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/slab.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/jiffies.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16 #include <linux/blktrace_api.h>
  17 #include "blk.h"
  18 #include "blk-cgroup.h"
  19
  20 /*
  21  * tunables
  22  */
  23 /* max queue in one round of service */
  24 static const int cfq_quantum = 8;
  25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  26 /* maximum backwards seek, in KiB */
  27 static const int cfq_back_max = 16 * 1024;
  28 /* penalty of a backwards seek */
  29 static const int cfq_back_penalty = 2;
  30 static const int cfq_slice_sync = HZ / 10;
  31 static int cfq_slice_async = HZ / 25;
  32 static const int cfq_slice_async_rq = 2;
  33 static int cfq_slice_idle = HZ / 125;
  34 static int cfq_group_idle = HZ / 125;
  35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
  36 static const int cfq_hist_divisor = 4;
  37
  38 /*
  39  * offset from end of service tree
  40  */
  41 #define CFQ_IDLE_DELAY          (HZ / 5)
  42
  43 /*
  44  * below this threshold, we consider thinktime immediate
  45  */
  46 #define CFQ_MIN_TT              (2)
  47
  48 #define CFQ_SLICE_SCALE         (5)
  49 #define CFQ_HW_QUEUE_MIN        (5)
  50 #define CFQ_SERVICE_SHIFT       12
  51
  52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  56
  57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  60
  61 static struct kmem_cache *cfq_pool;
  62
  63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  69
  70 struct cfq_ttime {
  71         unsigned long last_end_request;
  72
  73         unsigned long ttime_total;
  74         unsigned long ttime_samples;
  75         unsigned long ttime_mean;
  76 };
  77
  78 /*
  79  * Most of our rbtree usage is for sorting with min extraction, so
  80  * if we cache the leftmost node we don't have to walk down the tree
  81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  82  * move this into the elevator for the rq sorting as well.
  83  */
  84 struct cfq_rb_root {
  85         struct rb_root rb;
  86         struct rb_node *left;
  87         unsigned count;
  88         u64 min_vdisktime;
  89         struct cfq_ttime ttime;
  90 };
  91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  92                         .ttime = {.last_end_request = jiffies,},}
  93
  94 /*
  95  * Per process-grouping structure
  96  */
  97 struct cfq_queue {
  98         /* reference count */
  99         int ref;
 100         /* various state flags, see below */
 101         unsigned int flags;
 102         /* parent cfq_data */
 103         struct cfq_data *cfqd;
 104         /* service_tree member */
 105         struct rb_node rb_node;
 106         /* service_tree key */
 107         unsigned long rb_key;
 108         /* prio tree member */
 109         struct rb_node p_node;
 110         /* prio tree root we belong to, if any */
 111         struct rb_root *p_root;
 112         /* sorted list of pending requests */
 113         struct rb_root sort_list;
 114         /* if fifo isn't expired, next request to serve */
 115         struct request *next_rq;
 116         /* requests queued in sort_list */
 117         int queued[2];
 118         /* currently allocated requests */
 119         int allocated[2];
 120         /* fifo list of requests in sort_list */
 121         struct list_head fifo;
 122
 123         /* time when queue got scheduled in to dispatch first request. */
 124         unsigned long dispatch_start;
 125         unsigned int allocated_slice;
 126         unsigned int slice_dispatch;
 127         /* time when first request from queue completed and slice started. */
 128         unsigned long slice_start;
 129         unsigned long slice_end;
 130         long slice_resid;
 131
 132         /* pending priority requests */
 133         int prio_pending;
 134         /* number of requests that are on the dispatch list or inside driver */
 135         int dispatched;
 136
 137         /* io prio of this group */
 138         unsigned short ioprio, org_ioprio;
 139         unsigned short ioprio_class;
 140
 141         pid_t pid;
 142
 143         u32 seek_history;
 144         sector_t last_request_pos;
 145
 146         struct cfq_rb_root *service_tree;
 147         struct cfq_queue *new_cfqq;
 148         struct cfq_group *cfqg;
 149         /* Number of sectors dispatched from queue in single dispatch round */
 150         unsigned long nr_sectors;
 151 };
 152
 153 /*
 154  * First index in the service_trees.
 155  * IDLE is handled separately, so it has negative index
 156  */
 157 enum wl_class_t {
 158         BE_WORKLOAD = 0,
 159         RT_WORKLOAD = 1,
 160         IDLE_WORKLOAD = 2,
 161         CFQ_PRIO_NR,
 162 };
 163
 164 /*
 165  * Second index in the service_trees.
 166  */
 167 enum wl_type_t {
 168         ASYNC_WORKLOAD = 0,
 169         SYNC_NOIDLE_WORKLOAD = 1,
 170         SYNC_WORKLOAD = 2
 171 };
 172
 173 struct cfqg_stats {
 174 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 175         /* total bytes transferred */
 176         struct blkg_rwstat              service_bytes;
 177         /* total IOs serviced, post merge */
 178         struct blkg_rwstat              serviced;
 179         /* number of ios merged */
 180         struct blkg_rwstat              merged;
 181         /* total time spent on device in ns, may not be accurate w/ queueing */
 182         struct blkg_rwstat              service_time;
 183         /* total time spent waiting in scheduler queue in ns */
 184         struct blkg_rwstat              wait_time;
 185         /* number of IOs queued up */
 186         struct blkg_rwstat              queued;
 187         /* total sectors transferred */
 188         struct blkg_stat                sectors;
 189         /* total disk time and nr sectors dispatched by this group */
 190         struct blkg_stat                time;
 191 #ifdef CONFIG_DEBUG_BLK_CGROUP
 192         /* time not charged to this cgroup */
 193         struct blkg_stat                unaccounted_time;
 194         /* sum of number of ios queued across all samples */
 195         struct blkg_stat                avg_queue_size_sum;
 196         /* count of samples taken for average */
 197         struct blkg_stat                avg_queue_size_samples;
 198         /* how many times this group has been removed from service tree */
 199         struct blkg_stat                dequeue;
 200         /* total time spent waiting for it to be assigned a timeslice. */
 201         struct blkg_stat                group_wait_time;
 202         /* time spent idling for this blkcg_gq */
 203         struct blkg_stat                idle_time;
 204         /* total time with empty current active q with other requests queued */
 205         struct blkg_stat                empty_time;
 206         /* fields after this shouldn't be cleared on stat reset */
 207         uint64_t                        start_group_wait_time;
 208         uint64_t                        start_idle_time;
 209         uint64_t                        start_empty_time;
 210         uint16_t                        flags;
 211 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 212 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 213 };
 214
 215 /* This is per cgroup per device grouping structure */
 216 struct cfq_group {
 217         /* must be the first member */
 218         struct blkg_policy_data pd;
 219
 220         /* group service_tree member */
 221         struct rb_node rb_node;
 222
 223         /* group service_tree key */
 224         u64 vdisktime;
 225
 226         /*
 227          * The number of active cfqgs and sum of their weights under this
 228          * cfqg.  This covers this cfqg's leaf_weight and all children's
 229          * weights, but does not cover weights of further descendants.
 230          *
 231          * If a cfqg is on the service tree, it's active.  An active cfqg
 232          * also activates its parent and contributes to the children_weight
 233          * of the parent.
 234          */
 235         int nr_active;
 236         unsigned int children_weight;
 237
 238         /*
 239          * vfraction is the fraction of vdisktime that the tasks in this
 240          * cfqg are entitled to.  This is determined by compounding the
 241          * ratios walking up from this cfqg to the root.
 242          *
 243          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 244          * vfractions on a service tree is approximately 1.  The sum may
 245          * deviate a bit due to rounding errors and fluctuations caused by
 246          * cfqgs entering and leaving the service tree.
 247          */
 248         unsigned int vfraction;
 249
 250         /*
 251          * There are two weights - (internal) weight is the weight of this
 252          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 253          * this cfqg against the child cfqgs.  For the root cfqg, both
 254          * weights are kept in sync for backward compatibility.
 255          */
 256         unsigned int weight;
 257         unsigned int new_weight;
 258         unsigned int dev_weight;
 259
 260         unsigned int leaf_weight;
 261         unsigned int new_leaf_weight;
 262         unsigned int dev_leaf_weight;
 263
 264         /* number of cfqq currently on this group */
 265         int nr_cfqq;
 266
 267         /*
 268          * Per group busy queues average. Useful for workload slice calc. We
 269          * create the array for each prio class but at run time it is used
 270          * only for RT and BE class and slot for IDLE class remains unused.
 271          * This is primarily done to avoid confusion and a gcc warning.
 272          */
 273         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 274         /*
 275          * rr lists of queues with requests. We maintain service trees for
 276          * RT and BE classes. These trees are subdivided in subclasses
 277          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 278          * class there is no subclassification and all the cfq queues go on
 279          * a single tree service_tree_idle.
 280          * Counts are embedded in the cfq_rb_root
 281          */
 282         struct cfq_rb_root service_trees[2][3];
 283         struct cfq_rb_root service_tree_idle;
 284
 285         unsigned long saved_wl_slice;
 286         enum wl_type_t saved_wl_type;
 287         enum wl_class_t saved_wl_class;
 288
 289         /* number of requests that are on the dispatch list or inside driver */
 290         int dispatched;
 291         struct cfq_ttime ttime;
 292         struct cfqg_stats stats;        /* stats for this cfqg */
 293         struct cfqg_stats dead_stats;   /* stats pushed from dead children */
 294 };
 295
 296 struct cfq_io_cq {
 297         struct io_cq            icq;            /* must be the first member */
 298         struct cfq_queue        *cfqq[2];
 299         struct cfq_ttime        ttime;
 300         int                     ioprio;         /* the current ioprio */
 301 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 302         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 303 #endif
 304 };
 305
 306 /*
 307  * Per block device queue structure
 308  */
 309 struct cfq_data {
 310         struct request_queue *queue;
 311         /* Root service tree for cfq_groups */
 312         struct cfq_rb_root grp_service_tree;
 313         struct cfq_group *root_group;
 314
 315         /*
 316          * The priority currently being served
 317          */
 318         enum wl_class_t serving_wl_class;
 319         enum wl_type_t serving_wl_type;
 320         unsigned long workload_expires;
 321         struct cfq_group *serving_group;
 322
 323         /*
 324          * Each priority tree is sorted by next_request position.  These
 325          * trees are used when determining if two or more queues are
 326          * interleaving requests (see cfq_close_cooperator).
 327          */
 328         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 329
 330         unsigned int busy_queues;
 331         unsigned int busy_sync_queues;
 332
 333         int rq_in_driver;
 334         int rq_in_flight[2];
 335
 336         /*
 337          * queue-depth detection
 338          */
 339         int rq_queued;
 340         int hw_tag;
 341         /*
 342          * hw_tag can be
 343          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 344          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 345          *  0 => no NCQ
 346          */
 347         int hw_tag_est_depth;
 348         unsigned int hw_tag_samples;
 349
 350         /*
 351          * idle window management
 352          */
 353         struct timer_list idle_slice_timer;
 354         struct work_struct unplug_work;
 355
 356         struct cfq_queue *active_queue;
 357         struct cfq_io_cq *active_cic;
 358
 359         /*
 360          * async queue for each priority case
 361          */
 362         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 363         struct cfq_queue *async_idle_cfqq;
 364
 365         sector_t last_position;
 366
 367         /*
 368          * tunables, see top of file
 369          */
 370         unsigned int cfq_quantum;
 371         unsigned int cfq_fifo_expire[2];
 372         unsigned int cfq_back_penalty;
 373         unsigned int cfq_back_max;
 374         unsigned int cfq_slice[2];
 375         unsigned int cfq_slice_async_rq;
 376         unsigned int cfq_slice_idle;
 377         unsigned int cfq_group_idle;
 378         unsigned int cfq_latency;
 379         unsigned int cfq_target_latency;
 380
 381         /*
 382          * Fallback dummy cfqq for extreme OOM conditions
 383          */
 384         struct cfq_queue oom_cfqq;
 385
 386         unsigned long last_delayed_sync;
 387 };
 388
 389 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 390
 391 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 392                                             enum wl_class_t class,
 393                                             enum wl_type_t type)
 394 {
 395         if (!cfqg)
 396                 return NULL;
 397
 398         if (class == IDLE_WORKLOAD)
 399                 return &cfqg->service_tree_idle;
 400
 401         return &cfqg->service_trees[class][type];
 402 }
 403
 404 enum cfqq_state_flags {
 405         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 406         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 407         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 408         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 409         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 410         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 411         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 412         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 413         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 414         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 415         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 416         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 417         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 418 };
 419
 420 #define CFQ_CFQQ_FNS(name)                                              \
 421 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 422 {                                                                       \
 423         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 424 }                                                                       \
 425 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 426 {                                                                       \
 427         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 428 }                                                                       \
 429 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 430 {                                                                       \
 431         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 432 }
 433
 434 CFQ_CFQQ_FNS(on_rr);
 435 CFQ_CFQQ_FNS(wait_request);
 436 CFQ_CFQQ_FNS(must_dispatch);
 437 CFQ_CFQQ_FNS(must_alloc_slice);
 438 CFQ_CFQQ_FNS(fifo_expire);
 439 CFQ_CFQQ_FNS(idle_window);
 440 CFQ_CFQQ_FNS(prio_changed);
 441 CFQ_CFQQ_FNS(slice_new);
 442 CFQ_CFQQ_FNS(sync);
 443 CFQ_CFQQ_FNS(coop);
 444 CFQ_CFQQ_FNS(split_coop);
 445 CFQ_CFQQ_FNS(deep);
 446 CFQ_CFQQ_FNS(wait_busy);
 447 #undef CFQ_CFQQ_FNS
 448
 449 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 450 {
 451         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 452 }
 453
 454 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 455 {
 456         return pd_to_blkg(&cfqg->pd);
 457 }
 458
 459 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 460
 461 /* cfqg stats flags */
 462 enum cfqg_stats_flags {
 463         CFQG_stats_waiting = 0,
 464         CFQG_stats_idling,
 465         CFQG_stats_empty,
 466 };
 467
 468 #define CFQG_FLAG_FNS(name)                                             \
 469 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 470 {                                                                       \
 471         stats->flags |= (1 << CFQG_stats_##name);                       \
 472 }                                                                       \
 473 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 474 {                                                                       \
 475         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 476 }                                                                       \
 477 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 478 {                                                                       \
 479         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 480 }                                                                       \
 481
 482 CFQG_FLAG_FNS(waiting)
 483 CFQG_FLAG_FNS(idling)
 484 CFQG_FLAG_FNS(empty)
 485 #undef CFQG_FLAG_FNS
 486
 487 /* This should be called with the queue_lock held. */
 488 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 489 {
 490         unsigned long long now;
 491
 492         if (!cfqg_stats_waiting(stats))
 493                 return;
 494
 495         now = sched_clock();
 496         if (time_after64(now, stats->start_group_wait_time))
 497                 blkg_stat_add(&stats->group_wait_time,
 498                               now - stats->start_group_wait_time);
 499         cfqg_stats_clear_waiting(stats);
 500 }
 501
 502 /* This should be called with the queue_lock held. */
 503 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 504                                                  struct cfq_group *curr_cfqg)
 505 {
 506         struct cfqg_stats *stats = &cfqg->stats;
 507
 508         if (cfqg_stats_waiting(stats))
 509                 return;
 510         if (cfqg == curr_cfqg)
 511                 return;
 512         stats->start_group_wait_time = sched_clock();
 513         cfqg_stats_mark_waiting(stats);
 514 }
 515
 516 /* This should be called with the queue_lock held. */
 517 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 518 {
 519         unsigned long long now;
 520
 521         if (!cfqg_stats_empty(stats))
 522                 return;
 523
 524         now = sched_clock();
 525         if (time_after64(now, stats->start_empty_time))
 526                 blkg_stat_add(&stats->empty_time,
 527                               now - stats->start_empty_time);
 528         cfqg_stats_clear_empty(stats);
 529 }
 530
 531 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 532 {
 533         blkg_stat_add(&cfqg->stats.dequeue, 1);
 534 }
 535
 536 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 537 {
 538         struct cfqg_stats *stats = &cfqg->stats;
 539
 540         if (blkg_rwstat_total(&stats->queued))
 541                 return;
 542
 543         /*
 544          * group is already marked empty. This can happen if cfqq got new
 545          * request in parent group and moved to this group while being added
 546          * to service tree. Just ignore the event and move on.
