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          * Used to track any pending rt requests so we can pre-empt current
  89          * non-RT cfqq in service when this value is non-zero.
  90          */
  91         unsigned int busy_rt_queues;
  92
  93         int rq_in_driver;
  94         int sync_flight;
  95
  96         /*
  97          * queue-depth detection
  98          */
  99         int rq_queued;
 100         int hw_tag;
 101         int hw_tag_samples;
 102         int rq_in_driver_peak;
 103
 104         /*
 105          * idle window management
 106          */
 107         struct timer_list idle_slice_timer;
 108         struct work_struct unplug_work;
 109
 110         struct cfq_queue *active_queue;
 111         struct cfq_io_context *active_cic;
 112
 113         /*
 114          * async queue for each priority case
 115          */
 116         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 117         struct cfq_queue *async_idle_cfqq;
 118
 119         sector_t last_position;
 120         unsigned long last_end_request;
 121
 122         /*
 123          * tunables, see top of file
 124          */
 125         unsigned int cfq_quantum;
 126         unsigned int cfq_fifo_expire[2];
 127         unsigned int cfq_back_penalty;
 128         unsigned int cfq_back_max;
 129         unsigned int cfq_slice[2];
 130         unsigned int cfq_slice_async_rq;
 131         unsigned int cfq_slice_idle;
 132
 133         struct list_head cic_list;
 134 };
 135
 136 /*
 137  * Per process-grouping structure
 138  */
 139 struct cfq_queue {
 140         /* reference count */
 141         atomic_t ref;
 142         /* various state flags, see below */
 143         unsigned int flags;
 144         /* parent cfq_data */
 145         struct cfq_data *cfqd;
 146         /* service_tree member */
 147         struct rb_node rb_node;
 148         /* service_tree key */
 149         unsigned long rb_key;
 150         /* sorted list of pending requests */
 151         struct rb_root sort_list;
 152         /* if fifo isn't expired, next request to serve */
 153         struct request *next_rq;
 154         /* requests queued in sort_list */
 155         int queued[2];
 156         /* currently allocated requests */
 157         int allocated[2];
 158         /* fifo list of requests in sort_list */
 159         struct list_head fifo;
 160
 161         unsigned long slice_end;
 162         long slice_resid;
 163
 164         /* pending metadata requests */
 165         int meta_pending;
 166         /* number of requests that are on the dispatch list or inside driver */
 167         int dispatched;
 168
 169         /* io prio of this group */
 170         unsigned short ioprio, org_ioprio;
 171         unsigned short ioprio_class, org_ioprio_class;
 172
 173         pid_t pid;
 174 };
 175
 176 enum cfqq_state_flags {
 177         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 178         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 179         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
 180         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 181         CFQ_CFQQ_FLAG_must_dispatch,    /* must dispatch, even if expired */
 182         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 183         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 184         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 185         CFQ_CFQQ_FLAG_queue_new,        /* queue never been serviced */
 186         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 187         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 188 };
 189
 190 #define CFQ_CFQQ_FNS(name)                                              \
 191 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 192 {                                                                       \
 193         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 194 }                                                                       \
 195 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 196 {                                                                       \
 197         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 198 }                                                                       \
 199 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 200 {                                                                       \
 201         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 202 }
 203
 204 CFQ_CFQQ_FNS(on_rr);
 205 CFQ_CFQQ_FNS(wait_request);
 206 CFQ_CFQQ_FNS(must_alloc);
 207 CFQ_CFQQ_FNS(must_alloc_slice);
 208 CFQ_CFQQ_FNS(must_dispatch);
 209 CFQ_CFQQ_FNS(fifo_expire);
 210 CFQ_CFQQ_FNS(idle_window);
 211 CFQ_CFQQ_FNS(prio_changed);
 212 CFQ_CFQQ_FNS(queue_new);
 213 CFQ_CFQQ_FNS(slice_new);
 214 CFQ_CFQQ_FNS(sync);
 215 #undef CFQ_CFQQ_FNS
 216
 217 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 218         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 219 #define cfq_log(cfqd, fmt, args...)     \
 220         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 221
 222 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 223 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
 224                                        struct io_context *, gfp_t);
 225 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 226                                                 struct io_context *);
 227
 228 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 229                                             int is_sync)
 230 {
 231         return cic->cfqq[!!is_sync];
 232 }
 233
 234 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 235                                 struct cfq_queue *cfqq, int is_sync)
 236 {
 237         cic->cfqq[!!is_sync] = cfqq;
 238 }
 239
 240 /*
 241  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 242  * set (in which case it could also be direct WRITE).
 243  */
 244 static inline int cfq_bio_sync(struct bio *bio)
 245 {
 246         if (bio_data_dir(bio) == READ || bio_sync(bio))
 247                 return 1;
 248
 249         return 0;
 250 }
 251
 252 /*
 253  * scheduler run of queue, if there are requests pending and no one in the
 254  * driver that will restart queueing
 255  */
 256 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 257 {
 258         if (cfqd->busy_queues) {
 259                 cfq_log(cfqd, "schedule dispatch");
 260                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 261         }
 262 }
 263
 264 static int cfq_queue_empty(struct request_queue *q)
 265 {
 266         struct cfq_data *cfqd = q->elevator->elevator_data;
 267
 268         return !cfqd->busy_queues;
 269 }
 270
 271 /*
 272  * Scale schedule slice based on io priority. Use the sync time slice only
 273  * if a queue is marked sync and has sync io queued. A sync queue with async
 274  * io only, should not get full sync slice length.
 275  */
 276 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
 277                                  unsigned short prio)
 278 {
 279         const int base_slice = cfqd->cfq_slice[sync];
 280
 281         WARN_ON(prio >= IOPRIO_BE_NR);
 282
 283         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 284 }
 285
 286 static inline int
 287 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 288 {
 289         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 290 }
 291
 292 static inline void
 293 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 294 {
 295         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
 296         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 297 }
 298
 299 /*
 300  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 301  * isn't valid until the first request from the dispatch is activated
 302  * and the slice time set.
 303  */
 304 static inline int cfq_slice_used(struct cfq_queue *cfqq)
 305 {
 306         if (cfq_cfqq_slice_new(cfqq))
 307                 return 0;
 308         if (time_before(jiffies, cfqq->slice_end))
 309                 return 0;
 310
 311         return 1;
 312 }
 313
 314 /*
 315  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 316  * We choose the request that is closest to the head right now. Distance
 317  * behind the head is penalized and only allowed to a certain extent.
 318  */
 319 static struct request *
 320 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
 321 {
 322         sector_t last, s1, s2, d1 = 0, d2 = 0;
 323         unsigned long back_max;
 324 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 325 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 326         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 327
 328         if (rq1 == NULL || rq1 == rq2)
 329                 return rq2;
 330         if (rq2 == NULL)
 331                 return rq1;
 332
 333         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 334                 return rq1;
 335         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 336                 return rq2;
 337         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 338                 return rq1;
 339         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 340                 return rq2;
 341
 342         s1 = rq1->sector;
 343         s2 = rq2->sector;
 344
 345         last = cfqd->last_position;
 346
 347         /*
 348          * by definition, 1KiB is 2 sectors
 349          */
 350         back_max = cfqd->cfq_back_max * 2;
 351
 352         /*
 353          * Strict one way elevator _except_ in the case where we allow
 354          * short backward seeks which are biased as twice the cost of a
 355          * similar forward seek.
 356          */
 357         if (s1 >= last)
 358                 d1 = s1 - last;
 359         else if (s1 + back_max >= last)
 360                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 361         else
 362                 wrap |= CFQ_RQ1_WRAP;
 363
 364         if (s2 >= last)
 365                 d2 = s2 - last;
 366         else if (s2 + back_max >= last)
 367                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 368         else
 369                 wrap |= CFQ_RQ2_WRAP;
 370
 371         /* Found required data */
 372
 373         /*
 374          * By doing switch() on the bit mask "wrap" we avoid having to
 375          * check two variables for all permutations: --> faster!
 376          */
 377         switch (wrap) {
 378         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 379                 if (d1 < d2)
 380                         return rq1;
 381                 else if (d2 < d1)
 382                         return rq2;
 383                 else {
 384                         if (s1 >= s2)
 385                                 return rq1;
 386                         else
 387                                 return rq2;
 388                 }
 389
 390         case CFQ_RQ2_WRAP:
 391                 return rq1;
 392         case CFQ_RQ1_WRAP:
 393                 return rq2;
 394         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 395         default:
 396                 /*
 397                  * Since both rqs are wrapped,
 398                  * start with the one that's further behind head
 399                  * (--> only *one* back seek required),
 400                  * since back seek takes more time than forward.
