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 <linux/blk-cgroup.h>
  18 #include "blk.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 /* blkio-related constants */
  71 #define CFQ_WEIGHT_MIN          10
  72 #define CFQ_WEIGHT_MAX          1000
  73 #define CFQ_WEIGHT_DEFAULT      500
  74
  75 struct cfq_ttime {
  76         unsigned long last_end_request;
  77
  78         unsigned long ttime_total;
  79         unsigned long ttime_samples;
  80         unsigned long ttime_mean;
  81 };
  82
  83 /*
  84  * Most of our rbtree usage is for sorting with min extraction, so
  85  * if we cache the leftmost node we don't have to walk down the tree
  86  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  87  * move this into the elevator for the rq sorting as well.
  88  */
  89 struct cfq_rb_root {
  90         struct rb_root rb;
  91         struct rb_node *left;
  92         unsigned count;
  93         u64 min_vdisktime;
  94         struct cfq_ttime ttime;
  95 };
  96 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  97                         .ttime = {.last_end_request = jiffies,},}
  98
  99 /*
 100  * Per process-grouping structure
 101  */
 102 struct cfq_queue {
 103         /* reference count */
 104         int ref;
 105         /* various state flags, see below */
 106         unsigned int flags;
 107         /* parent cfq_data */
 108         struct cfq_data *cfqd;
 109         /* service_tree member */
 110         struct rb_node rb_node;
 111         /* service_tree key */
 112         unsigned long rb_key;
 113         /* prio tree member */
 114         struct rb_node p_node;
 115         /* prio tree root we belong to, if any */
 116         struct rb_root *p_root;
 117         /* sorted list of pending requests */
 118         struct rb_root sort_list;
 119         /* if fifo isn't expired, next request to serve */
 120         struct request *next_rq;
 121         /* requests queued in sort_list */
 122         int queued[2];
 123         /* currently allocated requests */
 124         int allocated[2];
 125         /* fifo list of requests in sort_list */
 126         struct list_head fifo;
 127
 128         /* time when queue got scheduled in to dispatch first request. */
 129         unsigned long dispatch_start;
 130         unsigned int allocated_slice;
 131         unsigned int slice_dispatch;
 132         /* time when first request from queue completed and slice started. */
 133         unsigned long slice_start;
 134         unsigned long slice_end;
 135         long slice_resid;
 136
 137         /* pending priority requests */
 138         int prio_pending;
 139         /* number of requests that are on the dispatch list or inside driver */
 140         int dispatched;
 141
 142         /* io prio of this group */
 143         unsigned short ioprio, org_ioprio;
 144         unsigned short ioprio_class;
 145
 146         pid_t pid;
 147
 148         u32 seek_history;
 149         sector_t last_request_pos;
 150
 151         struct cfq_rb_root *service_tree;
 152         struct cfq_queue *new_cfqq;
 153         struct cfq_group *cfqg;
 154         /* Number of sectors dispatched from queue in single dispatch round */
 155         unsigned long nr_sectors;
 156 };
 157
 158 /*
 159  * First index in the service_trees.
 160  * IDLE is handled separately, so it has negative index
 161  */
 162 enum wl_class_t {
 163         BE_WORKLOAD = 0,
 164         RT_WORKLOAD = 1,
 165         IDLE_WORKLOAD = 2,
 166         CFQ_PRIO_NR,
 167 };
 168
 169 /*
 170  * Second index in the service_trees.