 547          */
 548         if (cfqg_stats_empty(stats))
 549                 return;
 550
 551         stats->start_empty_time = sched_clock();
 552         cfqg_stats_mark_empty(stats);
 553 }
 554
 555 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 556 {
 557         struct cfqg_stats *stats = &cfqg->stats;
 558
 559         if (cfqg_stats_idling(stats)) {
 560                 unsigned long long now = sched_clock();
 561
 562                 if (time_after64(now, stats->start_idle_time))
 563                         blkg_stat_add(&stats->idle_time,
 564                                       now - stats->start_idle_time);
 565                 cfqg_stats_clear_idling(stats);
 566         }
 567 }
 568
 569 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 570 {
 571         struct cfqg_stats *stats = &cfqg->stats;
 572
 573         BUG_ON(cfqg_stats_idling(stats));
 574
 575         stats->start_idle_time = sched_clock();
 576         cfqg_stats_mark_idling(stats);
 577 }
 578
 579 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 580 {
 581         struct cfqg_stats *stats = &cfqg->stats;
 582
 583         blkg_stat_add(&stats->avg_queue_size_sum,
 584                       blkg_rwstat_total(&stats->queued));
 585         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 586         cfqg_stats_update_group_wait_time(stats);
 587 }
 588
 589 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 590
 591 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 592 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 593 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 594 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 595 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 596 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 597 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 598
 599 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 600
 601 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 602
 603 static struct blkcg_policy blkcg_policy_cfq;
 604
 605 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 606 {
 607         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 608 }
 609
 610 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 611 {
 612         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 613
 614         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 615 }
 616
 617 static inline void cfqg_get(struct cfq_group *cfqg)
 618 {
 619         return blkg_get(cfqg_to_blkg(cfqg));
 620 }
 621
 622 static inline void cfqg_put(struct cfq_group *cfqg)
 623 {
 624         return blkg_put(cfqg_to_blkg(cfqg));
 625 }
 626
 627 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 628         char __pbuf[128];                                               \
 629                                                                         \
 630         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 631         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 632                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 633                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 634                           __pbuf, ##args);                              \
 635 } while (0)
 636
 637 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 638         char __pbuf[128];                                               \
 639                                                                         \
 640         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 641         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 642 } while (0)
 643
 644 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 645                                             struct cfq_group *curr_cfqg, int rw)
 646 {
 647         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 648         cfqg_stats_end_empty_time(&cfqg->stats);
 649         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 650 }
 651
 652 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 653                         unsigned long time, unsigned long unaccounted_time)
 654 {
 655         blkg_stat_add(&cfqg->stats.time, time);
 656 #ifdef CONFIG_DEBUG_BLK_CGROUP
 657         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 658 #endif
 659 }
 660
 661 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 662 {
 663         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 664 }
 665
 666 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 667 {
 668         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 669 }
 670
 671 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 672                                               uint64_t bytes, int rw)
 673 {
 674         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
 675         blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
 676         blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
 677 }
 678
 679 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 680                         uint64_t start_time, uint64_t io_start_time, int rw)
 681 {
 682         struct cfqg_stats *stats = &cfqg->stats;
 683         unsigned long long now = sched_clock();
 684
 685         if (time_after64(now, io_start_time))
 686                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 687         if (time_after64(io_start_time, start_time))
 688                 blkg_rwstat_add(&stats->wait_time, rw,
 689                                 io_start_time - start_time);
 690 }
 691
 692 /* @stats = 0 */
 693 static void cfqg_stats_reset(struct cfqg_stats *stats)
 694 {
 695         /* queued stats shouldn't be cleared */
 696         blkg_rwstat_reset(&stats->service_bytes);
 697         blkg_rwstat_reset(&stats->serviced);
 698         blkg_rwstat_reset(&stats->merged);
 699         blkg_rwstat_reset(&stats->service_time);
 700         blkg_rwstat_reset(&stats->wait_time);
 701         blkg_stat_reset(&stats->time);
 702 #ifdef CONFIG_DEBUG_BLK_CGROUP
 703         blkg_stat_reset(&stats->unaccounted_time);
 704         blkg_stat_reset(&stats->avg_queue_size_sum);
 705         blkg_stat_reset(&stats->avg_queue_size_samples);
 706         blkg_stat_reset(&stats->dequeue);
 707         blkg_stat_reset(&stats->group_wait_time);
 708         blkg_stat_reset(&stats->idle_time);
 709         blkg_stat_reset(&stats->empty_time);
 710 #endif
 711 }
 712
 713 /* @to += @from */
 714 static void cfqg_stats_merge(struct cfqg_stats *to, struct cfqg_stats *from)
 715 {
 716         /* queued stats shouldn't be cleared */
 717         blkg_rwstat_merge(&to->service_bytes, &from->service_bytes);
 718         blkg_rwstat_merge(&to->serviced, &from->serviced);
 719         blkg_rwstat_merge(&to->merged, &from->merged);
 720         blkg_rwstat_merge(&to->service_time, &from->service_time);
 721         blkg_rwstat_merge(&to->wait_time, &from->wait_time);
 722         blkg_stat_merge(&from->time, &from->time);
 723 #ifdef CONFIG_DEBUG_BLK_CGROUP
 724         blkg_stat_merge(&to->unaccounted_time, &from->unaccounted_time);
 725         blkg_stat_merge(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 726         blkg_stat_merge(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 727         blkg_stat_merge(&to->dequeue, &from->dequeue);
 728         blkg_stat_merge(&to->group_wait_time, &from->group_wait_time);
 729         blkg_stat_merge(&to->idle_time, &from->idle_time);
 730         blkg_stat_merge(&to->empty_time, &from->empty_time);
 731 #endif
 732 }
 733
 734 /*
 735  * Transfer @cfqg's stats to its parent's dead_stats so that the ancestors'
 736  * recursive stats can still account for the amount used by this cfqg after
 737  * it's gone.
 738  */
 739 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 740 {
 741         struct cfq_group *parent = cfqg_parent(cfqg);
 742
 743         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 744
 745         if (unlikely(!parent))
 746                 return;
 747
 748         cfqg_stats_merge(&parent->dead_stats, &cfqg->stats);
 749         cfqg_stats_merge(&parent->dead_stats, &cfqg->dead_stats);
 750         cfqg_stats_reset(&cfqg->stats);
 751         cfqg_stats_reset(&cfqg->dead_stats);
 752 }
 753
 754 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 755
 756 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 757 static inline void cfqg_get(struct cfq_group *cfqg) { }
 758 static inline void cfqg_put(struct cfq_group *cfqg) { }
 759
 760 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 761         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 762                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 763                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 764                                 ##args)
 765 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 766
 767 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 768                         struct cfq_group *curr_cfqg, int rw) { }
 769 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 770                         unsigned long time, unsigned long unaccounted_time) { }
 771 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 772 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 773 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 774                                               uint64_t bytes, int rw) { }
 775 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 776                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 777
 778 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 779
 780 #define cfq_log(cfqd, fmt, args...)     \
 781         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 782
 783 /* Traverses through cfq group service trees */
 784 #define for_each_cfqg_st(cfqg, i, j, st) \
 785         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 786                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 787                         : &cfqg->service_tree_idle; \
 788                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 789                         (i == IDLE_WORKLOAD && j == 0); \
 790                         j++, st = i < IDLE_WORKLOAD ? \
 791                         &cfqg->service_trees[i][j]: NULL) \
 792
 793 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 794         struct cfq_ttime *ttime, bool group_idle)
 795 {
 796         unsigned long slice;
 797         if (!sample_valid(ttime->ttime_samples))
 798                 return false;
 799         if (group_idle)
 800                 slice = cfqd->cfq_group_idle;
 801         else
 802                 slice = cfqd->cfq_slice_idle;
 803         return ttime->ttime_mean > slice;
 804 }
 805
 806 static inline bool iops_mode(struct cfq_data *cfqd)
 807 {
 808         /*
 809          * If we are not idling on queues and it is a NCQ drive, parallel
 810          * execution of requests is on and measuring time is not possible
 811          * in most of the cases until and unless we drive shallower queue
 812          * depths and that becomes a performance bottleneck. In such cases
 813          * switch to start providing fairness in terms of number of IOs.
 814          */
 815         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 816                 return true;
 817         else
 818                 return false;
 819 }
 820
 821 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 822 {
 823         if (cfq_class_idle(cfqq))
 824                 return IDLE_WORKLOAD;
 825         if (cfq_class_rt(cfqq))
 826                 return RT_WORKLOAD;
 827         return BE_WORKLOAD;
 828 }
 829
 830
 831 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 832 {
 833         if (!cfq_cfqq_sync(cfqq))
 834                 return ASYNC_WORKLOAD;
 835         if (!cfq_cfqq_idle_window(cfqq))
 836                 return SYNC_NOIDLE_WORKLOAD;
 837         return SYNC_WORKLOAD;
 838 }
 839
 840 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 841                                         struct cfq_data *cfqd,
 842                                         struct cfq_group *cfqg)
 843 {
 844         if (wl_class == IDLE_WORKLOAD)
 845                 return cfqg->service_tree_idle.count;
 846
 847         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 848                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 849                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 850 }
 851
 852 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 853                                         struct cfq_group *cfqg)
 854 {
 855         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 856                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 857 }
 858
 859 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 860 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 861                                        struct cfq_io_cq *cic, struct bio *bio,
 862                                        gfp_t gfp_mask);
 863
 864 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 865 {
 866         /* cic->icq is the first member, %NULL will convert to %NULL */
 867         return container_of(icq, struct cfq_io_cq, icq);
 868 }
 869
 870 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 871                                                struct io_context *ioc)
 872 {
 873         if (ioc)
 874                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 875         return NULL;
 876 }
 877
 878 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 879 {
 880         return cic->cfqq[is_sync];
 881 }
 882
 883 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 884                                 bool is_sync)
 885 {
 886         cic->cfqq[is_sync] = cfqq;
 887 }
 888
 889 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 890 {
 891         return cic->icq.q->elevator->elevator_data;
 892 }
 893
 894 /*
 895  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 896  * set (in which case it could also be direct WRITE).
 897  */
 898 static inline bool cfq_bio_sync(struct bio *bio)
 899 {
 900         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 901 }
 902
 903 /*
 904  * scheduler run of queue, if there are requests pending and no one in the
 905  * driver that will restart queueing
 906  */
 907 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 908 {
 909         if (cfqd->busy_queues) {
 910                 cfq_log(cfqd, "schedule dispatch");
 911                 kblockd_schedule_work(&cfqd->unplug_work);
 912         }
 913 }
 914
 915 /*
 916  * Scale schedule slice based on io priority. Use the sync time slice only
 917  * if a queue is marked sync and has sync io queued. A sync queue with async
 918  * io only, should not get full sync slice length.
 919  */
 920 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 921                                  unsigned short prio)
 922 {
 923         const int base_slice = cfqd->cfq_slice[sync];
 924
 925         WARN_ON(prio >= IOPRIO_BE_NR);
 926
 927         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 928 }
 929
 930 static inline int
 931 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 932 {
 933         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 934 }
 935
 936 /**
 937  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 938  * @charge: disk time being charged
 939  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 940  *
 941  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 942  * scaling is inversely proportional.
 943  *
 944  * scaled = charge / vfraction
 945  *
 946  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 947  */
 948 static inline u64 cfqg_scale_charge(unsigned long charge,
 949                                     unsigned int vfraction)
 950 {
 951         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 952
 953         /* charge / vfraction */
 954         c <<= CFQ_SERVICE_SHIFT;
 955         do_div(c, vfraction);
 956         return c;
 957 }
 958
 959 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 960 {
 961         s64 delta = (s64)(vdisktime - min_vdisktime);
 962         if (delta > 0)
 963                 min_vdisktime = vdisktime;
 964
 965         return min_vdisktime;
 966 }
 967
 968 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 969 {
 970         s64 delta = (s64)(vdisktime - min_vdisktime);
 971         if (delta < 0)
 972                 min_vdisktime = vdisktime;
 973
 974         return min_vdisktime;
 975 }
 976
 977 static void update_min_vdisktime(struct cfq_rb_root *st)
 978 {
 979         struct cfq_group *cfqg;
 980
 981         if (st->left) {
 982                 cfqg = rb_entry_cfqg(st->left);
 983                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 984                                                   cfqg->vdisktime);
 985         }
 986 }
 987
 988 /*
 989  * get averaged number of queues of RT/BE priority.
 990  * average is updated, with a formula that gives more weight to higher numbers,
 991  * to quickly follows sudden increases and decrease slowly
 992  */
 993
 994 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
 995                                         struct cfq_group *cfqg, bool rt)
 996 {
 997         unsigned min_q, max_q;
 998         unsigned mult  = cfq_hist_divisor - 1;
 999         unsigned round = cfq_hist_divisor / 2;
1000         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1001
1002         min_q = min(cfqg->busy_queues_avg[rt], busy);
1003         max_q = max(cfqg->busy_queues_avg[rt], busy);
1004         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1005                 cfq_hist_divisor;
1006         return cfqg->busy_queues_avg[rt];
1007 }
1008
1009 static inline unsigned
1010 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1011 {
1012         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1013 }
1014
1015 static inline unsigned
1016 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1017 {
1018         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1019         if (cfqd->cfq_latency) {
1020                 /*
1021                  * interested queues (we consider only the ones with the same
1022                  * priority class in the cfq group)
1023                  */
1024                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1025                                                 cfq_class_rt(cfqq));
1026                 unsigned sync_slice = cfqd->cfq_slice[1];
1027                 unsigned expect_latency = sync_slice * iq;
1028                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1029
1030                 if (expect_latency > group_slice) {
1031                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1032                         /* scale low_slice according to IO priority
1033                          * and sync vs async */
1034                         unsigned low_slice =
1035                                 min(slice, base_low_slice * slice / sync_slice);
1036                         /* the adapted slice value is scaled to fit all iqs
1037                          * into the target latency */
1038                         slice = max(slice * group_slice / expect_latency,
1039                                     low_slice);
1040                 }
1041         }
1042         return slice;
1043 }
1044
1045 static inline void
1046 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1047 {
1048         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1049
1050         cfqq->slice_start = jiffies;
1051         cfqq->slice_end = jiffies + slice;
1052         cfqq->allocated_slice = slice;
1053         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1054 }
1055
1056 /*
1057  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1058  * isn't valid until the first request from the dispatch is activated
1059  * and the slice time set.
1060  */
1061 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1062 {
1063         if (cfq_cfqq_slice_new(cfqq))
1064                 return false;
1065         if (time_before(jiffies, cfqq->slice_end))
1066                 return false;
1067
1068         return true;
1069 }
1070
1071 /*
1072  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1073  * We choose the request that is closest to the head right now. Distance
1074  * behind the head is penalized and only allowed to a certain extent.
1075  */
1076 static struct request *
1077 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1078 {
1079         sector_t s1, s2, d1 = 0, d2 = 0;
1080         unsigned long back_max;
1081 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1082 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1083         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1084
1085         if (rq1 == NULL || rq1 == rq2)
1086                 return rq2;
1087         if (rq2 == NULL)
1088                 return rq1;
1089
1090         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1091                 return rq_is_sync(rq1) ? rq1 : rq2;
1092
1093         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1094                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1095
1096         s1 = blk_rq_pos(rq1);
1097         s2 = blk_rq_pos(rq2);
1098
1099         /*
1100          * by definition, 1KiB is 2 sectors
1101          */
1102         back_max = cfqd->cfq_back_max * 2;
1103
1104         /*
1105          * Strict one way elevator _except_ in the case where we allow
1106          * short backward seeks which are biased as twice the cost of a
1107          * similar forward seek.
1108          */
1109         if (s1 >= last)
1110                 d1 = s1 - last;
1111         else if (s1 + back_max >= last)
1112                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1113         else
1114                 wrap |= CFQ_RQ1_WRAP;
1115
1116         if (s2 >= last)
1117                 d2 = s2 - last;
1118         else if (s2 + back_max >= last)
1119                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1120         else
1121                 wrap |= CFQ_RQ2_WRAP;
1122
1123         /* Found required data */
1124
1125         /*
1126          * By doing switch() on the bit mask "wrap" we avoid having to
1127          * check two variables for all permutations: --> faster!
1128          */
1129         switch (wrap) {
1130         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1131                 if (d1 < d2)
1132                         return rq1;
1133                 else if (d2 < d1)
1134                         return rq2;
1135                 else {
1136                         if (s1 >= s2)
1137                                 return rq1;
1138                         else
1139                                 return rq2;
1140                 }
1141
1142         case CFQ_RQ2_WRAP:
1143                 return rq1;
1144         case CFQ_RQ1_WRAP:
1145                 return rq2;
1146         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1147         default:
1148                 /*
1149                  * Since both rqs are wrapped,
1150                  * start with the one that's further behind head
1151                  * (--> only *one* back seek required),
1152                  * since back seek takes more time than forward.