 401                  */
 402                 if (s1 <= s2)
 403                         return rq1;
 404                 else
 405                         return rq2;
 406         }
 407 }
 408
 409 /*
 410  * The below is leftmost cache rbtree addon
 411  */
 412 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 413 {
 414         if (!root->left)
 415                 root->left = rb_first(&root->rb);
 416
 417         if (root->left)
 418                 return rb_entry(root->left, struct cfq_queue, rb_node);
 419
 420         return NULL;
 421 }
 422
 423 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 424 {
 425         if (root->left == n)
 426                 root->left = NULL;
 427
 428         rb_erase(n, &root->rb);
 429         RB_CLEAR_NODE(n);
 430 }
 431
 432 /*
 433  * would be nice to take fifo expire time into account as well
 434  */
 435 static struct request *
 436 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 437                   struct request *last)
 438 {
 439         struct rb_node *rbnext = rb_next(&last->rb_node);
 440         struct rb_node *rbprev = rb_prev(&last->rb_node);
 441         struct request *next = NULL, *prev = NULL;
 442
 443         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 444
 445         if (rbprev)
 446                 prev = rb_entry_rq(rbprev);
 447
 448         if (rbnext)
 449                 next = rb_entry_rq(rbnext);
 450         else {
 451                 rbnext = rb_first(&cfqq->sort_list);
 452                 if (rbnext && rbnext != &last->rb_node)
 453                         next = rb_entry_rq(rbnext);
 454         }
 455
 456         return cfq_choose_req(cfqd, next, prev);
 457 }
 458
 459 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 460                                       struct cfq_queue *cfqq)
 461 {
 462         /*
 463          * just an approximation, should be ok.
 464          */
 465         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 466                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 467 }
 468
 469 /*
 470  * The cfqd->service_tree holds all pending cfq_queue's that have
 471  * requests waiting to be processed. It is sorted in the order that
 472  * we will service the queues.
 473  */
 474 static void cfq_service_tree_add(struct cfq_data *cfqd,
 475                                     struct cfq_queue *cfqq, int add_front)
 476 {
 477         struct rb_node **p, *parent;
 478         struct cfq_queue *__cfqq;
 479         unsigned long rb_key;
 480         int left;
 481
 482         if (cfq_class_idle(cfqq)) {
 483                 rb_key = CFQ_IDLE_DELAY;
 484                 parent = rb_last(&cfqd->service_tree.rb);
 485                 if (parent && parent != &cfqq->rb_node) {
 486                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 487                         rb_key += __cfqq->rb_key;
 488                 } else
 489                         rb_key += jiffies;
 490         } else if (!add_front) {
 491                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
 492                 rb_key += cfqq->slice_resid;
 493                 cfqq->slice_resid = 0;
 494         } else
 495                 rb_key = 0;
 496
 497         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
 498                 /*
 499                  * same position, nothing more to do
 500                  */
 501                 if (rb_key == cfqq->rb_key)
 502                         return;
 503
 504                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 505         }
 506
 507         left = 1;
 508         parent = NULL;
 509         p = &cfqd->service_tree.rb.rb_node;
 510         while (*p) {
 511                 struct rb_node **n;
 512
 513                 parent = *p;
 514                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
 515
 516                 /*
 517                  * sort RT queues first, we always want to give
 518                  * preference to them. IDLE queues goes to the back.
 519                  * after that, sort on the next service time.
 520                  */
 521                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
 522                         n = &(*p)->rb_left;
 523                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
 524                         n = &(*p)->rb_right;
 525                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
 526                         n = &(*p)->rb_left;
 527                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
 528                         n = &(*p)->rb_right;
 529                 else if (rb_key < __cfqq->rb_key)
 530                         n = &(*p)->rb_left;
 531                 else
 532                         n = &(*p)->rb_right;
 533
 534                 if (n == &(*p)->rb_right)
 535                         left = 0;
 536
 537                 p = n;
 538         }
 539
 540         if (left)
 541                 cfqd->service_tree.left = &cfqq->rb_node;
 542
 543         cfqq->rb_key = rb_key;
 544         rb_link_node(&cfqq->rb_node, parent, p);
 545         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
 546 }
 547
 548 /*
 549  * Update cfqq's position in the service tree.
 550  */
 551 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 552 {
 553         /*
 554          * Resorting requires the cfqq to be on the RR list already.
 555          */
 556         if (cfq_cfqq_on_rr(cfqq))
 557                 cfq_service_tree_add(cfqd, cfqq, 0);
 558 }
 559
 560 /*
 561  * add to busy list of queues for service, trying to be fair in ordering
 562  * the pending list according to last request service
 563  */
 564 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 565 {
 566         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
 567         BUG_ON(cfq_cfqq_on_rr(cfqq));
 568         cfq_mark_cfqq_on_rr(cfqq);
 569         cfqd->busy_queues++;
 570         if (cfq_class_rt(cfqq))
 571                 cfqd->busy_rt_queues++;
 572
 573         cfq_resort_rr_list(cfqd, cfqq);
 574 }
 575
 576 /*
 577  * Called when the cfqq no longer has requests pending, remove it from
 578  * the service tree.
 579  */
 580 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 581 {
 582         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
 583         BUG_ON(!cfq_cfqq_on_rr(cfqq));
 584         cfq_clear_cfqq_on_rr(cfqq);
 585
 586         if (!RB_EMPTY_NODE(&cfqq->rb_node))
 587                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
 588
 589         BUG_ON(!cfqd->busy_queues);
 590         cfqd->busy_queues--;
 591         if (cfq_class_rt(cfqq))
 592                 cfqd->busy_rt_queues--;
 593 }
 594
 595 /*
 596  * rb tree support functions
 597  */
 598 static void cfq_del_rq_rb(struct request *rq)
 599 {
 600         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 601         struct cfq_data *cfqd = cfqq->cfqd;
 602         const int sync = rq_is_sync(rq);
 603
 604         BUG_ON(!cfqq->queued[sync]);
 605         cfqq->queued[sync]--;
 606
 607         elv_rb_del(&cfqq->sort_list, rq);
 608
 609         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
 610                 cfq_del_cfqq_rr(cfqd, cfqq);
 611 }
 612
 613 static void cfq_add_rq_rb(struct request *rq)
 614 {
 615         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 616         struct cfq_data *cfqd = cfqq->cfqd;
 617         struct request *__alias;
 618
 619         cfqq->queued[rq_is_sync(rq)]++;
 620
 621         /*
 622          * looks a little odd, but the first insert might return an alias.
 623          * if that happens, put the alias on the dispatch list
 624          */
 625         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
 626                 cfq_dispatch_insert(cfqd->queue, __alias);
 627
 628         if (!cfq_cfqq_on_rr(cfqq))
 629                 cfq_add_cfqq_rr(cfqd, cfqq);
 630
 631         /*
 632          * check if this request is a better next-serve candidate
 633          */
 634         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
 635         BUG_ON(!cfqq->next_rq);
 636 }
 637
 638 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
 639 {
 640         elv_rb_del(&cfqq->sort_list, rq);
 641         cfqq->queued[rq_is_sync(rq)]--;
 642         cfq_add_rq_rb(rq);
 643 }
 644
 645 static struct request *
 646 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
 647 {
 648         struct task_struct *tsk = current;
 649         struct cfq_io_context *cic;
 650         struct cfq_queue *cfqq;
 651
 652         cic = cfq_cic_lookup(cfqd, tsk->io_context);
 653         if (!cic)
 654                 return NULL;
 655
 656         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 657         if (cfqq) {
 658                 sector_t sector = bio->bi_sector + bio_sectors(bio);
 659
 660                 return elv_rb_find(&cfqq->sort_list, sector);
 661         }
 662
 663         return NULL;
 664 }
 665
 666 static void cfq_activate_request(struct request_queue *q, struct request *rq)
 667 {
 668         struct cfq_data *cfqd = q->elevator->elevator_data;
 669
 670         cfqd->rq_in_driver++;
 671         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
 672                                                 cfqd->rq_in_driver);
 673
 674         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
 675 }
 676
 677 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
 678 {
 679         struct cfq_data *cfqd = q->elevator->elevator_data;
 680
 681         WARN_ON(!cfqd->rq_in_driver);
 682         cfqd->rq_in_driver--;
 683         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
 684                                                 cfqd->rq_in_driver);
 685 }
 686
 687 static void cfq_remove_request(struct request *rq)
 688 {
 689         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 690
 691         if (cfqq->next_rq == rq)
 692                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
 693
 694         list_del_init(&rq->queuelist);
 695         cfq_del_rq_rb(rq);
 696
 697         cfqq->cfqd->rq_queued--;
 698         if (rq_is_meta(rq)) {
 699                 WARN_ON(!cfqq->meta_pending);
 700                 cfqq->meta_pending--;
 701         }
 702 }
 703
 704 static int cfq_merge(struct request_queue *q, struct request **req,
 705                      struct bio *bio)
 706 {
 707         struct cfq_data *cfqd = q->elevator->elevator_data;
 708         struct request *__rq;
 709
 710         __rq = cfq_find_rq_fmerge(cfqd, bio);
 711         if (__rq && elv_rq_merge_ok(__rq, bio)) {
 712                 *req = __rq;
 713                 return ELEVATOR_FRONT_MERGE;
 714         }
 715
 716         return ELEVATOR_NO_MERGE;
 717 }
 718
 719 static void cfq_merged_request(struct request_queue *q, struct request *req,
 720                                int type)
 721 {
 722         if (type == ELEVATOR_FRONT_MERGE) {
 723                 struct cfq_queue *cfqq = RQ_CFQQ(req);
 724
 725                 cfq_reposition_rq_rb(cfqq, req);
 726         }
 727 }
 728
 729 static void
 730 cfq_merged_requests(struct request_queue *q, struct request *rq,
 731                     struct request *next)
 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 int cfq_allow_merge(struct request_queue *q, struct request *rq,
 744                            struct bio *bio)
 745 {
 746         struct cfq_data *cfqd = q->elevator->elevator_data;
 747         struct cfq_io_context *cic;
 748         struct cfq_queue *cfqq;
 749
 750         /*
 751          * Disallow merge of a sync bio into an async request.
