, flags);
2067
2068         if (!cic)
2069                 goto queue_fail;
2070
2071         cfqq = cic_to_cfqq(cic, is_sync);
2072         if (!cfqq) {
2073                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2074
2075                 if (!cfqq)
2076                         goto queue_fail;
2077
2078                 cic_set_cfqq(cic, cfqq, is_sync);
2079         }
2080
2081         cfqq->allocated[rw]++;
2082         cfq_clear_cfqq_must_alloc(cfqq);
2083         atomic_inc(&cfqq->ref);
2084
2085         spin_unlock_irqrestore(q->queue_lock, flags);
2086
2087         rq->elevator_private = cic;
2088         rq->elevator_private2 = cfqq;
2089         return 0;
2090
2091 queue_fail:
2092         if (cic)
2093                 put_io_context(cic->ioc);
2094
2095         cfq_schedule_dispatch(cfqd);
2096         spin_unlock_irqrestore(q->queue_lock, flags);
2097         cfq_log(cfqd, "set_request fail");
2098         return 1;
2099 }
2100
2101 static void cfq_kick_queue(struct work_struct *work)
2102 {
2103         struct cfq_data *cfqd =
2104                 container_of(work, struct cfq_data, unplug_work);
2105         struct request_queue *q = cfqd->queue;
2106         unsigned long flags;
2107
2108         spin_lock_irqsave(q->queue_lock, flags);
2109         blk_start_queueing(q);
2110         spin_unlock_irqrestore(q->queue_lock, flags);
2111 }
2112
2113 /*
2114  * Timer running if the active_queue is currently idling inside its time slice
2115  */
2116 static void cfq_idle_slice_timer(unsigned long data)
2117 {
2118         struct cfq_data *cfqd = (struct cfq_data *) data;
2119         struct cfq_queue *cfqq;
2120         unsigned long flags;
2121         int timed_out = 1;
2122
2123         cfq_log(cfqd, "idle timer fired");
2124
2125         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2126
2127         cfqq = cfqd->active_queue;
2128         if (cfqq) {
2129                 timed_out = 0;
2130
2131                 /*
2132                  * expired
2133                  */
2134                 if (cfq_slice_used(cfqq))
2135                         goto expire;
2136
2137                 /*
2138                  * only expire and reinvoke request handler, if there are
2139                  * other queues with pending requests
2140                  */
2141                 if (!cfqd->busy_queues)
2142                         goto out_cont;
2143
2144                 /*
2145                  * not expired and it has a request pending, let it dispatch
2146                  */
2147                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2148                         cfq_mark_cfqq_must_dispatch(cfqq);
2149                         goto out_kick;
2150                 }
2151         }
2152 expire:
2153         cfq_slice_expired(cfqd, timed_out);
2154 out_kick:
2155         cfq_schedule_dispatch(cfqd);
2156 out_cont:
2157         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2158 }
2159
2160 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2161 {
2162         del_timer_sync(&cfqd->idle_slice_timer);
2163         kblockd_flush_work(&cfqd->unplug_work);
2164 }
2165
2166 static void cfq_put_async_queues(struct cfq_data *cfqd)
2167 {
2168         int i;
2169
2170         for (i = 0; i < IOPRIO_BE_NR; i++) {
2171                 if (cfqd->async_cfqq[0][i])
2172                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2173                 if (cfqd->async_cfqq[1][i])
2174                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2175         }
2176
2177         if (cfqd->async_idle_cfqq)
2178                 cfq_put_queue(cfqd->async_idle_cfqq);
2179 }
2180
2181 static void cfq_exit_queue(elevator_t *e)
2182 {
2183         struct cfq_data *cfqd = e->elevator_data;
2184         struct request_queue *q = cfqd->queue;
2185
2186         cfq_shutdown_timer_wq(cfqd);
2187
2188         spin_lock_irq(q->queue_lock);
2189
2190         if (cfqd->active_queue)
2191                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2192
2193         while (!list_empty(&cfqd->cic_list)) {
2194                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2195                                                         struct cfq_io_context,
2196                                                         queue_list);
2197
2198                 __cfq_exit_single_io_context(cfqd, cic);
2199         }
2200
2201         cfq_put_async_queues(cfqd);
2202
2203         spin_unlock_irq(q->queue_lock);
2204
2205         cfq_shutdown_timer_wq(cfqd);
2206
2207         kfree(cfqd);
2208 }
2209
2210 static void *cfq_init_queue(struct request_queue *q)
2211 {
2212         struct cfq_data *cfqd;
2213
2214         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2215         if (!cfqd)
2216                 return NULL;
2217
2218         cfqd->service_tree = CFQ_RB_ROOT;
2219         INIT_LIST_HEAD(&cfqd->cic_list);
2220
2221         cfqd->queue = q;
2222
2223         init_timer(&cfqd->idle_slice_timer);
2224         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2225         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2226
2227         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2228
2229         cfqd->last_end_request = jiffies;
2230         cfqd->cfq_quantum = cfq_quantum;
2231         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2232         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2233         cfqd->cfq_back_max = cfq_back_max;
2234         cfqd->cfq_back_penalty = cfq_back_penalty;
2235         cfqd->cfq_slice[0] = cfq_slice_async;
2236         cfqd->cfq_slice[1] = cfq_slice_sync;
2237         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2238         cfqd->cfq_slice_idle = cfq_slice_idle;
2239         cfqd->hw_tag = 1;
2240
2241         return cfqd;
2242 }
2243
2244 static void cfq_slab_kill(void)
2245 {
2246         /*
2247          * Caller already ensured that pending RCU callbacks are completed,
2248          * so we should have no busy allocations at this point.