1153                  */
1154                 if (s1 <= s2)
1155                         return rq1;
1156                 else
1157                         return rq2;
1158         }
1159 }
1160
1161 /*
1162  * The below is leftmost cache rbtree addon
1163  */
1164 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1165 {
1166         /* Service tree is empty */
1167         if (!root->count)
1168                 return NULL;
1169
1170         if (!root->left)
1171                 root->left = rb_first(&root->rb);
1172
1173         if (root->left)
1174                 return rb_entry(root->left, struct cfq_queue, rb_node);
1175
1176         return NULL;
1177 }
1178
1179 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1180 {
1181         if (!root->left)
1182                 root->left = rb_first(&root->rb);
1183
1184         if (root->left)
1185                 return rb_entry_cfqg(root->left);
1186
1187         return NULL;
1188 }
1189
1190 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1191 {
1192         rb_erase(n, root);
1193         RB_CLEAR_NODE(n);
1194 }
1195
1196 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1197 {
1198         if (root->left == n)
1199                 root->left = NULL;
1200         rb_erase_init(n, &root->rb);
1201         --root->count;
1202 }
1203
1204 /*
1205  * would be nice to take fifo expire time into account as well
1206  */
1207 static struct request *
1208 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1209                   struct request *last)
1210 {
1211         struct rb_node *rbnext = rb_next(&last->rb_node);
1212         struct rb_node *rbprev = rb_prev(&last->rb_node);
1213         struct request *next = NULL, *prev = NULL;
1214
1215         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1216
1217         if (rbprev)
1218                 prev = rb_entry_rq(rbprev);
1219
1220         if (rbnext)
1221                 next = rb_entry_rq(rbnext);
1222         else {
1223                 rbnext = rb_first(&cfqq->sort_list);
1224                 if (rbnext && rbnext != &last->rb_node)
1225                         next = rb_entry_rq(rbnext);
1226         }
1227
1228         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1229 }
1230
1231 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1232                                       struct cfq_queue *cfqq)
1233 {
1234         /*
1235          * just an approximation, should be ok.
1236          */
1237         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1238                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1239 }
1240
1241 static inline s64
1242 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1243 {
1244         return cfqg->vdisktime - st->min_vdisktime;
1245 }
1246
1247 static void
1248 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1249 {
1250         struct rb_node **node = &st->rb.rb_node;
1251         struct rb_node *parent = NULL;
1252         struct cfq_group *__cfqg;
1253         s64 key = cfqg_key(st, cfqg);
1254         int left = 1;
1255
1256         while (*node != NULL) {
1257                 parent = *node;
1258                 __cfqg = rb_entry_cfqg(parent);
1259
1260                 if (key < cfqg_key(st, __cfqg))
1261                         node = &parent->rb_left;
1262                 else {
1263                         node = &parent->rb_right;
1264                         left = 0;
1265                 }
1266         }
1267
1268         if (left)
1269                 st->left = &cfqg->rb_node;
1270
1271         rb_link_node(&cfqg->rb_node, parent, node);
1272         rb_insert_color(&cfqg->rb_node, &st->rb);
1273 }
1274
1275 /*
1276  * This has to be called only on activation of cfqg
1277  */
1278 static void
1279 cfq_update_group_weight(struct cfq_group *cfqg)
1280 {
1281         if (cfqg->new_weight) {
1282                 cfqg->weight = cfqg->new_weight;
1283                 cfqg->new_weight = 0;
1284         }
1285 }
1286
1287 static void
1288 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1289 {
1290         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1291
1292         if (cfqg->new_leaf_weight) {
1293                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1294                 cfqg->new_leaf_weight = 0;
1295         }
1296 }
1297
1298 static void
1299 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1300 {
1301         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1302         struct cfq_group *pos = cfqg;
1303         struct cfq_group *parent;
1304         bool propagate;
1305
1306         /* add to the service tree */
1307         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1308
1309         /*
1310          * Update leaf_weight.  We cannot update weight at this point
1311          * because cfqg might already have been activated and is
1312          * contributing its current weight to the parent's child_weight.
1313          */
1314         cfq_update_group_leaf_weight(cfqg);
1315         __cfq_group_service_tree_add(st, cfqg);
1316
1317         /*
1318          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1319          * entitled to.  vfraction is calculated by walking the tree
1320          * towards the root calculating the fraction it has at each level.
1321          * The compounded ratio is how much vfraction @cfqg owns.
1322          *
1323          * Start with the proportion tasks in this cfqg has against active
1324          * children cfqgs - its leaf_weight against children_weight.
1325          */
1326         propagate = !pos->nr_active++;
1327         pos->children_weight += pos->leaf_weight;
1328         vfr = vfr * pos->leaf_weight / pos->children_weight;
1329
1330         /*
1331          * Compound ->weight walking up the tree.  Both activation and
1332          * vfraction calculation are done in the same loop.  Propagation
1333          * stops once an already activated node is met.  vfraction
1334          * calculation should always continue to the root.
1335          */
1336         while ((parent = cfqg_parent(pos))) {
1337                 if (propagate) {
1338                         cfq_update_group_weight(pos);
1339                         propagate = !parent->nr_active++;
1340                         parent->children_weight += pos->weight;
1341                 }
1342                 vfr = vfr * pos->weight / parent->children_weight;
1343                 pos = parent;
1344         }
1345
1346         cfqg->vfraction = max_t(unsigned, vfr, 1);
1347 }
1348
1349 static void
1350 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1351 {
1352         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1353         struct cfq_group *__cfqg;
1354         struct rb_node *n;
1355
1356         cfqg->nr_cfqq++;
1357         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1358                 return;
1359
1360         /*
1361          * Currently put the group at the end. Later implement something
1362          * so that groups get lesser vtime based on their weights, so that
1363          * if group does not loose all if it was not continuously backlogged.
1364          */
1365         n = rb_last(&st->rb);
1366         if (n) {
1367                 __cfqg = rb_entry_cfqg(n);
1368                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1369         } else
1370                 cfqg->vdisktime = st->min_vdisktime;
1371         cfq_group_service_tree_add(st, cfqg);
1372 }
1373
1374 static void
1375 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1376 {
1377         struct cfq_group *pos = cfqg;
1378         bool propagate;
1379
1380         /*
1381          * Undo activation from cfq_group_service_tree_add().  Deactivate
1382          * @cfqg and propagate deactivation upwards.
1383          */
1384         propagate = !--pos->nr_active;
1385         pos->children_weight -= pos->leaf_weight;
1386
1387         while (propagate) {
1388                 struct cfq_group *parent = cfqg_parent(pos);
1389
1390                 /* @pos has 0 nr_active at this point */
1391                 WARN_ON_ONCE(pos->children_weight);
1392                 pos->vfraction = 0;
1393
1394                 if (!parent)
1395                         break;
1396
1397                 propagate = !--parent->nr_active;
1398                 parent->children_weight -= pos->weight;
1399                 pos = parent;
1400         }
1401
1402         /* remove from the service tree */
1403         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1404                 cfq_rb_erase(&cfqg->rb_node, st);
1405 }
1406
1407 static void
1408 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1409 {
1410         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1411
1412         BUG_ON(cfqg->nr_cfqq < 1);
1413         cfqg->nr_cfqq--;
1414
1415         /* If there are other cfq queues under this group, don't delete it */
1416         if (cfqg->nr_cfqq)
1417                 return;
1418
1419         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1420         cfq_group_service_tree_del(st, cfqg);
1421         cfqg->saved_wl_slice = 0;
1422         cfqg_stats_update_dequeue(cfqg);
1423 }
1424
1425 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1426                                                 unsigned int *unaccounted_time)
1427 {
1428         unsigned int slice_used;
1429
1430         /*
1431          * Queue got expired before even a single request completed or
1432          * got expired immediately after first request completion.
1433          */
1434         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1435                 /*
1436                  * Also charge the seek time incurred to the group, otherwise
1437                  * if there are mutiple queues in the group, each can dispatch
1438                  * a single request on seeky media and cause lots of seek time
1439                  * and group will never know it.
1440                  */
1441                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1442                                         1);
1443         } else {
1444                 slice_used = jiffies - cfqq->slice_start;
1445                 if (slice_used > cfqq->allocated_slice) {
1446                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1447                         slice_used = cfqq->allocated_slice;
1448                 }
1449                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1450                         *unaccounted_time += cfqq->slice_start -
1451                                         cfqq->dispatch_start;
1452         }
1453
1454         return slice_used;
1455 }
1456
1457 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1458                                 struct cfq_queue *cfqq)
1459 {
1460         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1461         unsigned int used_sl, charge, unaccounted_sl = 0;
1462         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1463                         - cfqg->service_tree_idle.count;
1464         unsigned int vfr;
1465
1466         BUG_ON(nr_sync < 0);
1467         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1468
1469         if (iops_mode(cfqd))
1470                 charge = cfqq->slice_dispatch;
1471         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1472                 charge = cfqq->allocated_slice;
1473
1474         /*
1475          * Can't update vdisktime while on service tree and cfqg->vfraction
1476          * is valid only while on it.  Cache vfr, leave the service tree,
1477          * update vdisktime and go back on.  The re-addition to the tree
1478          * will also update the weights as necessary.
1479          */
1480         vfr = cfqg->vfraction;
1481         cfq_group_service_tree_del(st, cfqg);
1482         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1483         cfq_group_service_tree_add(st, cfqg);
1484
1485         /* This group is being expired. Save the context */
1486         if (time_after(cfqd->workload_expires, jiffies)) {
1487                 cfqg->saved_wl_slice = cfqd->workload_expires
1488                                                 - jiffies;
1489                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1490                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1491         } else
1492                 cfqg->saved_wl_slice = 0;
1493
1494         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1495                                         st->min_vdisktime);
1496         cfq_log_cfqq(cfqq->cfqd, cfqq,
1497                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1498                      used_sl, cfqq->slice_dispatch, charge,
1499                      iops_mode(cfqd), cfqq->nr_sectors);
1500         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1501         cfqg_stats_set_start_empty_time(cfqg);
1502 }
1503
1504 /**
1505  * cfq_init_cfqg_base - initialize base part of a cfq_group
1506  * @cfqg: cfq_group to initialize
1507  *
1508  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1509  * is enabled or not.
1510  */
1511 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1512 {
1513         struct cfq_rb_root *st;
1514         int i, j;
1515
1516         for_each_cfqg_st(cfqg, i, j, st)
1517                 *st = CFQ_RB_ROOT;
1518         RB_CLEAR_NODE(&cfqg->rb_node);
1519
1520         cfqg->ttime.last_end_request = jiffies;
1521 }
1522
1523 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1524 static void cfqg_stats_init(struct cfqg_stats *stats)
1525 {
1526         blkg_rwstat_init(&stats->service_bytes);
1527         blkg_rwstat_init(&stats->serviced);
1528         blkg_rwstat_init(&stats->merged);
1529         blkg_rwstat_init(&stats->service_time);
1530         blkg_rwstat_init(&stats->wait_time);
1531         blkg_rwstat_init(&stats->queued);
1532
1533         blkg_stat_init(&stats->sectors);
1534         blkg_stat_init(&stats->time);
1535
1536 #ifdef CONFIG_DEBUG_BLK_CGROUP
1537         blkg_stat_init(&stats->unaccounted_time);
1538         blkg_stat_init(&stats->avg_queue_size_sum);
1539         blkg_stat_init(&stats->avg_queue_size_samples);
1540         blkg_stat_init(&stats->dequeue);
1541         blkg_stat_init(&stats->group_wait_time);
1542         blkg_stat_init(&stats->idle_time);
1543         blkg_stat_init(&stats->empty_time);
1544 #endif
1545 }
1546
1547 static void cfq_pd_init(struct blkcg_gq *blkg)
1548 {
1549         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1550
1551         cfq_init_cfqg_base(cfqg);
1552         cfqg->weight = blkg->blkcg->cfq_weight;
1553         cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
1554         cfqg_stats_init(&cfqg->stats);
1555         cfqg_stats_init(&cfqg->dead_stats);
1556 }
1557
1558 static void cfq_pd_offline(struct blkcg_gq *blkg)
1559 {
1560         /*
1561          * @blkg is going offline and will be ignored by
1562          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1563          * that they don't get lost.  If IOs complete after this point, the
1564          * stats for them will be lost.  Oh well...
1565          */
1566         cfqg_stats_xfer_dead(blkg_to_cfqg(blkg));
1567 }
1568
1569 /* offset delta from cfqg->stats to cfqg->dead_stats */
1570 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1571                                         offsetof(struct cfq_group, stats);
1572
1573 /* to be used by recursive prfill, sums live and dead stats recursively */
1574 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1575 {
1576         u64 sum = 0;
1577
1578         sum += blkg_stat_recursive_sum(pd, off);
1579         sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1580         return sum;
1581 }
1582
1583 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1584 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1585                                                        int off)
1586 {
1587         struct blkg_rwstat a, b;
1588
1589         a = blkg_rwstat_recursive_sum(pd, off);
1590         b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1591         blkg_rwstat_merge(&a, &b);
1592         return a;
1593 }
1594
1595 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1596 {
1597         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1598
1599         cfqg_stats_reset(&cfqg->stats);
1600         cfqg_stats_reset(&cfqg->dead_stats);
1601 }
1602
1603 /*
1604  * Search for the cfq group current task belongs to. request_queue lock must
1605  * be held.