 752          */
 753         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
 754                 return 0;
 755
 756         /*
 757          * Lookup the cfqq that this bio will be queued with. Allow
 758          * merge only if rq is queued there.
 759          */
 760         cic = cfq_cic_lookup(cfqd, current->io_context);
 761         if (!cic)
 762                 return 0;
 763
 764         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
 765         if (cfqq == RQ_CFQQ(rq))
 766                 return 1;
 767
 768         return 0;
 769 }
 770
 771 static void __cfq_set_active_queue(struct cfq_data *cfqd,
 772                                    struct cfq_queue *cfqq)
 773 {
 774         if (cfqq) {
 775                 cfq_log_cfqq(cfqd, cfqq, "set_active");
 776                 cfqq->slice_end = 0;
 777                 cfq_clear_cfqq_must_alloc_slice(cfqq);
 778                 cfq_clear_cfqq_fifo_expire(cfqq);
 779                 cfq_mark_cfqq_slice_new(cfqq);
 780                 cfq_clear_cfqq_queue_new(cfqq);
 781         }
 782
 783         cfqd->active_queue = cfqq;
 784 }
 785
 786 /*
 787  * current cfqq expired its slice (or was too idle), select new one
 788  */
 789 static void
 790 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 791                     int timed_out)
 792 {
 793         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
 794
 795         if (cfq_cfqq_wait_request(cfqq))
 796                 del_timer(&cfqd->idle_slice_timer);
 797
 798         cfq_clear_cfqq_must_dispatch(cfqq);
 799         cfq_clear_cfqq_wait_request(cfqq);
 800
 801         /*
 802          * store what was left of this slice, if the queue idled/timed out
 803          */
 804         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
 805                 cfqq->slice_resid = cfqq->slice_end - jiffies;
 806                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
 807         }
 808
 809         cfq_resort_rr_list(cfqd, cfqq);
 810
 811         if (cfqq == cfqd->active_queue)
 812                 cfqd->active_queue = NULL;
 813
 814         if (cfqd->active_cic) {
 815                 put_io_context(cfqd->active_cic->ioc);
 816                 cfqd->active_cic = NULL;
 817         }
 818 }
 819
 820 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
 821 {
 822         struct cfq_queue *cfqq = cfqd->active_queue;
 823
 824         if (cfqq)
 825                 __cfq_slice_expired(cfqd, cfqq, timed_out);
 826 }
 827
 828 /*
 829  * Get next queue for service. Unless we have a queue preemption,
 830  * we'll simply select the first cfqq in the service tree.
 831  */
 832 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
 833 {
 834         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
 835                 return NULL;
 836
 837         return cfq_rb_first(&cfqd->service_tree);
 838 }
 839
 840 /*
 841  * Get and set a new active queue for service.
 842  */
 843 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
 844 {
 845         struct cfq_queue *cfqq;
 846
 847         cfqq = cfq_get_next_queue(cfqd);
 848         __cfq_set_active_queue(cfqd, cfqq);
 849         return cfqq;
 850 }
 851
 852 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
 853                                           struct request *rq)
 854 {
 855         if (rq->sector >= cfqd->last_position)
 856                 return rq->sector - cfqd->last_position;
 857         else
 858                 return cfqd->last_position - rq->sector;
 859 }
 860
 861 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
 862 {
 863         struct cfq_io_context *cic = cfqd->active_cic;
 864
 865         if (!sample_valid(cic->seek_samples))
 866                 return 0;
 867
 868         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
 869 }
 870
 871 static int cfq_close_cooperator(struct cfq_data *cfq_data,
 872                                 struct cfq_queue *cfqq)
 873 {
 874         /*
 875          * We should notice if some of the queues are cooperating, eg
 876          * working closely on the same area of the disk. In that case,
 877          * we can group them together and don't waste time idling.
 878          */
 879         return 0;
 880 }
 881
 882 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
 883
 884 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
 885 {
 886         struct cfq_queue *cfqq = cfqd->active_queue;
 887         struct cfq_io_context *cic;
 888         unsigned long sl;
 889
 890         /*
 891          * SSD device without seek penalty, disable idling. But only do so
 892          * for devices that support queuing, otherwise we still have a problem
 893          * with sync vs async workloads.
 894          */
 895         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
 896                 return;
 897
 898         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
 899         WARN_ON(cfq_cfqq_slice_new(cfqq));
 900
 901         /*
 902          * idle is disabled, either manually or by past process history
 903          */
 904         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
 905                 return;
 906
 907         /*
 908          * still requests with the driver, don't idle
 909          */
 910         if (cfqd->rq_in_driver)
 911                 return;
 912
 913         /*
 914          * task has exited, don't wait
 915          */
 916         cic = cfqd->active_cic;
 917         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
 918                 return;
 919
 920         /*
 921          * See if this prio level has a good candidate
 922          */
 923         if (cfq_close_cooperator(cfqd, cfqq) &&
 924             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
 925                 return;
 926
 927         cfq_mark_cfqq_must_dispatch(cfqq);
 928         cfq_mark_cfqq_wait_request(cfqq);
 929
 930         /*
 931          * we don't want to idle for seeks, but we do want to allow
 932          * fair distribution of slice time for a process doing back-to-back
 933          * seeks. so allow a little bit of time for him to submit a new rq
 934          */
 935         sl = cfqd->cfq_slice_idle;
 936         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
 937                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
 938
 939         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
 940         cfq_log(cfqd, "arm_idle: %lu", sl);
 941 }
 942
 943 /*
 944  * Move request from internal lists to the request queue dispatch list.
 945  */
 946 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
 947 {
 948         struct cfq_data *cfqd = q->elevator->elevator_data;
 949         struct cfq_queue *cfqq = RQ_CFQQ(rq);
 950
 951         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
 952
 953         cfq_remove_request(rq);
 954         cfqq->dispatched++;
 955         elv_dispatch_sort(q, rq);
 956
 957         if (cfq_cfqq_sync(cfqq))
 958                 cfqd->sync_flight++;
 959 }
 960
 961 /*
 962  * return expired entry, or NULL to just start from scratch in rbtree
 963  */
 964 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
 965 {
 966         struct cfq_data *cfqd = cfqq->cfqd;
 967         struct request *rq;
 968         int fifo;
 969
 970         if (cfq_cfqq_fifo_expire(cfqq))
 971                 return NULL;
 972
 973         cfq_mark_cfqq_fifo_expire(cfqq);
 974
 975         if (list_empty(&cfqq->fifo))
 976                 return NULL;
 977
 978         fifo = cfq_cfqq_sync(cfqq);
 979         rq = rq_entry_fifo(cfqq->fifo.next);
 980
 981         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
 982                 rq = NULL;
 983
 984         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
 985         return rq;
 986 }
 987
 988 static inline int
 989 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 990 {
 991         const int base_rq = cfqd->cfq_slice_async_rq;
 992
 993         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
 994
 995         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
 996 }
 997
 998 /*
 999  * Select a queue for service. If we have a current active queue,
1000  * check whether to continue servicing it, or retrieve and set a new one.
1001  */
1002 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1003 {
1004         struct cfq_queue *cfqq;
1005
1006         cfqq = cfqd->active_queue;
1007         if (!cfqq)
1008                 goto new_queue;
1009
1010         /*
1011          * The active queue has run out of time, expire it and select new.
1012          */
1013         if (cfq_slice_used(cfqq))
1014                 goto expire;
1015
1016         /*
1017          * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1018          * cfqq.