 171  */
 172 enum wl_type_t {
 173         ASYNC_WORKLOAD = 0,
 174         SYNC_NOIDLE_WORKLOAD = 1,
 175         SYNC_WORKLOAD = 2
 176 };
 177
 178 struct cfqg_stats {
 179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 180         /* total bytes transferred */
 181         struct blkg_rwstat              service_bytes;
 182         /* total IOs serviced, post merge */
 183         struct blkg_rwstat              serviced;
 184         /* number of ios merged */
 185         struct blkg_rwstat              merged;
 186         /* total time spent on device in ns, may not be accurate w/ queueing */
 187         struct blkg_rwstat              service_time;
 188         /* total time spent waiting in scheduler queue in ns */
 189         struct blkg_rwstat              wait_time;
 190         /* number of IOs queued up */
 191         struct blkg_rwstat              queued;
 192         /* total sectors transferred */
 193         struct blkg_stat                sectors;
 194         /* total disk time and nr sectors dispatched by this group */
 195         struct blkg_stat                time;
 196 #ifdef CONFIG_DEBUG_BLK_CGROUP
 197         /* time not charged to this cgroup */
 198         struct blkg_stat                unaccounted_time;
 199         /* sum of number of ios queued across all samples */
 200         struct blkg_stat                avg_queue_size_sum;
 201         /* count of samples taken for average */
 202         struct blkg_stat                avg_queue_size_samples;
 203         /* how many times this group has been removed from service tree */
 204         struct blkg_stat                dequeue;
 205         /* total time spent waiting for it to be assigned a timeslice. */
 206         struct blkg_stat                group_wait_time;
 207         /* time spent idling for this blkcg_gq */
 208         struct blkg_stat                idle_time;
 209         /* total time with empty current active q with other requests queued */
 210         struct blkg_stat                empty_time;
 211         /* fields after this shouldn't be cleared on stat reset */
 212         uint64_t                        start_group_wait_time;
 213         uint64_t                        start_idle_time;
 214         uint64_t                        start_empty_time;
 215         uint16_t                        flags;
 216 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 217 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 218 };
 219
 220 /* Per-cgroup data */
 221 struct cfq_group_data {
 222         /* must be the first member */
 223         struct blkcg_policy_data cpd;
 224
 225         unsigned int weight;
 226         unsigned int leaf_weight;
 227 };
 228
 229 /* This is per cgroup per device grouping structure */
 230 struct cfq_group {
 231         /* must be the first member */
 232         struct blkg_policy_data pd;
 233
 234         /* group service_tree member */
 235         struct rb_node rb_node;
 236
 237         /* group service_tree key */
 238         u64 vdisktime;
 239
 240         /*
 241          * The number of active cfqgs and sum of their weights under this
 242          * cfqg.  This covers this cfqg's leaf_weight and all children's
 243          * weights, but does not cover weights of further descendants.
 244          *
 245          * If a cfqg is on the service tree, it's active.  An active cfqg
 246          * also activates its parent and contributes to the children_weight
 247          * of the parent.
 248          */
 249         int nr_active;
 250         unsigned int children_weight;
 251
 252         /*
 253          * vfraction is the fraction of vdisktime that the tasks in this
 254          * cfqg are entitled to.  This is determined by compounding the
 255          * ratios walking up from this cfqg to the root.
 256          *
 257          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 258          * vfractions on a service tree is approximately 1.  The sum may
 259          * deviate a bit due to rounding errors and fluctuations caused by
 260          * cfqgs entering and leaving the service tree.
 261          */
 262         unsigned int vfraction;
 263
 264         /*
 265          * There are two weights - (internal) weight is the weight of this
 266          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 267          * this cfqg against the child cfqgs.  For the root cfqg, both
 268          * weights are kept in sync for backward compatibility.
 269          */
 270         unsigned int weight;
 271         unsigned int new_weight;
 272         unsigned int dev_weight;
 273
 274         unsigned int leaf_weight;
 275         unsigned int new_leaf_weight;
 276         unsigned int dev_leaf_weight;
 277
 278         /* number of cfqq currently on this group */
 279         int nr_cfqq;
 280
 281         /*
 282          * Per group busy queues average. Useful for workload slice calc. We
 283          * create the array for each prio class but at run time it is used
 284          * only for RT and BE class and slot for IDLE class remains unused.
 285          * This is primarily done to avoid confusion and a gcc warning.
 286          */
 287         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 288         /*
 289          * rr lists of queues with requests. We maintain service trees for
 290          * RT and BE classes. These trees are subdivided in subclasses
 291          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 292          * class there is no subclassification and all the cfq queues go on
 293          * a single tree service_tree_idle.
 294          * Counts are embedded in the cfq_rb_root
 295          */
 296         struct cfq_rb_root service_trees[2][3];
 297         struct cfq_rb_root service_tree_idle;
 298
 299         unsigned long saved_wl_slice;
 300         enum wl_type_t saved_wl_type;
 301         enum wl_class_t saved_wl_class;
 302
 303         /* number of requests that are on the dispatch list or inside driver */
 304         int dispatched;
 305         struct cfq_ttime ttime;
 306         struct cfqg_stats stats;        /* stats for this cfqg */
 307
 308         /* async queue for each priority case */
 309         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 310         struct cfq_queue *async_idle_cfqq;
 311
 312 };
 313
 314 struct cfq_io_cq {
 315         struct io_cq            icq;            /* must be the first member */
 316         struct cfq_queue        *cfqq[2];
 317         struct cfq_ttime        ttime;
 318         int                     ioprio;         /* the current ioprio */
 319 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 320         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 321 #endif
 322 };
 323
 324 /*
 325  * Per block device queue structure
 326  */
 327 struct cfq_data {
 328         struct request_queue *queue;
 329         /* Root service tree for cfq_groups */
 330         struct cfq_rb_root grp_service_tree;
 331         struct cfq_group *root_group;
 332
 333         /*
 334          * The priority currently being served
 335          */
 336         enum wl_class_t serving_wl_class;
 337         enum wl_type_t serving_wl_type;
 338         unsigned long workload_expires;
 339         struct cfq_group *serving_group;
 340
 341         /*
 342          * Each priority tree is sorted by next_request position.  These
 343          * trees are used when determining if two or more queues are
 344          * interleaving requests (see cfq_close_cooperator).