1606  */
1607 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1608                                                 struct blkcg *blkcg)
1609 {
1610         struct request_queue *q = cfqd->queue;
1611         struct cfq_group *cfqg = NULL;
1612
1613         /* avoid lookup for the common case where there's no blkcg */
1614         if (blkcg == &blkcg_root) {
1615                 cfqg = cfqd->root_group;
1616         } else {
1617                 struct blkcg_gq *blkg;
1618
1619                 blkg = blkg_lookup_create(blkcg, q);
1620                 if (!IS_ERR(blkg))
1621                         cfqg = blkg_to_cfqg(blkg);
1622         }
1623
1624         return cfqg;
1625 }
1626
1627 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1628 {
1629         /* Currently, all async queues are mapped to root group */
1630         if (!cfq_cfqq_sync(cfqq))
1631                 cfqg = cfqq->cfqd->root_group;
1632
1633         cfqq->cfqg = cfqg;
1634         /* cfqq reference on cfqg */
1635         cfqg_get(cfqg);
1636 }
1637
1638 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1639                                      struct blkg_policy_data *pd, int off)
1640 {
1641         struct cfq_group *cfqg = pd_to_cfqg(pd);
1642
1643         if (!cfqg->dev_weight)
1644                 return 0;
1645         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1646 }
1647
1648 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1649 {
1650         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1651                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1652                           0, false);
1653         return 0;
1654 }
1655
1656 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1657                                           struct blkg_policy_data *pd, int off)
1658 {
1659         struct cfq_group *cfqg = pd_to_cfqg(pd);
1660
1661         if (!cfqg->dev_leaf_weight)
1662                 return 0;
1663         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1664 }
1665
1666 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1667 {
1668         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1669                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1670                           0, false);
1671         return 0;
1672 }
1673
1674 static int cfq_print_weight(struct seq_file *sf, void *v)
1675 {
1676         seq_printf(sf, "%u\n", css_to_blkcg(seq_css(sf))->cfq_weight);
1677         return 0;
1678 }
1679
1680 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1681 {
1682         seq_printf(sf, "%u\n", css_to_blkcg(seq_css(sf))->cfq_leaf_weight);
1683         return 0;
1684 }
1685
1686 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1687                                         char *buf, size_t nbytes, loff_t off,
1688                                         bool is_leaf_weight)
1689 {
1690         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1691         struct blkg_conf_ctx ctx;
1692         struct cfq_group *cfqg;
1693         int ret;
1694
1695         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1696         if (ret)
1697                 return ret;
1698
1699         ret = -EINVAL;
1700         cfqg = blkg_to_cfqg(ctx.blkg);
1701         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1702                 if (!is_leaf_weight) {
1703                         cfqg->dev_weight = ctx.v;
1704                         cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1705                 } else {
1706                         cfqg->dev_leaf_weight = ctx.v;
1707                         cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1708                 }
1709                 ret = 0;
1710         }
1711
1712         blkg_conf_finish(&ctx);
1713         return ret ?: nbytes;
1714 }
1715
1716 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1717                                       char *buf, size_t nbytes, loff_t off)
1718 {
1719         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1720 }
1721
1722 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1723                                            char *buf, size_t nbytes, loff_t off)
1724 {
1725         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1726 }
1727
1728 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1729                             u64 val, bool is_leaf_weight)
1730 {
1731         struct blkcg *blkcg = css_to_blkcg(css);
1732         struct blkcg_gq *blkg;
1733
1734         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1735                 return -EINVAL;
1736
1737         spin_lock_irq(&blkcg->lock);
1738
1739         if (!is_leaf_weight)
1740                 blkcg->cfq_weight = val;
1741         else
1742                 blkcg->cfq_leaf_weight = val;
1743
1744         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1745                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1746
1747                 if (!cfqg)
1748                         continue;
1749
1750                 if (!is_leaf_weight) {
1751                         if (!cfqg->dev_weight)
1752                                 cfqg->new_weight = blkcg->cfq_weight;
1753                 } else {
1754                         if (!cfqg->dev_leaf_weight)
1755                                 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1756                 }
1757         }
1758
1759         spin_unlock_irq(&blkcg->lock);
1760         return 0;
1761 }
1762
1763 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1764                           u64 val)
1765 {
1766         return __cfq_set_weight(css, cft, val, false);
1767 }
1768
1769 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1770                                struct cftype *cft, u64 val)
1771 {
1772         return __cfq_set_weight(css, cft, val, true);
1773 }
1774
1775 static int cfqg_print_stat(struct seq_file *sf, void *v)
1776 {
1777         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1778                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1779         return 0;
1780 }
1781
1782 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1783 {
1784         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1785                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1786         return 0;
1787 }
1788
1789 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1790                                       struct blkg_policy_data *pd, int off)
1791 {
1792         u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1793
1794         return __blkg_prfill_u64(sf, pd, sum);
1795 }
1796
1797 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1798                                         struct blkg_policy_data *pd, int off)
1799 {
1800         struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1801
1802         return __blkg_prfill_rwstat(sf, pd, &sum);
1803 }
1804
1805 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1806 {
1807         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1808                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1809                           seq_cft(sf)->private, false);
1810         return 0;
1811 }
1812
1813 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1814 {
1815         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1816                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1817                           seq_cft(sf)->private, true);
1818         return 0;
1819 }
1820
1821 #ifdef CONFIG_DEBUG_BLK_CGROUP
1822 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1823                                       struct blkg_policy_data *pd, int off)
1824 {
1825         struct cfq_group *cfqg = pd_to_cfqg(pd);
1826         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1827         u64 v = 0;
1828
1829         if (samples) {
1830                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1831                 v = div64_u64(v, samples);
1832         }
1833         __blkg_prfill_u64(sf, pd, v);
1834         return 0;
1835 }
1836
1837 /* print avg_queue_size */
1838 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1839 {
1840         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1841                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1842                           0, false);
1843         return 0;
1844 }
1845 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1846
1847 static struct cftype cfq_blkcg_files[] = {
1848         /* on root, weight is mapped to leaf_weight */
1849         {
1850                 .name = "weight_device",
1851                 .flags = CFTYPE_ONLY_ON_ROOT,
1852                 .seq_show = cfqg_print_leaf_weight_device,
1853                 .write = cfqg_set_leaf_weight_device,
1854         },
1855         {
1856                 .name = "weight",
1857                 .flags = CFTYPE_ONLY_ON_ROOT,
1858                 .seq_show = cfq_print_leaf_weight,
1859                 .write_u64 = cfq_set_leaf_weight,
1860         },
1861
1862         /* no such mapping necessary for !roots */
1863         {
1864                 .name = "weight_device",
1865                 .flags = CFTYPE_NOT_ON_ROOT,
1866                 .seq_show = cfqg_print_weight_device,
1867                 .write = cfqg_set_weight_device,
1868         },
1869         {
1870                 .name = "weight",
1871                 .flags = CFTYPE_NOT_ON_ROOT,
1872                 .seq_show = cfq_print_weight,
1873                 .write_u64 = cfq_set_weight,
1874         },
1875
1876         {
1877                 .name = "leaf_weight_device",
1878                 .seq_show = cfqg_print_leaf_weight_device,
1879                 .write = cfqg_set_leaf_weight_device,
1880         },
1881         {
1882                 .name = "leaf_weight",
1883                 .seq_show = cfq_print_leaf_weight,
1884                 .write_u64 = cfq_set_leaf_weight,
1885         },
1886
1887         /* statistics, covers only the tasks in the cfqg */
1888         {
1889                 .name = "time",
1890                 .private = offsetof(struct cfq_group, stats.time),
1891                 .seq_show = cfqg_print_stat,
1892         },
1893         {
1894                 .name = "sectors",
1895                 .private = offsetof(struct cfq_group, stats.sectors),
1896                 .seq_show = cfqg_print_stat,
1897         },
1898         {
1899                 .name = "io_service_bytes",
1900                 .private = offsetof(struct cfq_group, stats.service_bytes),
1901                 .seq_show = cfqg_print_rwstat,
1902         },
1903         {
1904                 .name = "io_serviced",
1905                 .private = offsetof(struct cfq_group, stats.serviced),
1906                 .seq_show = cfqg_print_rwstat,
1907         },
1908         {
1909                 .name = "io_service_time",
1910                 .private = offsetof(struct cfq_group, stats.service_time),
1911                 .seq_show = cfqg_print_rwstat,
1912         },
1913         {
1914                 .name = "io_wait_time",
1915                 .private = offsetof(struct cfq_group, stats.wait_time),
1916                 .seq_show = cfqg_print_rwstat,
1917         },
1918         {
1919                 .name = "io_merged",
1920                 .private = offsetof(struct cfq_group, stats.merged),
1921                 .seq_show = cfqg_print_rwstat,
1922         },
1923         {
1924                 .name = "io_queued",
1925                 .private = offsetof(struct cfq_group, stats.queued),
1926                 .seq_show = cfqg_print_rwstat,
1927         },
1928
1929         /* the same statictics which cover the cfqg and its descendants */
1930         {
1931                 .name = "time_recursive",
1932                 .private = offsetof(struct cfq_group, stats.time),
1933                 .seq_show = cfqg_print_stat_recursive,
1934         },
1935         {
1936                 .name = "sectors_recursive",
1937                 .private = offsetof(struct cfq_group, stats.sectors),
1938                 .seq_show = cfqg_print_stat_recursive,
1939         },
1940         {
1941                 .name = "io_service_bytes_recursive",
1942                 .private = offsetof(struct cfq_group, stats.service_bytes),
1943                 .seq_show = cfqg_print_rwstat_recursive,
1944         },
1945         {
1946                 .name = "io_serviced_recursive",
1947                 .private = offsetof(struct cfq_group, stats.serviced),
1948                 .seq_show = cfqg_print_rwstat_recursive,
1949         },
1950         {
1951                 .name = "io_service_time_recursive",
1952                 .private = offsetof(struct cfq_group, stats.service_time),
1953                 .seq_show = cfqg_print_rwstat_recursive,
1954         },
1955         {
1956                 .name = "io_wait_time_recursive",
1957                 .private = offsetof(struct cfq_group, stats.wait_time),
1958                 .seq_show = cfqg_print_rwstat_recursive,
1959         },
1960         {
1961                 .name = "io_merged_recursive",
1962                 .private = offsetof(struct cfq_group, stats.merged),
1963                 .seq_show = cfqg_print_rwstat_recursive,
1964         },
1965         {
1966                 .name = "io_queued_recursive",
1967                 .private = offsetof(struct cfq_group, stats.queued),
1968                 .seq_show = cfqg_print_rwstat_recursive,
1969         },
1970 #ifdef CONFIG_DEBUG_BLK_CGROUP
1971         {
1972                 .name = "avg_queue_size",
1973                 .seq_show = cfqg_print_avg_queue_size,
1974         },
1975         {
1976                 .name = "group_wait_time",
1977                 .private = offsetof(struct cfq_group, stats.group_wait_time),
1978                 .seq_show = cfqg_print_stat,
1979         },
1980         {
1981                 .name = "idle_time",
1982                 .private = offsetof(struct cfq_group, stats.idle_time),
1983                 .seq_show = cfqg_print_stat,
1984         },
1985         {
1986                 .name = "empty_time",
1987                 .private = offsetof(struct cfq_group, stats.empty_time),
1988                 .seq_show = cfqg_print_stat,
1989         },
1990         {
1991                 .name = "dequeue",
1992                 .private = offsetof(struct cfq_group, stats.dequeue),
1993                 .seq_show = cfqg_print_stat,
1994         },
1995         {
1996                 .name = "unaccounted_time",
1997                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1998                 .seq_show = cfqg_print_stat,
1999         },
2000 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2001         { }     /* terminate */
2002 };
2003 #else /* GROUP_IOSCHED */
2004 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
2005                                                 struct blkcg *blkcg)
2006 {
2007         return cfqd->root_group;
2008 }
2009
2010 static inline void
2011 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2012         cfqq->cfqg = cfqg;
2013 }
2014
2015 #endif /* GROUP_IOSCHED */
2016
2017 /*
2018  * The cfqd->service_trees holds all pending cfq_queue's that have
2019  * requests waiting to be processed. It is sorted in the order that
2020  * we will service the queues.
2021  */
2022 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2023                                  bool add_front)
2024 {
2025         struct rb_node **p, *parent;
2026         struct cfq_queue *__cfqq;
2027         unsigned long rb_key;
2028         struct cfq_rb_root *st;
2029         int left;
2030         int new_cfqq = 1;
2031
2032         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2033         if (cfq_class_idle(cfqq)) {
2034                 rb_key = CFQ_IDLE_DELAY;
2035                 parent = rb_last(&st->rb);
2036                 if (parent && parent != &cfqq->rb_node) {
2037                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2038                         rb_key += __cfqq->rb_key;
2039                 } else
2040                         rb_key += jiffies;
2041         } else if (!add_front) {
2042                 /*
2043                  * Get our rb key offset. Subtract any residual slice
2044                  * value carried from last service. A negative resid
2045                  * count indicates slice overrun, and this should position
2046                  * the next service time further away in the tree.
2047                  */
2048                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2049                 rb_key -= cfqq->slice_resid;
2050                 cfqq->slice_resid = 0;
2051         } else {
2052                 rb_key = -HZ;
2053                 __cfqq = cfq_rb_first(st);
2054                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2055         }
2056
2057         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2058                 new_cfqq = 0;
2059                 /*
2060                  * same position, nothing more to do
2061                  */
2062                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2063                         return;
2064
2065                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2066                 cfqq->service_tree = NULL;
2067         }
2068
2069         left = 1;
2070         parent = NULL;
2071         cfqq->service_tree = st;
2072         p = &st->rb.rb_node;
2073         while (*p) {
2074                 parent = *p;
2075                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2076
2077                 /*
2078                  * sort by key, that represents service time.
2079                  */
2080                 if (time_before(rb_key, __cfqq->rb_key))
2081                         p = &parent->rb_left;
2082                 else {
2083                         p = &parent->rb_right;
2084                         left = 0;
2085                 }
2086         }
2087
2088         if (left)
2089                 st->left = &cfqq->rb_node;
2090
2091         cfqq->rb_key = rb_key;
2092         rb_link_node(&cfqq->rb_node, parent, p);
2093         rb_insert_color(&cfqq->rb_node, &st->rb);
2094         st->count++;
2095         if (add_front || !new_cfqq)
2096                 return;
2097         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2098 }
2099
2100 static struct cfq_queue *
2101 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2102                      sector_t sector, struct rb_node **ret_parent,
2103                      struct rb_node ***rb_link)
2104 {
2105         struct rb_node **p, *parent;
2106         struct cfq_queue *cfqq = NULL;
2107
2108         parent = NULL;
2109         p = &root->rb_node;
2110         while (*p) {
2111                 struct rb_node **n;
2112
2113                 parent = *p;
2114                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2115
2116                 /*
2117                  * Sort strictly based on sector.  Smallest to the left,
2118                  * largest to the right.
2119                  */
2120                 if (sector > blk_rq_pos(cfqq->next_rq))
2121                         n = &(*p)->rb_right;
2122                 else if (sector < blk_rq_pos(cfqq->next_rq))
2123                         n = &(*p)->rb_left;
2124                 else
2125                         break;
2126                 p = n;
2127                 cfqq = NULL;
2128         }
2129
2130         *ret_parent = parent;
2131         if (rb_link)
2132                 *rb_link = p;
2133         return cfqq;
2134 }
2135
2136 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2137 {
2138         struct rb_node **p, *parent;
2139         struct cfq_queue *__cfqq;
2140
2141         if (cfqq->p_root) {
2142                 rb_erase(&cfqq->p_node, cfqq->p_root);
2143                 cfqq->p_root = NULL;
2144         }
2145
2146         if (cfq_class_idle(cfqq))
2147                 return;
2148         if (!cfqq->next_rq)
2149                 return;
2150
2151         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2152         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2153                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2154         if (!__cfqq) {
2155                 rb_link_node(&cfqq->p_node, parent, p);
2156                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2157         } else
2158                 cfqq->p_root = NULL;
2159 }
2160
2161 /*
2162  * Update cfqq's position in the service tree.
2163  */
2164 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2165 {
2166         /*
2167          * Resorting requires the cfqq to be on the RR list already.
2168          */
2169         if (cfq_cfqq_on_rr(cfqq)) {
2170                 cfq_service_tree_add(cfqd, cfqq, 0);
2171                 cfq_prio_tree_add(cfqd, cfqq);
2172         }
2173 }
2174
2175 /*
2176  * add to busy list of queues for service, trying to be fair in ordering
2177  * the pending list according to last request service
2178  */
2179 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2180 {
2181         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2182         BUG_ON(cfq_cfqq_on_rr(cfqq));
2183         cfq_mark_cfqq_on_rr(cfqq);
2184         cfqd->busy_queues++;
2185         if (cfq_cfqq_sync(cfqq))
2186                 cfqd->busy_sync_queues++;
2187
2188         cfq_resort_rr_list(cfqd, cfqq);
2189 }
2190
2191 /*
2192  * Called when the cfqq no longer has requests pending, remove it from
2193  * the service tree.
2194  */
2195 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2196 {
2197         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2198         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2199         cfq_clear_cfqq_on_rr(cfqq);
2200
2201         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2202                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2203                 cfqq->service_tree = NULL;
2204         }
2205         if (cfqq->p_root) {
2206                 rb_erase(&cfqq->p_node, cfqq->p_root);
2207                 cfqq->p_root = NULL;
2208         }
2209
2210         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2211         BUG_ON(!cfqd->busy_queues);
2212         cfqd->busy_queues--;
2213         if (cfq_cfqq_sync(cfqq))
2214                 cfqd->busy_sync_queues--;
2215 }
2216
2217 /*
2218  * rb tree support functions
2219  */
2220 static void cfq_del_rq_rb(struct request *rq)
2221 {
2222         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2223         const int sync = rq_is_sync(rq);
2224
2225         BUG_ON(!cfqq->queued[sync]);
2226         cfqq->queued[sync]--;
2227
2228         elv_rb_del(&cfqq->sort_list, rq);
2229
2230         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2231                 /*
2232                  * Queue will be deleted from service tree when we actually
2233                  * expire it later. Right now just remove it from prio tree
2234                  * as it is empty.