1019          */
1020         if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1021                 /*
1022                  * We simulate this as cfqq timed out so that it gets to bank
1023                  * the remaining of its time slice.
1024                  */
1025                 cfq_log_cfqq(cfqd, cfqq, "preempt");
1026                 cfq_slice_expired(cfqd, 1);
1027                 goto new_queue;
1028         }
1029
1030         /*
1031          * The active queue has requests and isn't expired, allow it to
1032          * dispatch.
1033          */
1034         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1035                 goto keep_queue;
1036
1037         /*
1038          * No requests pending. If the active queue still has requests in
1039          * flight or is idling for a new request, allow either of these
1040          * conditions to happen (or time out) before selecting a new queue.
1041          */
1042         if (timer_pending(&cfqd->idle_slice_timer) ||
1043             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1044                 cfqq = NULL;
1045                 goto keep_queue;
1046         }
1047
1048 expire:
1049         cfq_slice_expired(cfqd, 0);
1050 new_queue:
1051         cfqq = cfq_set_active_queue(cfqd);
1052 keep_queue:
1053         return cfqq;
1054 }
1055
1056 /*
1057  * Dispatch some requests from cfqq, moving them to the request queue
1058  * dispatch list.
1059  */
1060 static int
1061 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1062                         int max_dispatch)
1063 {
1064         int dispatched = 0;
1065
1066         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1067
1068         do {
1069                 struct request *rq;
1070
1071                 /*
1072                  * follow expired path, else get first next available
1073                  */
1074                 rq = cfq_check_fifo(cfqq);
1075                 if (rq == NULL)
1076                         rq = cfqq->next_rq;
1077
1078                 /*
1079                  * finally, insert request into driver dispatch list
1080                  */
1081                 cfq_dispatch_insert(cfqd->queue, rq);
1082
1083                 dispatched++;
1084
1085                 if (!cfqd->active_cic) {
1086                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
1087                         cfqd->active_cic = RQ_CIC(rq);
1088                 }
1089
1090                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
1091                         break;
1092
1093                 /*
1094                  * If there is a non-empty RT cfqq waiting for current
1095                  * cfqq's timeslice to complete, pre-empt this cfqq
1096                  */
1097                 if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues)
1098                         break;
1099
1100         } while (dispatched < max_dispatch);
1101
1102         /*
1103          * expire an async queue immediately if it has used up its slice. idle
1104          * queue always expire after 1 dispatch round.
1105          */
1106         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1107             dispatched >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1108             cfq_class_idle(cfqq))) {
1109                 cfqq->slice_end = jiffies + 1;
1110                 cfq_slice_expired(cfqd, 0);
1111         }
1112
1113         return dispatched;
1114 }
1115
1116 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1117 {
1118         int dispatched = 0;
1119
1120         while (cfqq->next_rq) {
1121                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1122                 dispatched++;
1123         }
1124
1125         BUG_ON(!list_empty(&cfqq->fifo));
1126         return dispatched;
1127 }
1128
1129 /*
1130  * Drain our current requests. Used for barriers and when switching
1131  * io schedulers on-the-fly.
1132  */
1133 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1134 {
1135         struct cfq_queue *cfqq;
1136         int dispatched = 0;
1137
1138         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1139                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1140
1141         cfq_slice_expired(cfqd, 0);
1142
1143         BUG_ON(cfqd->busy_queues);
1144
1145         cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1146         return dispatched;
1147 }
1148
1149 static int cfq_dispatch_requests(struct request_queue *q, int force)
1150 {
1151         struct cfq_data *cfqd = q->elevator->elevator_data;
1152         struct cfq_queue *cfqq;
1153         int dispatched;
1154
1155         if (!cfqd->busy_queues)
1156                 return 0;
1157
1158         if (unlikely(force))
1159                 return cfq_forced_dispatch(cfqd);
1160
1161         dispatched = 0;
1162         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
1163                 int max_dispatch;
1164
1165                 max_dispatch = cfqd->cfq_quantum;
1166                 if (cfq_class_idle(cfqq))
1167                         max_dispatch = 1;
1168
1169                 if (cfqq->dispatched >= max_dispatch && cfqd->busy_queues > 1)
1170                         break;
1171
1172                 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1173                         break;
1174
1175                 cfq_clear_cfqq_must_dispatch(cfqq);
1176                 cfq_clear_cfqq_wait_request(cfqq);
1177                 del_timer(&cfqd->idle_slice_timer);
1178
1179                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1180         }
1181
1182         cfq_log(cfqd, "dispatched=%d", dispatched);
1183         return dispatched;
1184 }
1185
1186 /*
1187  * task holds one reference to the queue, dropped when task exits. each rq
1188  * in-flight on this queue also holds a reference, dropped when rq is freed.
1189  *
1190  * queue lock must be held here.
1191  */
1192 static void cfq_put_queue(struct cfq_queue *cfqq)
1193 {
1194         struct cfq_data *cfqd = cfqq->cfqd;
1195
1196         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1197
1198         if (!atomic_dec_and_test(&cfqq->ref))
1199                 return;
1200
1201         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1202         BUG_ON(rb_first(&cfqq->sort_list));
1203         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1204         BUG_ON(cfq_cfqq_on_rr(cfqq));
1205
1206         if (unlikely(cfqd->active_queue == cfqq)) {
1207                 __cfq_slice_expired(cfqd, cfqq, 0);
1208                 cfq_schedule_dispatch(cfqd);
1209         }
1210
1211         kmem_cache_free(cfq_pool, cfqq);
1212 }
1213
1214 /*
1215  * Must always be called with the rcu_read_lock() held
1216  */
1217 static void
1218 __call_for_each_cic(struct io_context *ioc,
1219                     void (*func)(struct io_context *, struct cfq_io_context *))
1220 {
1221         struct cfq_io_context *cic;
1222         struct hlist_node *n;
1223
1224         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1225                 func(ioc, cic);
1226 }
1227
1228 /*
1229  * Call func for each cic attached to this ioc.
1230  */
1231 static void
1232 call_for_each_cic(struct io_context *ioc,
1233                   void (*func)(struct io_context *, struct cfq_io_context *))
1234 {
1235         rcu_read_lock();
1236         __call_for_each_cic(ioc, func);
1237         rcu_read_unlock();
1238 }
1239
1240 static void cfq_cic_free_rcu(struct rcu_head *head)
1241 {
1242         struct cfq_io_context *cic;
1243
1244         cic = container_of(head, struct cfq_io_context, rcu_head);
1245
1246         kmem_cache_free(cfq_ioc_pool, cic);
1247         elv_ioc_count_dec(ioc_count);
1248
1249         if (ioc_gone) {
1250                 /*
1251                  * CFQ scheduler is exiting, grab exit lock and check
1252                  * the pending io context count. If it hits zero,
1253                  * complete ioc_gone and set it back to NULL
1254                  */
1255                 spin_lock(&ioc_gone_lock);
1256                 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1257                         complete(ioc_gone);
1258                         ioc_gone = NULL;
1259                 }
1260                 spin_unlock(&ioc_gone_lock);
1261         }
1262 }
1263
1264 static void cfq_cic_free(struct cfq_io_context *cic)
1265 {
1266         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1267 }
1268
1269 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1270 {
1271         unsigned long flags;
1272
1273         BUG_ON(!cic->dead_key);
1274
1275         spin_lock_irqsave(&ioc->lock, flags);
1276         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1277         hlist_del_rcu(&cic->cic_list);
1278         spin_unlock_irqrestore(&ioc->lock, flags);
1279
1280         cfq_cic_free(cic);
1281 }
1282
1283 /*
1284  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1285  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1286  * and ->trim() which is called with the task lock held
1287  */
1288 static void cfq_free_io_context(struct io_context *ioc)
1289 {
1290         /*
1291          * ioc->refcount is zero here, or we are called from elv_unregister(),
1292          * so no more cic's are allowed to be linked into this ioc.  So it
1293          * should be ok to iterate over the known list, we will see all cic's
1294          * since no new ones are added.