2249          */
2250         if (cfq_pool)
2251                 kmem_cache_destroy(cfq_pool);
2252         if (cfq_ioc_pool)
2253                 kmem_cache_destroy(cfq_ioc_pool);
2254 }
2255
2256 static int __init cfq_slab_setup(void)
2257 {
2258         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2259         if (!cfq_pool)
2260                 goto fail;
2261
2262         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2263         if (!cfq_ioc_pool)
2264                 goto fail;
2265
2266         return 0;
2267 fail:
2268         cfq_slab_kill();
2269         return -ENOMEM;
2270 }
2271
2272 /*
2273  * sysfs parts below -->
2274  */
2275 static ssize_t
2276 cfq_var_show(unsigned int var, char *page)
2277 {
2278         return sprintf(page, "%d\n", var);
2279 }
2280
2281 static ssize_t
2282 cfq_var_store(unsigned int *var, const char *page, size_t count)
2283 {
2284         char *p = (char *) page;
2285
2286         *var = simple_strtoul(p, &p, 10);
2287         return count;
2288 }
2289
2290 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2291 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2292 {                                                                       \
2293         struct cfq_data *cfqd = e->elevator_data;                       \
2294         unsigned int __data = __VAR;                                    \
2295         if (__CONV)                                                     \
2296                 __data = jiffies_to_msecs(__data);                      \
2297         return cfq_var_show(__data, (page));                            \
2298 }
2299 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2300 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2301 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2302 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2303 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2304 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2305 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2306 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2307 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2308 #undef SHOW_FUNCTION
2309
2310 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2311 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2312 {                                                                       \
2313         struct cfq_data *cfqd = e->elevator_data;                       \
2314         unsigned int __data;                                            \
2315         int ret = cfq_var_store(&__data, (page), count);                \
2316         if (__data < (MIN))                                             \
2317                 __data = (MIN);                                         \
2318         else if (__data > (MAX))                                        \
2319                 __data = (MAX);                                         \
2320         if (__CONV)                                                     \
2321                 *(__PTR) = msecs_to_jiffies(__data);                    \
2322         else                                                            \
2323                 *(__PTR) = __data;                                      \
2324         return ret;                                                     \
2325 }
2326 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2327 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2328                 UINT_MAX, 1);
2329 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2330                 UINT_MAX, 1);
2331 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2332 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2333                 UINT_MAX, 0);
2334 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2335 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2336 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2337 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2338                 UINT_MAX, 0);
2339 #undef STORE_FUNCTION
2340
2341 #define CFQ_ATTR(name) \
2342         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2343
2344 static struct elv_fs_entry cfq_attrs[] = {
2345         CFQ_ATTR(quantum),
2346         CFQ_ATTR(fifo_expire_sync),
2347         CFQ_ATTR(fifo_expire_async),
2348         CFQ_ATTR(back_seek_max),
2349         CFQ_ATTR(back_seek_penalty),
2350         CFQ_ATTR(slice_sync),
2351         CFQ_ATTR(slice_async),
2352         CFQ_ATTR(slice_async_rq),
2353         CFQ_ATTR(slice_idle),
2354         __ATTR_NULL
2355 };
2356
2357 static struct elevator_type iosched_cfq = {
2358         .ops = {
2359                 .elevator_merge_fn =            cfq_merge,
2360                 .elevator_merged_fn =           cfq_merged_request,
2361                 .elevator_merge_req_fn =        cfq_merged_requests,
2362                 .elevator_allow_merge_fn =      cfq_allow_merge,
2363                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2364                 .elevator_add_req_fn =          cfq_insert_request,
2365                 .elevator_activate_req_fn =     cfq_activate_request,
2366                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2367                 .elevator_queue_empty_fn =      cfq_queue_empty,
2368                 .elevator_completed_req_fn =    cfq_completed_request,
2369                 .elevator_former_req_fn =       elv_rb_former_request,
2370                 .elevator_latter_req_fn =       elv_rb_latter_request,
2371                 .elevator_set_req_fn =          cfq_set_request,
2372                 .elevator_put_req_fn =          cfq_put_request,
2373                 .elevator_may_queue_fn =        cfq_may_queue,
2374                 .elevator_init_fn =             cfq_init_queue,
2375                 .elevator_exit_fn =             cfq_exit_queue,
2376                 .trim =                         cfq_free_io_context,
2377         },
2378         .elevator_attrs =       cfq_attrs,
2379         .elevator_name =        "cfq",
2380         .elevator_owner =       THIS_MODULE,
2381 };
2382
2383 static int __init cfq_init(void)
2384 {
2385         /*
2386          * could be 0 on HZ < 1000 setups
2387          */
2388         if (!cfq_slice_async)
2389                 cfq_slice_async = 1;
2390         if (!cfq_slice_idle)
2391                 cfq_slice_idle = 1;
2392
2393         if (cfq_slab_setup())
2394                 return -ENOMEM;
2395
2396         elv_register(&iosched_cfq);
2397
2398         return 0;
2399 }
2400
2401 static void __exit cfq_exit(void)
2402 {
2403         DECLARE_COMPLETION_ONSTACK(all_gone);
2404         elv_unregister(&iosched_cfq);
2405         ioc_gone = &all_gone;
2406         /* ioc_gone's update must be visible before reading ioc_count */
2407         smp_wmb();
2408
2409         /*
2410          * this also protects us from entering cfq_slab_kill() with
2411          * pending RCU callbacks
2412          */
2413         if (elv_ioc_count_read(ioc_count))
2414                 wait_for_completion(&all_gone);
2415         cfq_slab_kill();
2416 }
2417
2418 module_init(cfq_init);
2419 module_exit(cfq_exit);
2420
2421 MODULE_AUTHOR("Jens Axboe");
2422 MODULE_LICENSE("GPL");
2423 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");