2235                  */
2236                 if (cfqq->p_root) {
2237                         rb_erase(&cfqq->p_node, cfqq->p_root);
2238                         cfqq->p_root = NULL;
2239                 }
2240         }
2241 }
2242
2243 static void cfq_add_rq_rb(struct request *rq)
2244 {
2245         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2246         struct cfq_data *cfqd = cfqq->cfqd;
2247         struct request *prev;
2248
2249         cfqq->queued[rq_is_sync(rq)]++;
2250
2251         elv_rb_add(&cfqq->sort_list, rq);
2252
2253         if (!cfq_cfqq_on_rr(cfqq))
2254                 cfq_add_cfqq_rr(cfqd, cfqq);
2255
2256         /*
2257          * check if this request is a better next-serve candidate
2258          */
2259         prev = cfqq->next_rq;
2260         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2261
2262         /*
2263          * adjust priority tree position, if ->next_rq changes
2264          */
2265         if (prev != cfqq->next_rq)
2266                 cfq_prio_tree_add(cfqd, cfqq);
2267
2268         BUG_ON(!cfqq->next_rq);
2269 }
2270
2271 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2272 {
2273         elv_rb_del(&cfqq->sort_list, rq);
2274         cfqq->queued[rq_is_sync(rq)]--;
2275         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2276         cfq_add_rq_rb(rq);
2277         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2278                                  rq->cmd_flags);
2279 }
2280
2281 static struct request *
2282 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2283 {
2284         struct task_struct *tsk = current;
2285         struct cfq_io_cq *cic;
2286         struct cfq_queue *cfqq;
2287
2288         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2289         if (!cic)
2290                 return NULL;
2291
2292         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2293         if (cfqq)
2294                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2295
2296         return NULL;
2297 }
2298
2299 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2300 {
2301         struct cfq_data *cfqd = q->elevator->elevator_data;
2302
2303         cfqd->rq_in_driver++;
2304         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2305                                                 cfqd->rq_in_driver);
2306
2307         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2308 }
2309
2310 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2311 {
2312         struct cfq_data *cfqd = q->elevator->elevator_data;
2313
2314         WARN_ON(!cfqd->rq_in_driver);
2315         cfqd->rq_in_driver--;
2316         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2317                                                 cfqd->rq_in_driver);
2318 }
2319
2320 static void cfq_remove_request(struct request *rq)
2321 {
2322         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2323
2324         if (cfqq->next_rq == rq)
2325                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2326
2327         list_del_init(&rq->queuelist);
2328         cfq_del_rq_rb(rq);
2329
2330         cfqq->cfqd->rq_queued--;
2331         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2332         if (rq->cmd_flags & REQ_PRIO) {
2333                 WARN_ON(!cfqq->prio_pending);
2334                 cfqq->prio_pending--;
2335         }
2336 }
2337
2338 static int cfq_merge(struct request_queue *q, struct request **req,
2339                      struct bio *bio)
2340 {
2341         struct cfq_data *cfqd = q->elevator->elevator_data;
2342         struct request *__rq;
2343
2344         __rq = cfq_find_rq_fmerge(cfqd, bio);
2345         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2346                 *req = __rq;
2347                 return ELEVATOR_FRONT_MERGE;
2348         }
2349
2350         return ELEVATOR_NO_MERGE;
2351 }
2352
2353 static void cfq_merged_request(struct request_queue *q, struct request *req,
2354                                int type)
2355 {
2356         if (type == ELEVATOR_FRONT_MERGE) {
2357                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2358
2359                 cfq_reposition_rq_rb(cfqq, req);
2360         }
2361 }
2362
2363 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2364                                 struct bio *bio)
2365 {
2366         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2367 }
2368
2369 static void
2370 cfq_merged_requests(struct request_queue *q, struct request *rq,
2371                     struct request *next)
2372 {
2373         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2374         struct cfq_data *cfqd = q->elevator->elevator_data;
2375
2376         /*
2377          * reposition in fifo if next is older than rq
2378          */
2379         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2380             time_before(next->fifo_time, rq->fifo_time) &&
2381             cfqq == RQ_CFQQ(next)) {
2382                 list_move(&rq->queuelist, &next->queuelist);
2383                 rq->fifo_time = next->fifo_time;
2384         }
2385
2386         if (cfqq->next_rq == next)
2387                 cfqq->next_rq = rq;
2388         cfq_remove_request(next);
2389         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2390
2391         cfqq = RQ_CFQQ(next);
2392         /*
2393          * all requests of this queue are merged to other queues, delete it
2394          * from the service tree. If it's the active_queue,
2395          * cfq_dispatch_requests() will choose to expire it or do idle
2396          */
2397         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2398             cfqq != cfqd->active_queue)
2399                 cfq_del_cfqq_rr(cfqd, cfqq);
2400 }
2401
2402 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2403                            struct bio *bio)
2404 {
2405         struct cfq_data *cfqd = q->elevator->elevator_data;
2406         struct cfq_io_cq *cic;
2407         struct cfq_queue *cfqq;
2408
2409         /*
2410          * Disallow merge of a sync bio into an async request.
2411          */
2412         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2413                 return false;
2414
2415         /*
2416          * Lookup the cfqq that this bio will be queued with and allow
2417          * merge only if rq is queued there.
2418          */
2419         cic = cfq_cic_lookup(cfqd, current->io_context);
2420         if (!cic)
2421                 return false;
2422
2423         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2424         return cfqq == RQ_CFQQ(rq);
2425 }
2426
2427 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2428 {
2429         del_timer(&cfqd->idle_slice_timer);
2430         cfqg_stats_update_idle_time(cfqq->cfqg);
2431 }
2432
2433 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2434                                    struct cfq_queue *cfqq)
2435 {
2436         if (cfqq) {
2437                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2438                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2439                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2440                 cfqq->slice_start = 0;
2441                 cfqq->dispatch_start = jiffies;
2442                 cfqq->allocated_slice = 0;
2443                 cfqq->slice_end = 0;
2444                 cfqq->slice_dispatch = 0;
2445                 cfqq->nr_sectors = 0;
2446
2447                 cfq_clear_cfqq_wait_request(cfqq);
2448                 cfq_clear_cfqq_must_dispatch(cfqq);
2449                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2450                 cfq_clear_cfqq_fifo_expire(cfqq);
2451                 cfq_mark_cfqq_slice_new(cfqq);
2452
2453                 cfq_del_timer(cfqd, cfqq);
2454         }
2455
2456         cfqd->active_queue = cfqq;
2457 }
2458
2459 /*
2460  * current cfqq expired its slice (or was too idle), select new one
2461  */
2462 static void
2463 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2464                     bool timed_out)
2465 {
2466         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2467
2468         if (cfq_cfqq_wait_request(cfqq))
2469                 cfq_del_timer(cfqd, cfqq);
2470
2471         cfq_clear_cfqq_wait_request(cfqq);
2472         cfq_clear_cfqq_wait_busy(cfqq);
2473
2474         /*
2475          * If this cfqq is shared between multiple processes, check to
2476          * make sure that those processes are still issuing I/Os within
2477          * the mean seek distance.  If not, it may be time to break the
2478          * queues apart again.
2479          */
2480         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2481                 cfq_mark_cfqq_split_coop(cfqq);
2482
2483         /*
2484          * store what was left of this slice, if the queue idled/timed out
2485          */
2486         if (timed_out) {
2487                 if (cfq_cfqq_slice_new(cfqq))
2488                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2489                 else
2490                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2491                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2492         }
2493
2494         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2495
2496         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2497                 cfq_del_cfqq_rr(cfqd, cfqq);
2498
2499         cfq_resort_rr_list(cfqd, cfqq);
2500
2501         if (cfqq == cfqd->active_queue)
2502                 cfqd->active_queue = NULL;
2503
2504         if (cfqd->active_cic) {
2505                 put_io_context(cfqd->active_cic->icq.ioc);
2506                 cfqd->active_cic = NULL;
2507         }
2508 }
2509
2510 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2511 {
2512         struct cfq_queue *cfqq = cfqd->active_queue;
2513
2514         if (cfqq)
2515                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2516 }
2517
2518 /*
2519  * Get next queue for service. Unless we have a queue preemption,
2520  * we'll simply select the first cfqq in the service tree.
2521  */
2522 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2523 {
2524         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2525                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2526
2527         if (!cfqd->rq_queued)
2528                 return NULL;
2529
2530         /* There is nothing to dispatch */
2531         if (!st)
2532                 return NULL;
2533         if (RB_EMPTY_ROOT(&st->rb))
2534                 return NULL;
2535         return cfq_rb_first(st);
2536 }
2537
2538 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2539 {
2540         struct cfq_group *cfqg;
2541         struct cfq_queue *cfqq;
2542         int i, j;
2543         struct cfq_rb_root *st;
2544
2545         if (!cfqd->rq_queued)
2546                 return NULL;
2547
2548         cfqg = cfq_get_next_cfqg(cfqd);
2549         if (!cfqg)
2550                 return NULL;
2551
2552         for_each_cfqg_st(cfqg, i, j, st)
2553                 if ((cfqq = cfq_rb_first(st)) != NULL)
2554                         return cfqq;
2555         return NULL;
2556 }
2557
2558 /*
2559  * Get and set a new active queue for service.
2560  */
2561 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2562                                               struct cfq_queue *cfqq)
2563 {
2564         if (!cfqq)
2565                 cfqq = cfq_get_next_queue(cfqd);
2566
2567         __cfq_set_active_queue(cfqd, cfqq);
2568         return cfqq;
2569 }
2570
2571 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2572                                           struct request *rq)
2573 {
2574         if (blk_rq_pos(rq) >= cfqd->last_position)
2575                 return blk_rq_pos(rq) - cfqd->last_position;
2576         else
2577                 return cfqd->last_position - blk_rq_pos(rq);
2578 }
2579
2580 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2581                                struct request *rq)
2582 {
2583         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2584 }
2585
2586 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2587                                     struct cfq_queue *cur_cfqq)
2588 {
2589         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2590         struct rb_node *parent, *node;
2591         struct cfq_queue *__cfqq;
2592         sector_t sector = cfqd->last_position;
2593
2594         if (RB_EMPTY_ROOT(root))
2595                 return NULL;
2596
2597         /*
2598          * First, if we find a request starting at the end of the last
2599          * request, choose it.
2600          */
2601         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2602         if (__cfqq)
2603                 return __cfqq;
2604
2605         /*
2606          * If the exact sector wasn't found, the parent of the NULL leaf
2607          * will contain the closest sector.
2608          */
2609         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2610         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2611                 return __cfqq;
2612
2613         if (blk_rq_pos(__cfqq->next_rq) < sector)
2614                 node = rb_next(&__cfqq->p_node);
2615         else
2616                 node = rb_prev(&__cfqq->p_node);
2617         if (!node)
2618                 return NULL;
2619
2620         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2621         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2622                 return __cfqq;
2623
2624         return NULL;
2625 }
2626
2627 /*
2628  * cfqd - obvious
2629  * cur_cfqq - passed in so that we don't decide that the current queue is
2630  *            closely cooperating with itself.
2631  *
2632  * So, basically we're assuming that that cur_cfqq has dispatched at least
2633  * one request, and that cfqd->last_position reflects a position on the disk
2634  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2635  * assumption.
2636  */
2637 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2638                                               struct cfq_queue *cur_cfqq)
2639 {
2640         struct cfq_queue *cfqq;
2641
2642         if (cfq_class_idle(cur_cfqq))
2643                 return NULL;
2644         if (!cfq_cfqq_sync(cur_cfqq))
2645                 return NULL;
2646         if (CFQQ_SEEKY(cur_cfqq))
2647                 return NULL;
2648
2649         /*
2650          * Don't search priority tree if it's the only queue in the group.
2651          */
2652         if (cur_cfqq->cfqg->nr_cfqq == 1)
2653                 return NULL;
2654
2655         /*
2656          * We should notice if some of the queues are cooperating, eg
2657          * working closely on the same area of the disk. In that case,
2658          * we can group them together and don't waste time idling.
2659          */
2660         cfqq = cfqq_close(cfqd, cur_cfqq);
2661         if (!cfqq)
2662                 return NULL;
2663
2664         /* If new queue belongs to different cfq_group, don't choose it */
2665         if (cur_cfqq->cfqg != cfqq->cfqg)
2666                 return NULL;
2667
2668         /*
2669          * It only makes sense to merge sync queues.
2670          */
2671         if (!cfq_cfqq_sync(cfqq))
2672                 return NULL;
2673         if (CFQQ_SEEKY(cfqq))
2674                 return NULL;
2675
2676         /*
2677          * Do not merge queues of different priority classes
2678          */
2679         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2680                 return NULL;
2681
2682         return cfqq;
2683 }
2684
2685 /*
2686  * Determine whether we should enforce idle window for this queue.
2687  */
2688
2689 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2690 {
2691         enum wl_class_t wl_class = cfqq_class(cfqq);
2692         struct cfq_rb_root *st = cfqq->service_tree;
2693
2694         BUG_ON(!st);
2695         BUG_ON(!st->count);
2696
2697         if (!cfqd->cfq_slice_idle)
2698                 return false;
2699
2700         /* We never do for idle class queues. */
2701         if (wl_class == IDLE_WORKLOAD)
2702                 return false;
2703
2704         /* We do for queues that were marked with idle window flag. */
2705         if (cfq_cfqq_idle_window(cfqq) &&
2706            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2707                 return true;
2708
2709         /*
2710          * Otherwise, we do only if they are the last ones
2711          * in their service tree.
2712          */
2713         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2714            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2715                 return true;
2716         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2717         return false;
2718 }
2719
2720 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2721 {
2722         struct cfq_queue *cfqq = cfqd->active_queue;
2723         struct cfq_io_cq *cic;
2724         unsigned long sl, group_idle = 0;
2725
2726         /*
2727          * SSD device without seek penalty, disable idling. But only do so
2728          * for devices that support queuing, otherwise we still have a problem
2729          * with sync vs async workloads.
2730          */
2731         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2732                 return;
2733
2734         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2735         WARN_ON(cfq_cfqq_slice_new(cfqq));
2736
2737         /*
2738          * idle is disabled, either manually or by past process history
2739          */
2740         if (!cfq_should_idle(cfqd, cfqq)) {
2741                 /* no queue idling. Check for group idling */
2742                 if (cfqd->cfq_group_idle)
2743                         group_idle = cfqd->cfq_group_idle;
2744                 else
2745                         return;
2746         }
2747
2748         /*
2749          * still active requests from this queue, don't idle
2750          */
2751         if (cfqq->dispatched)
2752                 return;
2753
2754         /*
2755          * task has exited, don't wait
2756          */
2757         cic = cfqd->active_cic;
2758         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2759                 return;
2760
2761         /*
2762          * If our average think time is larger than the remaining time
2763          * slice, then don't idle. This avoids overrunning the allotted
2764          * time slice.
2765          */
2766         if (sample_valid(cic->ttime.ttime_samples) &&
2767             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2768                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2769                              cic->ttime.ttime_mean);
2770                 return;
2771         }
2772
2773         /* There are other queues in the group, don't do group idle */
2774         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2775                 return;
2776
2777         cfq_mark_cfqq_wait_request(cfqq);
2778
2779         if (group_idle)
2780                 sl = cfqd->cfq_group_idle;
2781         else
2782                 sl = cfqd->cfq_slice_idle;
2783
2784         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2785         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2786         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2787                         group_idle ? 1 : 0);
2788 }
2789
2790 /*
2791  * Move request from internal lists to the request queue dispatch list.
2792  */
2793 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2794 {
2795         struct cfq_data *cfqd = q->elevator->elevator_data;
2796         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2797
2798         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2799
2800         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2801         cfq_remove_request(rq);
2802         cfqq->dispatched++;
2803         (RQ_CFQG(rq))->dispatched++;
2804         elv_dispatch_sort(q, rq);
2805
2806         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2807         cfqq->nr_sectors += blk_rq_sectors(rq);
2808         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2809 }
2810
2811 /*
2812  * return expired entry, or NULL to just start from scratch in rbtree
2813  */
2814 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2815 {
2816         struct request *rq = NULL;
2817
2818         if (cfq_cfqq_fifo_expire(cfqq))
2819                 return NULL;
2820
2821         cfq_mark_cfqq_fifo_expire(cfqq);
2822
2823         if (list_empty(&cfqq->fifo))
2824                 return NULL;
2825
2826         rq = rq_entry_fifo(cfqq->fifo.next);
2827         if (time_before(jiffies, rq->fifo_time))
2828                 rq = NULL;
2829
2830         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2831         return rq;
2832 }
2833
2834 static inline int
2835 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2836 {
2837         const int base_rq = cfqd->cfq_slice_async_rq;
2838
2839         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2840
2841         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2842 }
2843
2844 /*
2845  * Must be called with the queue_lock held.
2846  */
2847 static int cfqq_process_refs(struct cfq_queue *cfqq)
2848 {
2849         int process_refs, io_refs;
2850
2851         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2852         process_refs = cfqq->ref - io_refs;
2853         BUG_ON(process_refs < 0);
2854         return process_refs;
2855 }
2856
2857 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2858 {
2859         int process_refs, new_process_refs;
2860         struct cfq_queue *__cfqq;
2861
2862         /*
2863          * If there are no process references on the new_cfqq, then it is
2864          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2865          * chain may have dropped their last reference (not just their
2866          * last process reference).
2867          */
2868         if (!cfqq_process_refs(new_cfqq))
2869                 return;
2870
2871         /* Avoid a circular list and skip interim queue merges */
2872         while ((__cfqq = new_cfqq->new_cfqq)) {
2873                 if (__cfqq == cfqq)
2874                         return;
2875                 new_cfqq = __cfqq;
2876         }
2877
2878         process_refs = cfqq_process_refs(cfqq);
2879         new_process_refs = cfqq_process_refs(new_cfqq);
2880         /*
2881          * If the process for the cfqq has gone away, there is no
2882          * sense in merging the queues.
2883          */
2884         if (process_refs == 0 || new_process_refs == 0)
2885                 return;
2886
2887         /*
2888          * Merge in the direction of the lesser amount of work.