1295          */
1296         __call_for_each_cic(ioc, cic_free_func);
1297 }
1298
1299 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1300 {
1301         if (unlikely(cfqq == cfqd->active_queue)) {
1302                 __cfq_slice_expired(cfqd, cfqq, 0);
1303                 cfq_schedule_dispatch(cfqd);
1304         }
1305
1306         cfq_put_queue(cfqq);
1307 }
1308
1309 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1310                                          struct cfq_io_context *cic)
1311 {
1312         struct io_context *ioc = cic->ioc;
1313
1314         list_del_init(&cic->queue_list);
1315
1316         /*
1317          * Make sure key == NULL is seen for dead queues
1318          */
1319         smp_wmb();
1320         cic->dead_key = (unsigned long) cic->key;
1321         cic->key = NULL;
1322
1323         if (ioc->ioc_data == cic)
1324                 rcu_assign_pointer(ioc->ioc_data, NULL);
1325
1326         if (cic->cfqq[ASYNC]) {
1327                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1328                 cic->cfqq[ASYNC] = NULL;
1329         }
1330
1331         if (cic->cfqq[SYNC]) {
1332                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1333                 cic->cfqq[SYNC] = NULL;
1334         }
1335 }
1336
1337 static void cfq_exit_single_io_context(struct io_context *ioc,
1338                                        struct cfq_io_context *cic)
1339 {
1340         struct cfq_data *cfqd = cic->key;
1341
1342         if (cfqd) {
1343                 struct request_queue *q = cfqd->queue;
1344                 unsigned long flags;
1345
1346                 spin_lock_irqsave(q->queue_lock, flags);
1347
1348                 /*
1349                  * Ensure we get a fresh copy of the ->key to prevent
1350                  * race between exiting task and queue
1351                  */
1352                 smp_read_barrier_depends();
1353                 if (cic->key)
1354                         __cfq_exit_single_io_context(cfqd, cic);
1355
1356                 spin_unlock_irqrestore(q->queue_lock, flags);
1357         }
1358 }
1359
1360 /*
1361  * The process that ioc belongs to has exited, we need to clean up
1362  * and put the internal structures we have that belongs to that process.
1363  */
1364 static void cfq_exit_io_context(struct io_context *ioc)
1365 {
1366         call_for_each_cic(ioc, cfq_exit_single_io_context);
1367 }
1368
1369 static struct cfq_io_context *
1370 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1371 {
1372         struct cfq_io_context *cic;
1373
1374         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1375                                                         cfqd->queue->node);
1376         if (cic) {
1377                 cic->last_end_request = jiffies;
1378                 INIT_LIST_HEAD(&cic->queue_list);
1379                 INIT_HLIST_NODE(&cic->cic_list);
1380                 cic->dtor = cfq_free_io_context;
1381                 cic->exit = cfq_exit_io_context;
1382                 elv_ioc_count_inc(ioc_count);
1383         }
1384
1385         return cic;
1386 }
1387
1388 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1389 {
1390         struct task_struct *tsk = current;
1391         int ioprio_class;
1392
1393         if (!cfq_cfqq_prio_changed(cfqq))
1394                 return;
1395
1396         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1397         switch (ioprio_class) {
1398         default:
1399                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1400         case IOPRIO_CLASS_NONE:
1401                 /*
1402                  * no prio set, inherit CPU scheduling settings
1403                  */
1404                 cfqq->ioprio = task_nice_ioprio(tsk);
1405                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1406                 break;
1407         case IOPRIO_CLASS_RT:
1408                 cfqq->ioprio = task_ioprio(ioc);
1409                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1410                 break;
1411         case IOPRIO_CLASS_BE:
1412                 cfqq->ioprio = task_ioprio(ioc);
1413                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1414                 break;
1415         case IOPRIO_CLASS_IDLE:
1416                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1417                 cfqq->ioprio = 7;
1418                 cfq_clear_cfqq_idle_window(cfqq);
1419                 break;
1420         }
1421
1422         /*
1423          * keep track of original prio settings in case we have to temporarily
1424          * elevate the priority of this queue
1425          */
1426         cfqq->org_ioprio = cfqq->ioprio;
1427         cfqq->org_ioprio_class = cfqq->ioprio_class;
1428         cfq_clear_cfqq_prio_changed(cfqq);
1429 }
1430
1431 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1432 {
1433         struct cfq_data *cfqd = cic->key;
1434         struct cfq_queue *cfqq;
1435         unsigned long flags;
1436
1437         if (unlikely(!cfqd))
1438                 return;
1439
1440         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1441
1442         cfqq = cic->cfqq[ASYNC];
1443         if (cfqq) {
1444                 struct cfq_queue *new_cfqq;
1445                 new_cfqq = cfq_get_queue(cfqd, ASYNC, cic->ioc, GFP_ATOMIC);
1446                 if (new_cfqq) {
1447                         cic->cfqq[ASYNC] = new_cfqq;
1448                         cfq_put_queue(cfqq);
1449                 }
1450         }
1451
1452         cfqq = cic->cfqq[SYNC];
1453         if (cfqq)
1454                 cfq_mark_cfqq_prio_changed(cfqq);
1455
1456         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1457 }
1458
1459 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1460 {
1461         call_for_each_cic(ioc, changed_ioprio);
1462         ioc->ioprio_changed = 0;
1463 }
1464
1465 static struct cfq_queue *
1466 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1467                      struct io_context *ioc, gfp_t gfp_mask)
1468 {
1469         struct cfq_queue *cfqq, *new_cfqq = NULL;
1470         struct cfq_io_context *cic;
1471
1472 retry:
1473         cic = cfq_cic_lookup(cfqd, ioc);
1474         /* cic always exists here */
1475         cfqq = cic_to_cfqq(cic, is_sync);
1476
1477         if (!cfqq) {
1478                 if (new_cfqq) {
1479                         cfqq = new_cfqq;
1480                         new_cfqq = NULL;
1481                 } else if (gfp_mask & __GFP_WAIT) {
1482                         /*
1483                          * Inform the allocator of the fact that we will
1484                          * just repeat this allocation if it fails, to allow
1485                          * the allocator to do whatever it needs to attempt to
1486                          * free memory.
1487                          */
1488                         spin_unlock_irq(cfqd->queue->queue_lock);
1489                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1490                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1491                                         cfqd->queue->node);
1492                         spin_lock_irq(cfqd->queue->queue_lock);
1493                         goto retry;
1494                 } else {
1495                         cfqq = kmem_cache_alloc_node(cfq_pool,
1496                                         gfp_mask | __GFP_ZERO,
1497                                         cfqd->queue->node);
1498                         if (!cfqq)
1499                                 goto out;
1500                 }
1501
1502                 RB_CLEAR_NODE(&cfqq->rb_node);
1503                 INIT_LIST_HEAD(&cfqq->fifo);
1504
1505                 atomic_set(&cfqq->ref, 0);
1506                 cfqq->cfqd = cfqd;
1507
1508                 cfq_mark_cfqq_prio_changed(cfqq);
1509                 cfq_mark_cfqq_queue_new(cfqq);
1510
1511                 cfq_init_prio_data(cfqq, ioc);
1512
1513                 if (is_sync) {
1514                         if (!cfq_class_idle(cfqq))
1515                                 cfq_mark_cfqq_idle_window(cfqq);
1516                         cfq_mark_cfqq_sync(cfqq);
1517                 }
1518                 cfqq->pid = current->pid;
1519                 cfq_log_cfqq(cfqd, cfqq, "alloced");
1520         }
1521
1522         if (new_cfqq)
1523                 kmem_cache_free(cfq_pool, new_cfqq);
1524
1525 out:
1526         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1527         return cfqq;
1528 }
1529
1530 static struct cfq_queue **
1531 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1532 {
1533         switch (ioprio_class) {
1534         case IOPRIO_CLASS_RT:
1535                 return &cfqd->async_cfqq[0][ioprio];
1536         case IOPRIO_CLASS_BE:
1537                 return &cfqd->async_cfqq[1][ioprio];
1538         case IOPRIO_CLASS_IDLE:
1539                 return &cfqd->async_idle_cfqq;
1540         default:
1541                 BUG();
1542         }
1543 }
1544
1545 static struct cfq_queue *
1546 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1547               gfp_t gfp_mask)
1548 {
1549         const int ioprio = task_ioprio(ioc);
1550         const int ioprio_class = task_ioprio_class(ioc);
1551         struct cfq_queue **async_cfqq = NULL;
1552         struct cfq_queue *cfqq = NULL;
1553
1554         if (!is_sync) {
1555                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1556                 cfqq = *async_cfqq;
1557         }
1558
1559         if (!cfqq) {
1560                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1561                 if (!cfqq)
1562                         return NULL;
1563         }
1564
1565         /*
1566          * pin the queue now that it's allocated, scheduler exit will prune it
1567          */
1568         if (!is_sync && !(*async_cfqq)) {
1569                 atomic_inc(&cfqq->ref);
1570                 *async_cfqq = cfqq;
1571         }
1572
1573         atomic_inc(&cfqq->ref);
1574         return cfqq;
1575 }
1576
1577 /*
1578  * We drop cfq io contexts lazily, so we may find a dead one.