2889          */
2890         if (new_process_refs >= process_refs) {
2891                 cfqq->new_cfqq = new_cfqq;
2892                 new_cfqq->ref += process_refs;
2893         } else {
2894                 new_cfqq->new_cfqq = cfqq;
2895                 cfqq->ref += new_process_refs;
2896         }
2897 }
2898
2899 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2900                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2901 {
2902         struct cfq_queue *queue;
2903         int i;
2904         bool key_valid = false;
2905         unsigned long lowest_key = 0;
2906         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2907
2908         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2909                 /* select the one with lowest rb_key */
2910                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2911                 if (queue &&
2912                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2913                         lowest_key = queue->rb_key;
2914                         cur_best = i;
2915                         key_valid = true;
2916                 }
2917         }
2918
2919         return cur_best;
2920 }
2921
2922 static void
2923 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2924 {
2925         unsigned slice;
2926         unsigned count;
2927         struct cfq_rb_root *st;
2928         unsigned group_slice;
2929         enum wl_class_t original_class = cfqd->serving_wl_class;
2930
2931         /* Choose next priority. RT > BE > IDLE */
2932         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2933                 cfqd->serving_wl_class = RT_WORKLOAD;
2934         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2935                 cfqd->serving_wl_class = BE_WORKLOAD;
2936         else {
2937                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2938                 cfqd->workload_expires = jiffies + 1;
2939                 return;
2940         }
2941
2942         if (original_class != cfqd->serving_wl_class)
2943                 goto new_workload;
2944
2945         /*
2946          * For RT and BE, we have to choose also the type
2947          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2948          * expiration time
2949          */
2950         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2951         count = st->count;
2952
2953         /*
2954          * check workload expiration, and that we still have other queues ready
2955          */
2956         if (count && !time_after(jiffies, cfqd->workload_expires))
2957                 return;
2958
2959 new_workload:
2960         /* otherwise select new workload type */
2961         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2962                                         cfqd->serving_wl_class);
2963         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2964         count = st->count;
2965
2966         /*
2967          * the workload slice is computed as a fraction of target latency
2968          * proportional to the number of queues in that workload, over
2969          * all the queues in the same priority class
2970          */
2971         group_slice = cfq_group_slice(cfqd, cfqg);
2972
2973         slice = group_slice * count /
2974                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2975                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2976                                         cfqg));
2977
2978         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2979                 unsigned int tmp;
2980
2981                 /*
2982                  * Async queues are currently system wide. Just taking
2983                  * proportion of queues with-in same group will lead to higher
2984                  * async ratio system wide as generally root group is going
2985                  * to have higher weight. A more accurate thing would be to
2986                  * calculate system wide asnc/sync ratio.
2987                  */
2988                 tmp = cfqd->cfq_target_latency *
2989                         cfqg_busy_async_queues(cfqd, cfqg);
2990                 tmp = tmp/cfqd->busy_queues;
2991                 slice = min_t(unsigned, slice, tmp);
2992
2993                 /* async workload slice is scaled down according to
2994                  * the sync/async slice ratio. */
2995                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2996         } else
2997                 /* sync workload slice is at least 2 * cfq_slice_idle */
2998                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2999
3000         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3001         cfq_log(cfqd, "workload slice:%d", slice);
3002         cfqd->workload_expires = jiffies + slice;
3003 }
3004
3005 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3006 {
3007         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3008         struct cfq_group *cfqg;
3009
3010         if (RB_EMPTY_ROOT(&st->rb))
3011                 return NULL;
3012         cfqg = cfq_rb_first_group(st);
3013         update_min_vdisktime(st);
3014         return cfqg;
3015 }
3016
3017 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3018 {
3019         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3020
3021         cfqd->serving_group = cfqg;
3022
3023         /* Restore the workload type data */
3024         if (cfqg->saved_wl_slice) {
3025                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3026                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3027                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3028         } else
3029                 cfqd->workload_expires = jiffies - 1;
3030
3031         choose_wl_class_and_type(cfqd, cfqg);
3032 }
3033
3034 /*
3035  * Select a queue for service. If we have a current active queue,
3036  * check whether to continue servicing it, or retrieve and set a new one.
3037  */
3038 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3039 {
3040         struct cfq_queue *cfqq, *new_cfqq = NULL;
3041
3042         cfqq = cfqd->active_queue;
3043         if (!cfqq)
3044                 goto new_queue;
3045
3046         if (!cfqd->rq_queued)
3047                 return NULL;
3048
3049         /*
3050          * We were waiting for group to get backlogged. Expire the queue
3051          */
3052         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3053                 goto expire;
3054
3055         /*
3056          * The active queue has run out of time, expire it and select new.
3057          */
3058         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3059                 /*
3060                  * If slice had not expired at the completion of last request
3061                  * we might not have turned on wait_busy flag. Don't expire
3062                  * the queue yet. Allow the group to get backlogged.
3063                  *
3064                  * The very fact that we have used the slice, that means we
3065                  * have been idling all along on this queue and it should be
3066                  * ok to wait for this request to complete.
3067                  */
3068                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3069                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3070                         cfqq = NULL;
3071                         goto keep_queue;
3072                 } else
3073                         goto check_group_idle;
3074         }
3075
3076         /*
3077          * The active queue has requests and isn't expired, allow it to
3078          * dispatch.
3079          */
3080         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3081                 goto keep_queue;
3082
3083         /*
3084          * If another queue has a request waiting within our mean seek
3085          * distance, let it run.  The expire code will check for close
3086          * cooperators and put the close queue at the front of the service
3087          * tree.  If possible, merge the expiring queue with the new cfqq.
3088          */
3089         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3090         if (new_cfqq) {
3091                 if (!cfqq->new_cfqq)
3092                         cfq_setup_merge(cfqq, new_cfqq);
3093                 goto expire;
3094         }
3095
3096         /*
3097          * No requests pending. If the active queue still has requests in
3098          * flight or is idling for a new request, allow either of these
3099          * conditions to happen (or time out) before selecting a new queue.
3100          */
3101         if (timer_pending(&cfqd->idle_slice_timer)) {
3102                 cfqq = NULL;
3103                 goto keep_queue;
3104         }
3105
3106         /*
3107          * This is a deep seek queue, but the device is much faster than
3108          * the queue can deliver, don't idle
3109          **/
3110         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3111             (cfq_cfqq_slice_new(cfqq) ||
3112             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3113                 cfq_clear_cfqq_deep(cfqq);
3114                 cfq_clear_cfqq_idle_window(cfqq);
3115         }
3116
3117         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3118                 cfqq = NULL;
3119                 goto keep_queue;
3120         }
3121
3122         /*
3123          * If group idle is enabled and there are requests dispatched from
3124          * this group, wait for requests to complete.
3125          */
3126 check_group_idle:
3127         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3128             cfqq->cfqg->dispatched &&
3129             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3130                 cfqq = NULL;
3131                 goto keep_queue;
3132         }
3133
3134 expire:
3135         cfq_slice_expired(cfqd, 0);
3136 new_queue:
3137         /*
3138          * Current queue expired. Check if we have to switch to a new
3139          * service tree
3140          */
3141         if (!new_cfqq)
3142                 cfq_choose_cfqg(cfqd);
3143
3144         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3145 keep_queue:
3146         return cfqq;
3147 }
3148
3149 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3150 {
3151         int dispatched = 0;
3152
3153         while (cfqq->next_rq) {
3154                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3155                 dispatched++;
3156         }
3157
3158         BUG_ON(!list_empty(&cfqq->fifo));
3159
3160         /* By default cfqq is not expired if it is empty. Do it explicitly */
3161         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3162         return dispatched;
3163 }
3164
3165 /*
3166  * Drain our current requests. Used for barriers and when switching
3167  * io schedulers on-the-fly.
3168  */
3169 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3170 {
3171         struct cfq_queue *cfqq;
3172         int dispatched = 0;
3173
3174         /* Expire the timeslice of the current active queue first */
3175         cfq_slice_expired(cfqd, 0);
3176         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3177                 __cfq_set_active_queue(cfqd, cfqq);
3178                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3179         }
3180
3181         BUG_ON(cfqd->busy_queues);
3182
3183         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3184         return dispatched;
3185 }
3186
3187 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3188         struct cfq_queue *cfqq)
3189 {
3190         /* the queue hasn't finished any request, can't estimate */
3191         if (cfq_cfqq_slice_new(cfqq))
3192                 return true;
3193         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3194                 cfqq->slice_end))
3195                 return true;
3196
3197         return false;
3198 }
3199
3200 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3201 {
3202         unsigned int max_dispatch;
3203
3204         /*
3205          * Drain async requests before we start sync IO
3206          */
3207         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3208                 return false;
3209
3210         /*
3211          * If this is an async queue and we have sync IO in flight, let it wait
3212          */
3213         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3214                 return false;
3215
3216         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3217         if (cfq_class_idle(cfqq))
3218                 max_dispatch = 1;
3219
3220         /*
3221          * Does this cfqq already have too much IO in flight?
3222          */
3223         if (cfqq->dispatched >= max_dispatch) {
3224                 bool promote_sync = false;
3225                 /*
3226                  * idle queue must always only have a single IO in flight
3227                  */
3228                 if (cfq_class_idle(cfqq))
3229                         return false;
3230
3231                 /*
3232                  * If there is only one sync queue
3233                  * we can ignore async queue here and give the sync
3234                  * queue no dispatch limit. The reason is a sync queue can
3235                  * preempt async queue, limiting the sync queue doesn't make
3236                  * sense. This is useful for aiostress test.
3237                  */
3238                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3239                         promote_sync = true;
3240
3241                 /*
3242                  * We have other queues, don't allow more IO from this one
3243                  */
3244                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3245                                 !promote_sync)
3246                         return false;
3247
3248                 /*
3249                  * Sole queue user, no limit
3250                  */
3251                 if (cfqd->busy_queues == 1 || promote_sync)
3252                         max_dispatch = -1;
3253                 else
3254                         /*
3255                          * Normally we start throttling cfqq when cfq_quantum/2
3256                          * requests have been dispatched. But we can drive
3257                          * deeper queue depths at the beginning of slice
3258                          * subjected to upper limit of cfq_quantum.
3259                          * */
3260                         max_dispatch = cfqd->cfq_quantum;
3261         }
3262
3263         /*
3264          * Async queues must wait a bit before being allowed dispatch.
3265          * We also ramp up the dispatch depth gradually for async IO,
3266          * based on the last sync IO we serviced
3267          */
3268         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3269                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3270                 unsigned int depth;
3271
3272                 depth = last_sync / cfqd->cfq_slice[1];
3273                 if (!depth && !cfqq->dispatched)
3274                         depth = 1;
3275                 if (depth < max_dispatch)
3276                         max_dispatch = depth;
3277         }
3278
3279         /*
3280          * If we're below the current max, allow a dispatch
3281          */
3282         return cfqq->dispatched < max_dispatch;
3283 }
3284
3285 /*
3286  * Dispatch a request from cfqq, moving them to the request queue
3287  * dispatch list.
3288  */
3289 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3290 {
3291         struct request *rq;
3292
3293         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3294
3295         if (!cfq_may_dispatch(cfqd, cfqq))
3296                 return false;
3297
3298         /*
3299          * follow expired path, else get first next available
3300          */
3301         rq = cfq_check_fifo(cfqq);
3302         if (!rq)
3303                 rq = cfqq->next_rq;
3304
3305         /*
3306          * insert request into driver dispatch list
3307          */
3308         cfq_dispatch_insert(cfqd->queue, rq);
3309
3310         if (!cfqd->active_cic) {
3311                 struct cfq_io_cq *cic = RQ_CIC(rq);
3312
3313                 atomic_long_inc(&cic->icq.ioc->refcount);
3314                 cfqd->active_cic = cic;
3315         }
3316
3317         return true;
3318 }
3319
3320 /*
3321  * Find the cfqq that we need to service and move a request from that to the
3322  * dispatch list
3323  */
3324 static int cfq_dispatch_requests(struct request_queue *q, int force)
3325 {
3326         struct cfq_data *cfqd = q->elevator->elevator_data;
3327         struct cfq_queue *cfqq;
3328
3329         if (!cfqd->busy_queues)
3330                 return 0;
3331
3332         if (unlikely(force))
3333                 return cfq_forced_dispatch(cfqd);
3334
3335         cfqq = cfq_select_queue(cfqd);
3336         if (!cfqq)
3337                 return 0;
3338
3339         /*
3340          * Dispatch a request from this cfqq, if it is allowed
3341          */
3342         if (!cfq_dispatch_request(cfqd, cfqq))
3343                 return 0;
3344
3345         cfqq->slice_dispatch++;
3346         cfq_clear_cfqq_must_dispatch(cfqq);
3347
3348         /*
3349          * expire an async queue immediately if it has used up its slice. idle
3350          * queue always expire after 1 dispatch round.
3351          */
3352         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3353             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3354             cfq_class_idle(cfqq))) {
3355                 cfqq->slice_end = jiffies + 1;
3356                 cfq_slice_expired(cfqd, 0);
3357         }
3358
3359         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3360         return 1;
3361 }
3362
3363 /*
3364  * task holds one reference to the queue, dropped when task exits. each rq
3365  * in-flight on this queue also holds a reference, dropped when rq is freed.
3366  *
3367  * Each cfq queue took a reference on the parent group. Drop it now.
3368  * queue lock must be held here.
3369  */
3370 static void cfq_put_queue(struct cfq_queue *cfqq)
3371 {
3372         struct cfq_data *cfqd = cfqq->cfqd;
3373         struct cfq_group *cfqg;
3374
3375         BUG_ON(cfqq->ref <= 0);
3376
3377         cfqq->ref--;
3378         if (cfqq->ref)
3379                 return;
3380
3381         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3382         BUG_ON(rb_first(&cfqq->sort_list));
3383         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3384         cfqg = cfqq->cfqg;
3385
3386         if (unlikely(cfqd->active_queue == cfqq)) {
3387                 __cfq_slice_expired(cfqd, cfqq, 0);
3388                 cfq_schedule_dispatch(cfqd);
3389         }
3390
3391         BUG_ON(cfq_cfqq_on_rr(cfqq));
3392         kmem_cache_free(cfq_pool, cfqq);
3393         cfqg_put(cfqg);
3394 }
3395
3396 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3397 {
3398         struct cfq_queue *__cfqq, *next;
3399
3400         /*
3401          * If this queue was scheduled to merge with another queue, be
3402          * sure to drop the reference taken on that queue (and others in
3403          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3404          */
3405         __cfqq = cfqq->new_cfqq;
3406         while (__cfqq) {
3407                 if (__cfqq == cfqq) {
3408                         WARN(1, "cfqq->new_cfqq loop detected\n");
3409                         break;
3410                 }
3411                 next = __cfqq->new_cfqq;
3412                 cfq_put_queue(__cfqq);
3413                 __cfqq = next;
3414         }
3415 }
3416
3417 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3418 {
3419         if (unlikely(cfqq == cfqd->active_queue)) {
3420                 __cfq_slice_expired(cfqd, cfqq, 0);
3421                 cfq_schedule_dispatch(cfqd);
3422         }
3423
3424         cfq_put_cooperator(cfqq);
3425
3426         cfq_put_queue(cfqq);
3427 }
3428
3429 static void cfq_init_icq(struct io_cq *icq)
3430 {
3431         struct cfq_io_cq *cic = icq_to_cic(icq);
3432
3433         cic->ttime.last_end_request = jiffies;
3434 }
3435
3436 static void cfq_exit_icq(struct io_cq *icq)
3437 {
3438         struct cfq_io_cq *cic = icq_to_cic(icq);
3439         struct cfq_data *cfqd = cic_to_cfqd(cic);
3440
3441         if (cic->cfqq[BLK_RW_ASYNC]) {
3442                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3443                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3444         }
3445
3446         if (cic->cfqq[BLK_RW_SYNC]) {
3447                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3448                 cic->cfqq[BLK_RW_SYNC] = NULL;
3449         }
3450 }
3451
3452 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3453 {
3454         struct task_struct *tsk = current;
3455         int ioprio_class;
3456
3457         if (!cfq_cfqq_prio_changed(cfqq))
3458                 return;
3459
3460         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3461         switch (ioprio_class) {
3462         default:
3463                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3464         case IOPRIO_CLASS_NONE:
3465                 /*
3466                  * no prio set, inherit CPU scheduling settings
3467                  */
3468                 cfqq->ioprio = task_nice_ioprio(tsk);
3469                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3470                 break;
3471         case IOPRIO_CLASS_RT:
3472                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3473                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3474                 break;
3475         case IOPRIO_CLASS_BE:
3476                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3477                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3478                 break;
3479         case IOPRIO_CLASS_IDLE:
3480                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3481                 cfqq->ioprio = 7;
3482                 cfq_clear_cfqq_idle_window(cfqq);
3483                 break;
3484         }
3485
3486         /*
3487          * keep track of original prio settings in case we have to temporarily
3488          * elevate the priority of this queue
3489          */
3490         cfqq->org_ioprio = cfqq->ioprio;
3491         cfq_clear_cfqq_prio_changed(cfqq);
3492 }
3493
3494 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3495 {
3496         int ioprio = cic->icq.ioc->ioprio;
3497         struct cfq_data *cfqd = cic_to_cfqd(cic);
3498         struct cfq_queue *cfqq;
3499
3500         /*
3501          * Check whether ioprio has changed.  The condition may trigger
3502          * spuriously on a newly created cic but there's no harm.