1579  */
1580 static void
1581 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1582                   struct cfq_io_context *cic)
1583 {
1584         unsigned long flags;
1585
1586         WARN_ON(!list_empty(&cic->queue_list));
1587
1588         spin_lock_irqsave(&ioc->lock, flags);
1589
1590         BUG_ON(ioc->ioc_data == cic);
1591
1592         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1593         hlist_del_rcu(&cic->cic_list);
1594         spin_unlock_irqrestore(&ioc->lock, flags);
1595
1596         cfq_cic_free(cic);
1597 }
1598
1599 static struct cfq_io_context *
1600 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1601 {
1602         struct cfq_io_context *cic;
1603         unsigned long flags;
1604         void *k;
1605
1606         if (unlikely(!ioc))
1607                 return NULL;
1608
1609         rcu_read_lock();
1610
1611         /*
1612          * we maintain a last-hit cache, to avoid browsing over the tree
1613          */
1614         cic = rcu_dereference(ioc->ioc_data);
1615         if (cic && cic->key == cfqd) {
1616                 rcu_read_unlock();
1617                 return cic;
1618         }
1619
1620         do {
1621                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1622                 rcu_read_unlock();
1623                 if (!cic)
1624                         break;
1625                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1626                 k = cic->key;
1627                 if (unlikely(!k)) {
1628                         cfq_drop_dead_cic(cfqd, ioc, cic);
1629                         rcu_read_lock();
1630                         continue;
1631                 }
1632
1633                 spin_lock_irqsave(&ioc->lock, flags);
1634                 rcu_assign_pointer(ioc->ioc_data, cic);
1635                 spin_unlock_irqrestore(&ioc->lock, flags);
1636                 break;
1637         } while (1);
1638
1639         return cic;
1640 }
1641
1642 /*
1643  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1644  * the process specific cfq io context when entered from the block layer.
1645  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1646  */
1647 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1648                         struct cfq_io_context *cic, gfp_t gfp_mask)
1649 {
1650         unsigned long flags;
1651         int ret;
1652
1653         ret = radix_tree_preload(gfp_mask);
1654         if (!ret) {
1655                 cic->ioc = ioc;
1656                 cic->key = cfqd;
1657
1658                 spin_lock_irqsave(&ioc->lock, flags);
1659                 ret = radix_tree_insert(&ioc->radix_root,
1660                                                 (unsigned long) cfqd, cic);
1661                 if (!ret)
1662                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1663                 spin_unlock_irqrestore(&ioc->lock, flags);
1664
1665                 radix_tree_preload_end();
1666
1667                 if (!ret) {
1668                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1669                         list_add(&cic->queue_list, &cfqd->cic_list);
1670                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1671                 }
1672         }
1673
1674         if (ret)
1675                 printk(KERN_ERR "cfq: cic link failed!\n");
1676
1677         return ret;
1678 }
1679
1680 /*
1681  * Setup general io context and cfq io context. There can be several cfq
1682  * io contexts per general io context, if this process is doing io to more
1683  * than one device managed by cfq.
1684  */
1685 static struct cfq_io_context *
1686 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1687 {
1688         struct io_context *ioc = NULL;
1689         struct cfq_io_context *cic;
1690
1691         might_sleep_if(gfp_mask & __GFP_WAIT);
1692
1693         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1694         if (!ioc)
1695                 return NULL;
1696
1697         cic = cfq_cic_lookup(cfqd, ioc);
1698         if (cic)
1699                 goto out;
1700
1701         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1702         if (cic == NULL)
1703                 goto err;
1704
1705         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1706                 goto err_free;
1707
1708 out:
1709         smp_read_barrier_depends();
1710         if (unlikely(ioc->ioprio_changed))
1711                 cfq_ioc_set_ioprio(ioc);
1712
1713         return cic;
1714 err_free:
1715         cfq_cic_free(cic);
1716 err:
1717         put_io_context(ioc);
1718         return NULL;
1719 }
1720
1721 static void
1722 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1723 {
1724         unsigned long elapsed = jiffies - cic->last_end_request;
1725         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1726
1727         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1728         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1729         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1730 }
1731
1732 static void
1733 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1734                        struct request *rq)
1735 {
1736         sector_t sdist;
1737         u64 total;
1738
1739         if (cic->last_request_pos < rq->sector)
1740                 sdist = rq->sector - cic->last_request_pos;
1741         else
1742                 sdist = cic->last_request_pos - rq->sector;
1743
1744         /*
1745          * Don't allow the seek distance to get too large from the
1746          * odd fragment, pagein, etc
1747          */
1748         if (cic->seek_samples <= 60) /* second&third seek */
1749                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1750         else
1751                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1752
1753         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1754         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1755         total = cic->seek_total + (cic->seek_samples/2);
1756         do_div(total, cic->seek_samples);
1757         cic->seek_mean = (sector_t)total;
1758 }
1759
1760 /*
1761  * Disable idle window if the process thinks too long or seeks so much that
1762  * it doesn't matter
1763  */
1764 static void
1765 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1766                        struct cfq_io_context *cic)
1767 {
1768         int old_idle, enable_idle;
1769
1770         /*
1771          * Don't idle for async or idle io prio class
1772          */
1773         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1774                 return;
1775
1776         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1777
1778         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1779             (cfqd->hw_tag && CIC_SEEKY(cic)))
1780                 enable_idle = 0;
1781         else if (sample_valid(cic->ttime_samples)) {
1782                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1783                         enable_idle = 0;
1784                 else
1785                         enable_idle = 1;
1786         }
1787
1788         if (old_idle != enable_idle) {
1789                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1790                 if (enable_idle)
1791                         cfq_mark_cfqq_idle_window(cfqq);
1792                 else
1793                         cfq_clear_cfqq_idle_window(cfqq);
1794         }
1795 }
1796
1797 /*
1798  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1799  * no or if we aren't sure, a 1 will cause a preempt.
1800  */
1801 static int
1802 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1803                    struct request *rq)
1804 {
1805         struct cfq_queue *cfqq;
1806
1807         cfqq = cfqd->active_queue;
1808         if (!cfqq)
1809                 return 0;
1810
1811         if (cfq_slice_used(cfqq))
1812                 return 1;
1813
1814         if (cfq_class_idle(new_cfqq))
1815                 return 0;
1816
1817         if (cfq_class_idle(cfqq))
1818                 return 1;
1819
1820         /*
1821          * if the new request is sync, but the currently running queue is
1822          * not, let the sync request have priority.
1823          */
1824         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1825                 return 1;
1826
1827         /*
1828          * So both queues are sync. Let the new request get disk time if
1829          * it's a metadata request and the current queue is doing regular IO.
1830          */
1831         if (rq_is_meta(rq) && !cfqq->meta_pending)
1832                 return 1;
1833
1834         /*
1835          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1836          */
1837         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1838                 return 1;
1839
1840         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1841                 return 0;
1842
1843         /*
1844          * if this request is as-good as one we would expect from the
1845          * current cfqq, let it preempt
1846          */
1847         if (cfq_rq_close(cfqd, rq))
1848                 return 1;
1849
1850         return 0;
1851 }
1852
1853 /*
1854  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1855  * let it have half of its nominal slice.
1856  */
1857 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1858 {
1859         cfq_log_cfqq(cfqd, cfqq, "preempt");
1860         cfq_slice_expired(cfqd, 1);
1861
1862         /*
1863          * Put the new queue at the front of the of the current list,
1864          * so we know that it will be selected next.
1865          */
1866         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1867
1868         cfq_service_tree_add(cfqd, cfqq, 1);
1869
1870         cfqq->slice_end = 0;
1871         cfq_mark_cfqq_slice_new(cfqq);
1872 }
1873
1874 /*
1875  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1876  * something we should do about it
1877  */
1878 static void
1879 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1880                 struct request *rq)
1881 {
1882         struct cfq_io_context *cic = RQ_CIC(rq);
1883
1884         cfqd->rq_queued++;
1885         if (rq_is_meta(rq))
1886                 cfqq->meta_pending++;
1887
1888         cfq_update_io_thinktime(cfqd, cic);
1889         cfq_update_io_seektime(cfqd, cic, rq);
1890         cfq_update_idle_window(cfqd, cfqq, cic);
1891
1892         cic->last_request_pos = rq->sector + rq->nr_sectors;
1893
1894         if (cfqq == cfqd->active_queue) {
1895                 /*
1896                  * if we are waiting for a request for this queue, let it rip
1897                  * immediately and flag that we must not expire this queue
1898                  * just now
1899                  */
1900                 if (cfq_cfqq_wait_request(cfqq)) {
1901                         cfq_mark_cfqq_must_dispatch(cfqq);
1902                         del_timer(&cfqd->idle_slice_timer);
1903                         blk_start_queueing(cfqd->queue);
1904                 }
1905         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1906                 /*
1907                  * not the active queue - expire current slice if it is
1908                  * idle and has expired it's mean thinktime or this new queue
1909                  * has some old slice time left and is of higher priority or
1910                  * this new queue is RT and the current one is BE
1911                  */
1912                 cfq_preempt_queue(cfqd, cfqq);
1913                 cfq_mark_cfqq_must_dispatch(cfqq);
1914                 blk_start_queueing(cfqd->queue);
1915         }
1916 }
1917
1918 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1919 {
1920         struct cfq_data *cfqd = q->elevator->elevator_data;
1921         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1922
1923         cfq_log_cfqq(cfqd, cfqq, "insert_request");
1924         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1925
1926         cfq_add_rq_rb(rq);
1927
1928         list_add_tail(&rq->queuelist, &cfqq->fifo);
1929
1930         cfq_rq_enqueued(cfqd, cfqq, rq);
1931 }
1932
1933 /*
1934  * Update hw_tag based on peak queue depth over 50 samples under
1935  * sufficient load.
1936  */
1937 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1938 {
1939         if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1940                 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1941
1942         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1943             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1944                 return;
1945
1946         if (cfqd->hw_tag_samples++ < 50)
1947                 return;
1948
1949         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1950                 cfqd->hw_tag = 1;
1951         else
1952                 cfqd->hw_tag = 0;
1953
1954         cfqd->hw_tag_samples = 0;
1955         cfqd->rq_in_driver_peak = 0;
1956 }
1957
1958 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1959 {
1960         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1961         struct cfq_data *cfqd = cfqq->cfqd;
1962         const int sync = rq_is_sync(rq);
1963         unsigned long now;
1964
1965         now = jiffies;
1966         cfq_log_cfqq(cfqd, cfqq, "complete");
1967
1968         cfq_update_hw_tag(cfqd);
1969
1970         WARN_ON(!cfqd->rq_in_driver);
1971         WARN_ON(!cfqq->dispatched);
1972         cfqd->rq_in_driver--;
1973         cfqq->dispatched--;
1974
1975         if (cfq_cfqq_sync(cfqq))
1976                 cfqd->sync_flight--;
1977
1978         if (!cfq_class_idle(cfqq))
1979                 cfqd->last_end_request = now;
1980
1981         if (sync)
1982                 RQ_CIC(rq)->last_end_request = now;
1983
1984         /*
1985          * If this is the active queue, check if it needs to be expired,
1986          * or if we want to idle in case it has no pending requests.
1987          */
1988         if (cfqd->active_queue == cfqq) {
1989                 if (cfq_cfqq_slice_new(cfqq)) {
1990                         cfq_set_prio_slice(cfqd, cfqq);
1991                         cfq_clear_cfqq_slice_new(cfqq);
1992                 }
1993                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
1994                         cfq_slice_expired(cfqd, 1);
1995                 else if (sync && !rq_noidle(rq) &&
1996                          RB_EMPTY_ROOT(&cfqq->sort_list)) {
1997                         cfq_arm_slice_timer(cfqd);
1998                 }
1999         }
2000
2001         if (!cfqd->rq_in_driver)
2002                 cfq_schedule_dispatch(cfqd);
2003 }
2004
2005 /*
2006  * we temporarily boost lower priority queues if they are holding fs exclusive
2007  * resources. they are boosted to normal prio (CLASS_BE/4)
2008  */
2009 static void cfq_prio_boost(struct cfq_queue *cfqq)
2010 {
2011         if (has_fs_excl()) {
2012                 /*
2013                  * boost idle prio on transactions that would lock out other
2014                  * users of the filesystem
2015                  */
2016                 if (cfq_class_idle(cfqq))
2017                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2018                 if (cfqq->ioprio > IOPRIO_NORM)
2019                         cfqq->ioprio = IOPRIO_NORM;
2020         } else {
2021                 /*
2022                  * check if we need to unboost the queue
2023                  */
2024                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2025                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2026                 if (cfqq->ioprio != cfqq->org_ioprio)
2027                         cfqq->ioprio = cfqq->org_ioprio;
2028         }
2029 }
2030
2031 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2032 {
2033         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2034             !cfq_cfqq_must_alloc_slice(cfqq)) {
2035                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2036                 return ELV_MQUEUE_MUST;
2037         }
2038
2039         return ELV_MQUEUE_MAY;
2040 }
2041
2042 static int cfq_may_queue(struct request_queue *q, int rw)
2043 {
2044         struct cfq_data *cfqd = q->elevator->elevator_data;
2045         struct task_struct *tsk = current;
2046         struct cfq_io_context *cic;
2047         struct cfq_queue *cfqq;
2048
2049         /*
2050          * don't force setup of a queue from here, as a call to may_queue
2051          * does not necessarily imply that a request actually will be queued.
2052          * so just lookup a possibly existing queue, or return 'may queue'
2053          * if that fails
2054          */
2055         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2056         if (!cic)
2057                 return ELV_MQUEUE_MAY;
2058
2059         cfqq = cic_to_cfqq(cic, rw & REQ_RW_SYNC);
2060         if (cfqq) {
2061                 cfq_init_prio_data(cfqq, cic->ioc);
2062                 cfq_prio_boost(cfqq);
2063
2064                 return __cfq_may_queue(cfqq);
2065         }
2066
2067         return ELV_MQUEUE_MAY;
2068 }
2069
2070 /*
2071  * queue lock held here
2072  */
2073 static void cfq_put_request(struct request *rq)
2074 {
2075         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2076
2077         if (cfqq) {
2078                 const int rw = rq_data_dir(rq);
2079
2080                 BUG_ON(!cfqq->allocated[rw]);
2081                 cfqq->allocated[rw]--;
2082
2083                 put_io_context(RQ_CIC(rq)->ioc);
2084
2085                 rq->elevator_private = NULL;
2086                 rq->elevator_private2 = NULL;
2087
2088                 cfq_put_queue(cfqq);
2089         }
2090 }
2091
2092 /*
2093  * Allocate cfq data structures associated with this request.
2094  */
2095 static int
2096 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2097 {
2098         struct cfq_data *cfqd = q->elevator->elevator_data;
2099         struct cfq_io_context *cic;
2100         const int rw = rq_data_dir(rq);
2101         const int is_sync = rq_is_sync(rq);
2102         struct cfq_queue *cfqq;
2103         unsigned long flags;
2104
2105         might_sleep_if(gfp_mask & __GFP_WAIT);
2106
2107         cic = cfq_get_io_context(cfqd, gfp_mask);
2108
2109         spin_lock_irqsave(q->queue_lock, flags);
2110
2111         if (!cic)
2112                 goto queue_fail;
2113
2114         cfqq = cic_to_cfqq(cic, is_sync);
2115         if (!cfqq) {
2116                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2117
2118                 if (!cfqq)
2119                         goto queue_fail;
2120
2121                 cic_set_cfqq(cic, cfqq, is_sync);
2122         }
2123
2124         cfqq->allocated[rw]++;
2125         cfq_clear_cfqq_must_alloc(cfqq);
2126         atomic_inc(&cfqq->ref);
2127
2128         spin_unlock_irqrestore(q->queue_lock, flags);
2129
2130         rq->elevator_private = cic;
2131         rq->elevator_private2 = cfqq;
2132         return 0;
2133
2134 queue_fail:
2135         if (cic)
2136                 put_io_context(cic->ioc);
2137
2138         cfq_schedule_dispatch(cfqd);
2139         spin_unlock_irqrestore(q->queue_lock, flags);
2140         cfq_log(cfqd, "set_request fail");
2141         return 1;
2142 }
2143
2144 static void cfq_kick_queue(struct work_struct *work)
2145 {
2146         struct cfq_data *cfqd =
2147                 container_of(work, struct cfq_data, unplug_work);
2148         struct request_queue *q = cfqd->queue;
2149         unsigned long flags;
2150
2151         spin_lock_irqsave(q->queue_lock, flags);
2152         blk_start_queueing(q);
2153         spin_unlock_irqrestore(q->queue_lock, flags);
2154 }
2155
2156 /*
2157  * Timer running if the active_queue is currently idling inside its time slice
2158  */
2159 static void cfq_idle_slice_timer(unsigned long data)
2160 {
2161         struct cfq_data *cfqd = (struct cfq_data *) data;
2162         struct cfq_queue *cfqq;
2163         unsigned long flags;
2164         int timed_out = 1;
2165
2166         cfq_log(cfqd, "idle timer fired");
2167
2168         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2169
2170         cfqq = cfqd->active_queue;
2171         if (cfqq) {
2172                 timed_out = 0;
2173
2174                 /*
2175                  * expired
2176                  */
2177                 if (cfq_slice_used(cfqq))
2178                         goto expire;
2179
2180                 /*
2181                  * only expire and reinvoke request handler, if there are
2182                  * other queues with pending requests
2183                  */
2184                 if (!cfqd->busy_queues)
2185                         goto out_cont;
2186
2187                 /*
2188                  * not expired and it has a request pending, let it dispatch
2189                  */
2190                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
2191                         cfq_mark_cfqq_must_dispatch(cfqq);
2192                         goto out_kick;
2193                 }
2194         }
2195 expire:
2196         cfq_slice_expired(cfqd, timed_out);
2197 out_kick:
2198         cfq_schedule_dispatch(cfqd);
2199 out_cont:
2200         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2201 }
2202
2203 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2204 {
2205         del_timer_sync(&cfqd->idle_slice_timer);
2206         cancel_work_sync(&cfqd->unplug_work);
2207 }
2208
2209 static void cfq_put_async_queues(struct cfq_data *cfqd)
2210 {
2211         int i;
2212
2213         for (i = 0; i < IOPRIO_BE_NR; i++) {
2214                 if (cfqd->async_cfqq[0][i])
2215                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2216                 if (cfqd->async_cfqq[1][i])
2217                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2218         }
2219
2220         if (cfqd->async_idle_cfqq)