3503          */
3504         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3505                 return;
3506
3507         cfqq = cic->cfqq[BLK_RW_ASYNC];
3508         if (cfqq) {
3509                 struct cfq_queue *new_cfqq;
3510                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3511                                          GFP_ATOMIC);
3512                 if (new_cfqq) {
3513                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3514                         cfq_put_queue(cfqq);
3515                 }
3516         }
3517
3518         cfqq = cic->cfqq[BLK_RW_SYNC];
3519         if (cfqq)
3520                 cfq_mark_cfqq_prio_changed(cfqq);
3521
3522         cic->ioprio = ioprio;
3523 }
3524
3525 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3526                           pid_t pid, bool is_sync)
3527 {
3528         RB_CLEAR_NODE(&cfqq->rb_node);
3529         RB_CLEAR_NODE(&cfqq->p_node);
3530         INIT_LIST_HEAD(&cfqq->fifo);
3531
3532         cfqq->ref = 0;
3533         cfqq->cfqd = cfqd;
3534
3535         cfq_mark_cfqq_prio_changed(cfqq);
3536
3537         if (is_sync) {
3538                 if (!cfq_class_idle(cfqq))
3539                         cfq_mark_cfqq_idle_window(cfqq);
3540                 cfq_mark_cfqq_sync(cfqq);
3541         }
3542         cfqq->pid = pid;
3543 }
3544
3545 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3546 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3547 {
3548         struct cfq_data *cfqd = cic_to_cfqd(cic);
3549         struct cfq_queue *sync_cfqq;
3550         uint64_t serial_nr;
3551
3552         rcu_read_lock();
3553         serial_nr = bio_blkcg(bio)->css.serial_nr;
3554         rcu_read_unlock();
3555
3556         /*
3557          * Check whether blkcg has changed.  The condition may trigger
3558          * spuriously on a newly created cic but there's no harm.
3559          */
3560         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3561                 return;
3562
3563         sync_cfqq = cic_to_cfqq(cic, 1);
3564         if (sync_cfqq) {
3565                 /*
3566                  * Drop reference to sync queue. A new sync queue will be
3567                  * assigned in new group upon arrival of a fresh request.
3568                  */
3569                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3570                 cic_set_cfqq(cic, NULL, 1);
3571                 cfq_put_queue(sync_cfqq);
3572         }
3573
3574         cic->blkcg_serial_nr = serial_nr;
3575 }
3576 #else
3577 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3578 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3579
3580 static struct cfq_queue *
3581 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3582                      struct bio *bio, gfp_t gfp_mask)
3583 {
3584         struct blkcg *blkcg;
3585         struct cfq_queue *cfqq, *new_cfqq = NULL;
3586         struct cfq_group *cfqg;
3587
3588 retry:
3589         rcu_read_lock();
3590
3591         blkcg = bio_blkcg(bio);
3592         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3593         cfqq = cic_to_cfqq(cic, is_sync);
3594
3595         /*
3596          * Always try a new alloc if we fell back to the OOM cfqq
3597          * originally, since it should just be a temporary situation.
3598          */
3599         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3600                 cfqq = NULL;
3601                 if (new_cfqq) {
3602                         cfqq = new_cfqq;
3603                         new_cfqq = NULL;
3604                 } else if (gfp_mask & __GFP_WAIT) {
3605                         rcu_read_unlock();
3606                         spin_unlock_irq(cfqd->queue->queue_lock);
3607                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3608                                         gfp_mask | __GFP_ZERO,
3609                                         cfqd->queue->node);
3610                         spin_lock_irq(cfqd->queue->queue_lock);
3611                         if (new_cfqq)
3612                                 goto retry;
3613                         else
3614                                 return &cfqd->oom_cfqq;
3615                 } else {
3616                         cfqq = kmem_cache_alloc_node(cfq_pool,
3617                                         gfp_mask | __GFP_ZERO,
3618                                         cfqd->queue->node);
3619                 }
3620
3621                 if (cfqq) {
3622                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3623                         cfq_init_prio_data(cfqq, cic);
3624                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3625                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3626                 } else
3627                         cfqq = &cfqd->oom_cfqq;
3628         }
3629
3630         if (new_cfqq)
3631                 kmem_cache_free(cfq_pool, new_cfqq);
3632
3633         rcu_read_unlock();
3634         return cfqq;
3635 }
3636
3637 static struct cfq_queue **
3638 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3639 {
3640         switch (ioprio_class) {
3641         case IOPRIO_CLASS_RT:
3642                 return &cfqd->async_cfqq[0][ioprio];
3643         case IOPRIO_CLASS_NONE:
3644                 ioprio = IOPRIO_NORM;
3645                 /* fall through */
3646         case IOPRIO_CLASS_BE:
3647                 return &cfqd->async_cfqq[1][ioprio];
3648         case IOPRIO_CLASS_IDLE:
3649                 return &cfqd->async_idle_cfqq;
3650         default:
3651                 BUG();
3652         }
3653 }
3654
3655 static struct cfq_queue *
3656 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3657               struct bio *bio, gfp_t gfp_mask)
3658 {
3659         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3660         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3661         struct cfq_queue **async_cfqq = NULL;
3662         struct cfq_queue *cfqq = NULL;
3663
3664         if (!is_sync) {
3665                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3666                 cfqq = *async_cfqq;
3667         }
3668
3669         if (!cfqq)
3670                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3671
3672         /*
3673          * pin the queue now that it's allocated, scheduler exit will prune it
3674          */
3675         if (!is_sync && !(*async_cfqq)) {
3676                 cfqq->ref++;
3677                 *async_cfqq = cfqq;
3678         }
3679
3680         cfqq->ref++;
3681         return cfqq;
3682 }
3683
3684 static void
3685 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3686 {
3687         unsigned long elapsed = jiffies - ttime->last_end_request;
3688         elapsed = min(elapsed, 2UL * slice_idle);
3689
3690         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3691         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3692         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3693 }
3694
3695 static void
3696 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3697                         struct cfq_io_cq *cic)
3698 {
3699         if (cfq_cfqq_sync(cfqq)) {
3700                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3701                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3702                         cfqd->cfq_slice_idle);
3703         }
3704 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3705         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3706 #endif
3707 }
3708
3709 static void
3710 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3711                        struct request *rq)
3712 {
3713         sector_t sdist = 0;
3714         sector_t n_sec = blk_rq_sectors(rq);
3715         if (cfqq->last_request_pos) {
3716                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3717                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3718                 else
3719                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3720         }
3721
3722         cfqq->seek_history <<= 1;
3723         if (blk_queue_nonrot(cfqd->queue))
3724                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3725         else
3726                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3727 }
3728
3729 /*
3730  * Disable idle window if the process thinks too long or seeks so much that
3731  * it doesn't matter
3732  */
3733 static void
3734 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3735                        struct cfq_io_cq *cic)
3736 {
3737         int old_idle, enable_idle;
3738
3739         /*
3740          * Don't idle for async or idle io prio class
3741          */
3742         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3743                 return;
3744
3745         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3746
3747         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3748                 cfq_mark_cfqq_deep(cfqq);
3749
3750         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3751                 enable_idle = 0;
3752         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3753                  !cfqd->cfq_slice_idle ||
3754                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3755                 enable_idle = 0;
3756         else if (sample_valid(cic->ttime.ttime_samples)) {
3757                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3758                         enable_idle = 0;
3759                 else
3760                         enable_idle = 1;
3761         }
3762
3763         if (old_idle != enable_idle) {
3764                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3765                 if (enable_idle)
3766                         cfq_mark_cfqq_idle_window(cfqq);
3767                 else
3768                         cfq_clear_cfqq_idle_window(cfqq);
3769         }
3770 }
3771
3772 /*
3773  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3774  * no or if we aren't sure, a 1 will cause a preempt.
3775  */
3776 static bool
3777 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3778                    struct request *rq)
3779 {
3780         struct cfq_queue *cfqq;
3781
3782         cfqq = cfqd->active_queue;
3783         if (!cfqq)
3784                 return false;
3785
3786         if (cfq_class_idle(new_cfqq))
3787                 return false;
3788
3789         if (cfq_class_idle(cfqq))
3790                 return true;
3791
3792         /*
3793          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3794          */
3795         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3796                 return false;
3797
3798         /*
3799          * if the new request is sync, but the currently running queue is
3800          * not, let the sync request have priority.
3801          */
3802         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3803                 return true;
3804
3805         if (new_cfqq->cfqg != cfqq->cfqg)
3806                 return false;
3807
3808         if (cfq_slice_used(cfqq))
3809                 return true;
3810
3811         /* Allow preemption only if we are idling on sync-noidle tree */
3812         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3813             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3814             new_cfqq->service_tree->count == 2 &&
3815             RB_EMPTY_ROOT(&cfqq->sort_list))
3816                 return true;
3817
3818         /*
3819          * So both queues are sync. Let the new request get disk time if
3820          * it's a metadata request and the current queue is doing regular IO.
3821          */
3822         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3823                 return true;
3824
3825         /*
3826          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3827          */
3828         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3829                 return true;
3830
3831         /* An idle queue should not be idle now for some reason */
3832         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3833                 return true;
3834
3835         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3836                 return false;
3837
3838         /*
3839          * if this request is as-good as one we would expect from the
3840          * current cfqq, let it preempt
3841          */
3842         if (cfq_rq_close(cfqd, cfqq, rq))
3843                 return true;
3844
3845         return false;
3846 }
3847
3848 /*
3849  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3850  * let it have half of its nominal slice.
3851  */
3852 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3853 {
3854         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3855
3856         cfq_log_cfqq(cfqd, cfqq, "preempt");
3857         cfq_slice_expired(cfqd, 1);
3858
3859         /*
3860          * workload type is changed, don't save slice, otherwise preempt
3861          * doesn't happen
3862          */
3863         if (old_type != cfqq_type(cfqq))
3864                 cfqq->cfqg->saved_wl_slice = 0;
3865
3866         /*
3867          * Put the new queue at the front of the of the current list,
3868          * so we know that it will be selected next.
3869          */
3870         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3871
3872         cfq_service_tree_add(cfqd, cfqq, 1);
3873
3874         cfqq->slice_end = 0;
3875         cfq_mark_cfqq_slice_new(cfqq);
3876 }
3877
3878 /*
3879  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3880  * something we should do about it
3881  */
3882 static void
3883 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3884                 struct request *rq)
3885 {
3886         struct cfq_io_cq *cic = RQ_CIC(rq);
3887
3888         cfqd->rq_queued++;
3889         if (rq->cmd_flags & REQ_PRIO)
3890                 cfqq->prio_pending++;
3891
3892         cfq_update_io_thinktime(cfqd, cfqq, cic);
3893         cfq_update_io_seektime(cfqd, cfqq, rq);
3894         cfq_update_idle_window(cfqd, cfqq, cic);
3895
3896         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3897
3898         if (cfqq == cfqd->active_queue) {
3899                 /*
3900                  * Remember that we saw a request from this process, but
3901                  * don't start queuing just yet. Otherwise we risk seeing lots
3902                  * of tiny requests, because we disrupt the normal plugging
3903                  * and merging. If the request is already larger than a single
3904                  * page, let it rip immediately. For that case we assume that
3905                  * merging is already done. Ditto for a busy system that
3906                  * has other work pending, don't risk delaying until the
3907                  * idle timer unplug to continue working.
3908                  */
3909                 if (cfq_cfqq_wait_request(cfqq)) {
3910                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3911                             cfqd->busy_queues > 1) {
3912                                 cfq_del_timer(cfqd, cfqq);
3913                                 cfq_clear_cfqq_wait_request(cfqq);
3914                                 __blk_run_queue(cfqd->queue);
3915                         } else {
3916                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3917                                 cfq_mark_cfqq_must_dispatch(cfqq);
3918                         }
3919                 }
3920         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3921                 /*
3922                  * not the active queue - expire current slice if it is
3923                  * idle and has expired it's mean thinktime or this new queue
3924                  * has some old slice time left and is of higher priority or
3925                  * this new queue is RT and the current one is BE
3926                  */
3927                 cfq_preempt_queue(cfqd, cfqq);
3928                 __blk_run_queue(cfqd->queue);
3929         }
3930 }
3931
3932 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3933 {
3934         struct cfq_data *cfqd = q->elevator->elevator_data;
3935         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3936
3937         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3938         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3939
3940         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
3941         list_add_tail(&rq->queuelist, &cfqq->fifo);
3942         cfq_add_rq_rb(rq);
3943         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3944                                  rq->cmd_flags);
3945         cfq_rq_enqueued(cfqd, cfqq, rq);
3946 }
3947
3948 /*
3949  * Update hw_tag based on peak queue depth over 50 samples under
3950  * sufficient load.
3951  */
3952 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3953 {
3954         struct cfq_queue *cfqq = cfqd->active_queue;
3955
3956         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3957                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3958
3959         if (cfqd->hw_tag == 1)
3960                 return;
3961
3962         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3963             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3964                 return;
3965
3966         /*
3967          * If active queue hasn't enough requests and can idle, cfq might not
3968          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3969          * case
3970          */
3971         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3972             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3973             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3974                 return;
3975
3976         if (cfqd->hw_tag_samples++ < 50)
3977                 return;
3978
3979         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3980                 cfqd->hw_tag = 1;
3981         else
3982                 cfqd->hw_tag = 0;
3983 }
3984
3985 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3986 {
3987         struct cfq_io_cq *cic = cfqd->active_cic;
3988
3989         /* If the queue already has requests, don't wait */
3990         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3991                 return false;
3992
3993         /* If there are other queues in the group, don't wait */
3994         if (cfqq->cfqg->nr_cfqq > 1)
3995                 return false;
3996
3997         /* the only queue in the group, but think time is big */
3998         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3999                 return false;
4000
4001         if (cfq_slice_used(cfqq))
4002                 return true;
4003
4004         /* if slice left is less than think time, wait busy */
4005         if (cic && sample_valid(cic->ttime.ttime_samples)
4006             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4007                 return true;
4008
4009         /*
4010          * If think times is less than a jiffy than ttime_mean=0 and above
4011          * will not be true. It might happen that slice has not expired yet
4012          * but will expire soon (4-5 ns) during select_queue(). To cover the
4013          * case where think time is less than a jiffy, mark the queue wait
4014          * busy if only 1 jiffy is left in the slice.
4015          */
4016         if (cfqq->slice_end - jiffies == 1)
4017                 return true;
4018
4019         return false;
4020 }
4021
4022 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4023 {
4024         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4025         struct cfq_data *cfqd = cfqq->cfqd;
4026         const int sync = rq_is_sync(rq);
4027         unsigned long now;
4028
4029         now = jiffies;
4030         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4031                      !!(rq->cmd_flags & REQ_NOIDLE));
4032
4033         cfq_update_hw_tag(cfqd);
4034
4035         WARN_ON(!cfqd->rq_in_driver);
4036         WARN_ON(!cfqq->dispatched);
4037         cfqd->rq_in_driver--;
4038         cfqq->dispatched--;
4039         (RQ_CFQG(rq))->dispatched--;
4040         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4041                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4042
4043         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4044
4045         if (sync) {
4046                 struct cfq_rb_root *st;
4047
4048                 RQ_CIC(rq)->ttime.last_end_request = now;
4049
4050                 if (cfq_cfqq_on_rr(cfqq))
4051                         st = cfqq->service_tree;
4052                 else
4053                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4054                                         cfqq_type(cfqq));
4055
4056                 st->ttime.last_end_request = now;
4057                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4058                         cfqd->last_delayed_sync = now;
4059         }
4060
4061 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4062         cfqq->cfqg->ttime.last_end_request = now;
4063 #endif
4064
4065         /*
4066          * If this is the active queue, check if it needs to be expired,
4067          * or if we want to idle in case it has no pending requests.
4068          */
4069         if (cfqd->active_queue == cfqq) {
4070                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4071
4072                 if (cfq_cfqq_slice_new(cfqq)) {
4073                         cfq_set_prio_slice(cfqd, cfqq);
4074                         cfq_clear_cfqq_slice_new(cfqq);
4075                 }
4076
4077                 /*
4078                  * Should we wait for next request to come in before we expire
4079                  * the queue.
4080                  */
4081                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4082                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4083                         if (!cfqd->cfq_slice_idle)
4084                                 extend_sl = cfqd->cfq_group_idle;
4085                         cfqq->slice_end = jiffies + extend_sl;
4086                         cfq_mark_cfqq_wait_busy(cfqq);
4087                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4088                 }
4089
4090                 /*
4091                  * Idling is not enabled on:
4092                  * - expired queues
4093                  * - idle-priority queues
4094                  * - async queues
4095                  * - queues with still some requests queued
4096                  * - when there is a close cooperator
4097                  */
4098                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4099                         cfq_slice_expired(cfqd, 1);
4100                 else if (sync && cfqq_empty &&
4101                          !cfq_close_cooperator(cfqd, cfqq)) {
4102                         cfq_arm_slice_timer(cfqd);
4103                 }
4104         }
4105
4106         if (!cfqd->rq_in_driver)
4107                 cfq_schedule_dispatch(cfqd);
4108 }
4109
4110 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4111 {
4112         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4113                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4114                 return ELV_MQUEUE_MUST;
4115         }
4116
4117         return ELV_MQUEUE_MAY;
4118 }
4119
4120 static int cfq_may_queue(struct request_queue *q, int rw)
4121 {
4122         struct cfq_data *cfqd = q->elevator->elevator_data;
4123         struct task_struct *tsk = current;
4124         struct cfq_io_cq *cic;
4125         struct cfq_queue *cfqq;
4126
4127         /*
4128          * don't force setup of a queue from here, as a call to may_queue
4129          * does not necessarily imply that a request actually will be queued.