2221                 cfq_put_queue(cfqd->async_idle_cfqq);
2222 }
2223
2224 static void cfq_exit_queue(struct elevator_queue *e)
2225 {
2226         struct cfq_data *cfqd = e->elevator_data;
2227         struct request_queue *q = cfqd->queue;
2228
2229         cfq_shutdown_timer_wq(cfqd);
2230
2231         spin_lock_irq(q->queue_lock);
2232
2233         if (cfqd->active_queue)
2234                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2235
2236         while (!list_empty(&cfqd->cic_list)) {
2237                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2238                                                         struct cfq_io_context,
2239                                                         queue_list);
2240
2241                 __cfq_exit_single_io_context(cfqd, cic);
2242         }
2243
2244         cfq_put_async_queues(cfqd);
2245
2246         spin_unlock_irq(q->queue_lock);
2247
2248         cfq_shutdown_timer_wq(cfqd);
2249
2250         kfree(cfqd);
2251 }
2252
2253 static void *cfq_init_queue(struct request_queue *q)
2254 {
2255         struct cfq_data *cfqd;
2256
2257         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2258         if (!cfqd)
2259                 return NULL;
2260
2261         cfqd->service_tree = CFQ_RB_ROOT;
2262         INIT_LIST_HEAD(&cfqd->cic_list);
2263
2264         cfqd->queue = q;
2265
2266         init_timer(&cfqd->idle_slice_timer);
2267         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2268         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2269
2270         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2271
2272         cfqd->last_end_request = jiffies;
2273         cfqd->cfq_quantum = cfq_quantum;
2274         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2275         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2276         cfqd->cfq_back_max = cfq_back_max;
2277         cfqd->cfq_back_penalty = cfq_back_penalty;
2278         cfqd->cfq_slice[0] = cfq_slice_async;
2279         cfqd->cfq_slice[1] = cfq_slice_sync;
2280         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2281         cfqd->cfq_slice_idle = cfq_slice_idle;
2282         cfqd->hw_tag = 1;
2283
2284         return cfqd;
2285 }
2286
2287 static void cfq_slab_kill(void)
2288 {
2289         /*
2290          * Caller already ensured that pending RCU callbacks are completed,
2291          * so we should have no busy allocations at this point.
2292          */
2293         if (cfq_pool)
2294                 kmem_cache_destroy(cfq_pool);
2295         if (cfq_ioc_pool)
2296                 kmem_cache_destroy(cfq_ioc_pool);
2297 }
2298
2299 static int __init cfq_slab_setup(void)
2300 {
2301         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2302         if (!cfq_pool)
2303                 goto fail;
2304
2305         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2306         if (!cfq_ioc_pool)
2307                 goto fail;
2308
2309         return 0;
2310 fail:
2311         cfq_slab_kill();
2312         return -ENOMEM;
2313 }
2314
2315 /*
2316  * sysfs parts below -->
2317  */
2318 static ssize_t
2319 cfq_var_show(unsigned int var, char *page)
2320 {
2321         return sprintf(page, "%d\n", var);
2322 }
2323
2324 static ssize_t
2325 cfq_var_store(unsigned int *var, const char *page, size_t count)
2326 {
2327         char *p = (char *) page;
2328
2329         *var = simple_strtoul(p, &p, 10);
2330         return count;
2331 }
2332
2333 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2334 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2335 {                                                                       \
2336         struct cfq_data *cfqd = e->elevator_data;                       \
2337         unsigned int __data = __VAR;                                    \
2338         if (__CONV)                                                     \
2339                 __data = jiffies_to_msecs(__data);                      \
2340         return cfq_var_show(__data, (page));                            \
2341 }
2342 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2343 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2344 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2345 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2346 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2347 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2348 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2349 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2350 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2351 #undef SHOW_FUNCTION
2352
2353 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2354 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2355 {                                                                       \
2356         struct cfq_data *cfqd = e->elevator_data;                       \
2357         unsigned int __data;                                            \
2358         int ret = cfq_var_store(&__data, (page), count);                \
2359         if (__data < (MIN))                                             \
2360                 __data = (MIN);                                         \
2361         else if (__data > (MAX))                                        \
2362                 __data = (MAX);                                         \
2363         if (__CONV)                                                     \
2364                 *(__PTR) = msecs_to_jiffies(__data);                    \
2365         else                                                            \
2366                 *(__PTR) = __data;                                      \
2367         return ret;                                                     \
2368 }
2369 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2370 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2371                 UINT_MAX, 1);
2372 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2373                 UINT_MAX, 1);
2374 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2375 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2376                 UINT_MAX, 0);
2377 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2378 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2379 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2380 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2381                 UINT_MAX, 0);
2382 #undef STORE_FUNCTION
2383
2384 #define CFQ_ATTR(name) \
2385         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2386
2387 static struct elv_fs_entry cfq_attrs[] = {
2388         CFQ_ATTR(quantum),
2389         CFQ_ATTR(fifo_expire_sync),
2390         CFQ_ATTR(fifo_expire_async),
2391         CFQ_ATTR(back_seek_max),
2392         CFQ_ATTR(back_seek_penalty),
2393         CFQ_ATTR(slice_sync),
2394         CFQ_ATTR(slice_async),
2395         CFQ_ATTR(slice_async_rq),
2396         CFQ_ATTR(slice_idle),
2397         __ATTR_NULL
2398 };
2399
2400 static struct elevator_type iosched_cfq = {
2401         .ops = {
2402                 .elevator_merge_fn =            cfq_merge,
2403                 .elevator_merged_fn =           cfq_merged_request,
2404                 .elevator_merge_req_fn =        cfq_merged_requests,
2405                 .elevator_allow_merge_fn =      cfq_allow_merge,
2406                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2407                 .elevator_add_req_fn =          cfq_insert_request,
2408                 .elevator_activate_req_fn =     cfq_activate_request,
2409                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2410                 .elevator_queue_empty_fn =      cfq_queue_empty,
2411                 .elevator_completed_req_fn =    cfq_completed_request,
2412                 .elevator_former_req_fn =       elv_rb_former_request,
2413                 .elevator_latter_req_fn =       elv_rb_latter_request,
2414                 .elevator_set_req_fn =          cfq_set_request,
2415                 .elevator_put_req_fn =          cfq_put_request,
2416                 .elevator_may_queue_fn =        cfq_may_queue,
2417                 .elevator_init_fn =             cfq_init_queue,
2418                 .elevator_exit_fn =             cfq_exit_queue,
2419                 .trim =                         cfq_free_io_context,
2420         },
2421         .elevator_attrs =       cfq_attrs,
2422         .elevator_name =        "cfq",
2423         .elevator_owner =       THIS_MODULE,
2424 };
2425
2426 static int __init cfq_init(void)
2427 {
2428         /*
2429          * could be 0 on HZ < 1000 setups
2430          */
2431         if (!cfq_slice_async)
2432                 cfq_slice_async = 1;
2433         if (!cfq_slice_idle)
2434                 cfq_slice_idle = 1;
2435
2436         if (cfq_slab_setup())
2437                 return -ENOMEM;
2438
2439         elv_register(&iosched_cfq);
2440
2441         return 0;
2442 }
2443
2444 static void __exit cfq_exit(void)
2445 {
2446         DECLARE_COMPLETION_ONSTACK(all_gone);
2447         elv_unregister(&iosched_cfq);
2448         ioc_gone = &all_gone;
2449         /* ioc_gone's update must be visible before reading ioc_count */
2450         smp_wmb();
2451
2452         /*
2453          * this also protects us from entering cfq_slab_kill() with
2454          * pending RCU callbacks
2455          */
2456         if (elv_ioc_count_read(ioc_count))
2457                 wait_for_completion(&all_gone);
2458         cfq_slab_kill();
2459 }
2460
2461 module_init(cfq_init);
2462 module_exit(cfq_exit);
2463
2464 MODULE_AUTHOR("Jens Axboe");
2465 MODULE_LICENSE("GPL");
2466 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");