4130          * so just lookup a possibly existing queue, or return 'may queue'
4131          * if that fails
4132          */
4133         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4134         if (!cic)
4135                 return ELV_MQUEUE_MAY;
4136
4137         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4138         if (cfqq) {
4139                 cfq_init_prio_data(cfqq, cic);
4140
4141                 return __cfq_may_queue(cfqq);
4142         }
4143
4144         return ELV_MQUEUE_MAY;
4145 }
4146
4147 /*
4148  * queue lock held here
4149  */
4150 static void cfq_put_request(struct request *rq)
4151 {
4152         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4153
4154         if (cfqq) {
4155                 const int rw = rq_data_dir(rq);
4156
4157                 BUG_ON(!cfqq->allocated[rw]);
4158                 cfqq->allocated[rw]--;
4159
4160                 /* Put down rq reference on cfqg */
4161                 cfqg_put(RQ_CFQG(rq));
4162                 rq->elv.priv[0] = NULL;
4163                 rq->elv.priv[1] = NULL;
4164
4165                 cfq_put_queue(cfqq);
4166         }
4167 }
4168
4169 static struct cfq_queue *
4170 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4171                 struct cfq_queue *cfqq)
4172 {
4173         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4174         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4175         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4176         cfq_put_queue(cfqq);
4177         return cic_to_cfqq(cic, 1);
4178 }
4179
4180 /*
4181  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4182  * was the last process referring to said cfqq.
4183  */
4184 static struct cfq_queue *
4185 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4186 {
4187         if (cfqq_process_refs(cfqq) == 1) {
4188                 cfqq->pid = current->pid;
4189                 cfq_clear_cfqq_coop(cfqq);
4190                 cfq_clear_cfqq_split_coop(cfqq);
4191                 return cfqq;
4192         }
4193
4194         cic_set_cfqq(cic, NULL, 1);
4195
4196         cfq_put_cooperator(cfqq);
4197
4198         cfq_put_queue(cfqq);
4199         return NULL;
4200 }
4201 /*
4202  * Allocate cfq data structures associated with this request.
4203  */
4204 static int
4205 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4206                 gfp_t gfp_mask)
4207 {
4208         struct cfq_data *cfqd = q->elevator->elevator_data;
4209         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4210         const int rw = rq_data_dir(rq);
4211         const bool is_sync = rq_is_sync(rq);
4212         struct cfq_queue *cfqq;
4213
4214         might_sleep_if(gfp_mask & __GFP_WAIT);
4215
4216         spin_lock_irq(q->queue_lock);
4217
4218         check_ioprio_changed(cic, bio);
4219         check_blkcg_changed(cic, bio);
4220 new_queue:
4221         cfqq = cic_to_cfqq(cic, is_sync);
4222         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4223                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4224                 cic_set_cfqq(cic, cfqq, is_sync);
4225         } else {
4226                 /*
4227                  * If the queue was seeky for too long, break it apart.
4228                  */
4229                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4230                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4231                         cfqq = split_cfqq(cic, cfqq);
4232                         if (!cfqq)
4233                                 goto new_queue;
4234                 }
4235
4236                 /*
4237                  * Check to see if this queue is scheduled to merge with
4238                  * another, closely cooperating queue.  The merging of
4239                  * queues happens here as it must be done in process context.
4240                  * The reference on new_cfqq was taken in merge_cfqqs.
4241                  */
4242                 if (cfqq->new_cfqq)
4243                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4244         }
4245
4246         cfqq->allocated[rw]++;
4247
4248         cfqq->ref++;
4249         cfqg_get(cfqq->cfqg);
4250         rq->elv.priv[0] = cfqq;
4251         rq->elv.priv[1] = cfqq->cfqg;
4252         spin_unlock_irq(q->queue_lock);
4253         return 0;
4254 }
4255
4256 static void cfq_kick_queue(struct work_struct *work)
4257 {
4258         struct cfq_data *cfqd =
4259                 container_of(work, struct cfq_data, unplug_work);
4260         struct request_queue *q = cfqd->queue;
4261
4262         spin_lock_irq(q->queue_lock);
4263         __blk_run_queue(cfqd->queue);
4264         spin_unlock_irq(q->queue_lock);
4265 }
4266
4267 /*
4268  * Timer running if the active_queue is currently idling inside its time slice
4269  */
4270 static void cfq_idle_slice_timer(unsigned long data)
4271 {
4272         struct cfq_data *cfqd = (struct cfq_data *) data;
4273         struct cfq_queue *cfqq;
4274         unsigned long flags;
4275         int timed_out = 1;
4276
4277         cfq_log(cfqd, "idle timer fired");
4278
4279         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4280
4281         cfqq = cfqd->active_queue;
4282         if (cfqq) {
4283                 timed_out = 0;
4284
4285                 /*
4286                  * We saw a request before the queue expired, let it through
4287                  */
4288                 if (cfq_cfqq_must_dispatch(cfqq))
4289                         goto out_kick;
4290
4291                 /*
4292                  * expired
4293                  */
4294                 if (cfq_slice_used(cfqq))
4295                         goto expire;
4296
4297                 /*
4298                  * only expire and reinvoke request handler, if there are
4299                  * other queues with pending requests
4300                  */
4301                 if (!cfqd->busy_queues)
4302                         goto out_cont;
4303
4304                 /*
4305                  * not expired and it has a request pending, let it dispatch
4306                  */
4307                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4308                         goto out_kick;
4309
4310                 /*
4311                  * Queue depth flag is reset only when the idle didn't succeed
4312                  */
4313                 cfq_clear_cfqq_deep(cfqq);
4314         }
4315 expire:
4316         cfq_slice_expired(cfqd, timed_out);
4317 out_kick:
4318         cfq_schedule_dispatch(cfqd);
4319 out_cont:
4320         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4321 }
4322
4323 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4324 {
4325         del_timer_sync(&cfqd->idle_slice_timer);
4326         cancel_work_sync(&cfqd->unplug_work);
4327 }
4328
4329 static void cfq_put_async_queues(struct cfq_data *cfqd)
4330 {
4331         int i;
4332
4333         for (i = 0; i < IOPRIO_BE_NR; i++) {
4334                 if (cfqd->async_cfqq[0][i])
4335                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4336                 if (cfqd->async_cfqq[1][i])
4337                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4338         }
4339
4340         if (cfqd->async_idle_cfqq)
4341                 cfq_put_queue(cfqd->async_idle_cfqq);
4342 }
4343
4344 static void cfq_exit_queue(struct elevator_queue *e)
4345 {
4346         struct cfq_data *cfqd = e->elevator_data;
4347         struct request_queue *q = cfqd->queue;
4348
4349         cfq_shutdown_timer_wq(cfqd);
4350
4351         spin_lock_irq(q->queue_lock);
4352
4353         if (cfqd->active_queue)
4354                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4355
4356         cfq_put_async_queues(cfqd);
4357
4358         spin_unlock_irq(q->queue_lock);
4359
4360         cfq_shutdown_timer_wq(cfqd);
4361
4362 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4363         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4364 #else
4365         kfree(cfqd->root_group);
4366 #endif
4367         kfree(cfqd);
4368 }
4369
4370 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4371 {
4372         struct cfq_data *cfqd;
4373         struct blkcg_gq *blkg __maybe_unused;
4374         int i, ret;
4375         struct elevator_queue *eq;
4376
4377         eq = elevator_alloc(q, e);
4378         if (!eq)
4379                 return -ENOMEM;
4380
4381         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4382         if (!cfqd) {
4383                 kobject_put(&eq->kobj);
4384                 return -ENOMEM;
4385         }
4386         eq->elevator_data = cfqd;
4387
4388         cfqd->queue = q;
4389         spin_lock_irq(q->queue_lock);
4390         q->elevator = eq;
4391         spin_unlock_irq(q->queue_lock);
4392
4393         /* Init root service tree */
4394         cfqd->grp_service_tree = CFQ_RB_ROOT;
4395
4396         /* Init root group and prefer root group over other groups by default */
4397 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4398         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4399         if (ret)
4400                 goto out_free;
4401
4402         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4403 #else
4404         ret = -ENOMEM;
4405         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4406                                         GFP_KERNEL, cfqd->queue->node);
4407         if (!cfqd->root_group)
4408                 goto out_free;
4409
4410         cfq_init_cfqg_base(cfqd->root_group);
4411 #endif
4412         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4413         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4414
4415         /*
4416          * Not strictly needed (since RB_ROOT just clears the node and we
4417          * zeroed cfqd on alloc), but better be safe in case someone decides
4418          * to add magic to the rb code
4419          */
4420         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4421                 cfqd->prio_trees[i] = RB_ROOT;
4422
4423         /*
4424          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4425          * Grab a permanent reference to it, so that the normal code flow
4426          * will not attempt to free it.  oom_cfqq is linked to root_group
4427          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4428          * the reference from linking right away.
4429          */
4430         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4431         cfqd->oom_cfqq.ref++;
4432
4433         spin_lock_irq(q->queue_lock);
4434         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4435         cfqg_put(cfqd->root_group);
4436         spin_unlock_irq(q->queue_lock);
4437
4438         init_timer(&cfqd->idle_slice_timer);
4439         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4440         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4441
4442         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4443
4444         cfqd->cfq_quantum = cfq_quantum;
4445         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4446         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4447         cfqd->cfq_back_max = cfq_back_max;
4448         cfqd->cfq_back_penalty = cfq_back_penalty;
4449         cfqd->cfq_slice[0] = cfq_slice_async;
4450         cfqd->cfq_slice[1] = cfq_slice_sync;
4451         cfqd->cfq_target_latency = cfq_target_latency;
4452         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4453         cfqd->cfq_slice_idle = cfq_slice_idle;
4454         cfqd->cfq_group_idle = cfq_group_idle;
4455         cfqd->cfq_latency = 1;
4456         cfqd->hw_tag = -1;
4457         /*
4458          * we optimistically start assuming sync ops weren't delayed in last
4459          * second, in order to have larger depth for async operations.
4460          */
4461         cfqd->last_delayed_sync = jiffies - HZ;
4462         return 0;
4463
4464 out_free:
4465         kfree(cfqd);
4466         kobject_put(&eq->kobj);
4467         return ret;
4468 }
4469
4470 /*
4471  * sysfs parts below -->
4472  */
4473 static ssize_t
4474 cfq_var_show(unsigned int var, char *page)
4475 {
4476         return sprintf(page, "%u\n", var);
4477 }
4478
4479 static ssize_t
4480 cfq_var_store(unsigned int *var, const char *page, size_t count)
4481 {
4482         char *p = (char *) page;
4483
4484         *var = simple_strtoul(p, &p, 10);
4485         return count;
4486 }
4487
4488 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4489 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4490 {                                                                       \
4491         struct cfq_data *cfqd = e->elevator_data;                       \
4492         unsigned int __data = __VAR;                                    \
4493         if (__CONV)                                                     \
4494                 __data = jiffies_to_msecs(__data);                      \
4495         return cfq_var_show(__data, (page));                            \
4496 }
4497 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4498 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4499 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4500 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4501 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4502 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4503 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4504 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4505 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4506 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4507 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4508 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4509 #undef SHOW_FUNCTION
4510
4511 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4512 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4513 {                                                                       \
4514         struct cfq_data *cfqd = e->elevator_data;                       \
4515         unsigned int __data;                                            \
4516         int ret = cfq_var_store(&__data, (page), count);                \
4517         if (__data < (MIN))                                             \
4518                 __data = (MIN);                                         \
4519         else if (__data > (MAX))                                        \
4520                 __data = (MAX);                                         \
4521         if (__CONV)                                                     \
4522                 *(__PTR) = msecs_to_jiffies(__data);                    \
4523         else                                                            \
4524                 *(__PTR) = __data;                                      \
4525         return ret;                                                     \
4526 }
4527 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4528 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4529                 UINT_MAX, 1);
4530 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4531                 UINT_MAX, 1);
4532 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4533 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4534                 UINT_MAX, 0);
4535 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4536 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4537 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4538 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4539 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4540                 UINT_MAX, 0);
4541 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4542 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4543 #undef STORE_FUNCTION
4544
4545 #define CFQ_ATTR(name) \
4546         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4547
4548 static struct elv_fs_entry cfq_attrs[] = {
4549         CFQ_ATTR(quantum),
4550         CFQ_ATTR(fifo_expire_sync),
4551         CFQ_ATTR(fifo_expire_async),
4552         CFQ_ATTR(back_seek_max),
4553         CFQ_ATTR(back_seek_penalty),
4554         CFQ_ATTR(slice_sync),
4555         CFQ_ATTR(slice_async),
4556         CFQ_ATTR(slice_async_rq),
4557         CFQ_ATTR(slice_idle),
4558         CFQ_ATTR(group_idle),
4559         CFQ_ATTR(low_latency),
4560         CFQ_ATTR(target_latency),
4561         __ATTR_NULL
4562 };
4563
4564 static struct elevator_type iosched_cfq = {
4565         .ops = {
4566                 .elevator_merge_fn =            cfq_merge,
4567                 .elevator_merged_fn =           cfq_merged_request,
4568                 .elevator_merge_req_fn =        cfq_merged_requests,
4569                 .elevator_allow_merge_fn =      cfq_allow_merge,
4570                 .elevator_bio_merged_fn =       cfq_bio_merged,
4571                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4572                 .elevator_add_req_fn =          cfq_insert_request,
4573                 .elevator_activate_req_fn =     cfq_activate_request,
4574                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4575                 .elevator_completed_req_fn =    cfq_completed_request,
4576                 .elevator_former_req_fn =       elv_rb_former_request,
4577                 .elevator_latter_req_fn =       elv_rb_latter_request,
4578                 .elevator_init_icq_fn =         cfq_init_icq,
4579                 .elevator_exit_icq_fn =         cfq_exit_icq,
4580                 .elevator_set_req_fn =          cfq_set_request,
4581                 .elevator_put_req_fn =          cfq_put_request,
4582                 .elevator_may_queue_fn =        cfq_may_queue,
4583                 .elevator_init_fn =             cfq_init_queue,
4584                 .elevator_exit_fn =             cfq_exit_queue,
4585         },
4586         .icq_size       =       sizeof(struct cfq_io_cq),
4587         .icq_align      =       __alignof__(struct cfq_io_cq),
4588         .elevator_attrs =       cfq_attrs,
4589         .elevator_name  =       "cfq",
4590         .elevator_owner =       THIS_MODULE,
4591 };
4592
4593 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4594 static struct blkcg_policy blkcg_policy_cfq = {
4595         .pd_size                = sizeof(struct cfq_group),
4596         .cftypes                = cfq_blkcg_files,
4597
4598         .pd_init_fn             = cfq_pd_init,
4599         .pd_offline_fn          = cfq_pd_offline,
4600         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4601 };
4602 #endif
4603
4604 static int __init cfq_init(void)
4605 {
4606         int ret;
4607
4608         /*
4609          * could be 0 on HZ < 1000 setups
4610          */
4611         if (!cfq_slice_async)
4612                 cfq_slice_async = 1;
4613         if (!cfq_slice_idle)
4614                 cfq_slice_idle = 1;
4615
4616 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4617         if (!cfq_group_idle)
4618                 cfq_group_idle = 1;
4619
4620         ret = blkcg_policy_register(&blkcg_policy_cfq);
4621         if (ret)
4622                 return ret;
4623 #else
4624         cfq_group_idle = 0;
4625 #endif
4626
4627         ret = -ENOMEM;
4628         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4629         if (!cfq_pool)
4630                 goto err_pol_unreg;
4631
4632         ret = elv_register(&iosched_cfq);
4633         if (ret)
4634                 goto err_free_pool;
4635
4636         return 0;
4637
4638 err_free_pool:
4639         kmem_cache_destroy(cfq_pool);
4640 err_pol_unreg:
4641 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4642         blkcg_policy_unregister(&blkcg_policy_cfq);
4643 #endif
4644         return ret;
4645 }
4646
4647 static void __exit cfq_exit(void)
4648 {
4649 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4650         blkcg_policy_unregister(&blkcg_policy_cfq);
4651 #endif
4652         elv_unregister(&iosched_cfq);
4653         kmem_cache_destroy(cfq_pool);
4654 }
4655
4656 module_init(cfq_init);
4657 module_exit(cfq_exit);
4658
4659 MODULE_AUTHOR("Jens Axboe");
4660 MODULE_LICENSE("GPL");
4661 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");