 345          */
 346         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 347
 348         unsigned int busy_queues;
 349         unsigned int busy_sync_queues;
 350
 351         int rq_in_driver;
 352         int rq_in_flight[2];
 353
 354         /*
 355          * queue-depth detection
 356          */
 357         int rq_queued;
 358         int hw_tag;
 359         /*
 360          * hw_tag can be
 361          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 362          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 363          *  0 => no NCQ
 364          */
 365         int hw_tag_est_depth;
 366         unsigned int hw_tag_samples;
 367
 368         /*
 369          * idle window management
 370          */
 371         struct timer_list idle_slice_timer;
 372         struct work_struct unplug_work;
 373
 374         struct cfq_queue *active_queue;
 375         struct cfq_io_cq *active_cic;
 376
 377         sector_t last_position;
 378
 379         /*
 380          * tunables, see top of file
 381          */
 382         unsigned int cfq_quantum;
 383         unsigned int cfq_fifo_expire[2];
 384         unsigned int cfq_back_penalty;
 385         unsigned int cfq_back_max;
 386         unsigned int cfq_slice[2];
 387         unsigned int cfq_slice_async_rq;
 388         unsigned int cfq_slice_idle;
 389         unsigned int cfq_group_idle;
 390         unsigned int cfq_latency;
 391         unsigned int cfq_target_latency;
 392
 393         /*
 394          * Fallback dummy cfqq for extreme OOM conditions
 395          */
 396         struct cfq_queue oom_cfqq;
 397
 398         unsigned long last_delayed_sync;
 399 };
 400
 401 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 402 static void cfq_put_queue(struct cfq_queue *cfqq);
 403
 404 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 405                                             enum wl_class_t class,
 406                                             enum wl_type_t type)
 407 {
 408         if (!cfqg)
 409                 return NULL;
 410
 411         if (class == IDLE_WORKLOAD)
 412                 return &cfqg->service_tree_idle;
 413
 414         return &cfqg->service_trees[class][type];
 415 }
 416
 417 enum cfqq_state_flags {
 418         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 419         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 420         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 421         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 422         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 423         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 424         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 425         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 426         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 427         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 428         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 429         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 430         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 431 };
 432
 433 #define CFQ_CFQQ_FNS(name)                                              \
 434 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 435 {                                                                       \
 436         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 437 }                                                                       \
 438 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 439 {                                                                       \
 440         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 441 }                                                                       \
 442 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 443 {                                                                       \
 444         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 445 }
 446
 447 CFQ_CFQQ_FNS(on_rr);
 448 CFQ_CFQQ_FNS(wait_request);
 449 CFQ_CFQQ_FNS(must_dispatch);
 450 CFQ_CFQQ_FNS(must_alloc_slice);
 451 CFQ_CFQQ_FNS(fifo_expire);
 452 CFQ_CFQQ_FNS(idle_window);
 453 CFQ_CFQQ_FNS(prio_changed);
 454 CFQ_CFQQ_FNS(slice_new);
 455 CFQ_CFQQ_FNS(sync);
 456 CFQ_CFQQ_FNS(coop);
 457 CFQ_CFQQ_FNS(split_coop);
 458 CFQ_CFQQ_FNS(deep);
 459 CFQ_CFQQ_FNS(wait_busy);
 460 #undef CFQ_CFQQ_FNS
 461
 462 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 463
 464 /* cfqg stats flags */
 465 enum cfqg_stats_flags {
 466         CFQG_stats_waiting = 0,
 467         CFQG_stats_idling,
 468         CFQG_stats_empty,
 469 };
 470
 471 #define CFQG_FLAG_FNS(name)                                             \
 472 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 473 {                                                                       \
 474         stats->flags |= (1 << CFQG_stats_##name);                       \
 475 }                                                                       \
 476 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 477 {                                                                       \
 478         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 479 }                                                                       \
 480 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 481 {                                                                       \
 482         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 483 }                                                                       \
 484
 485 CFQG_FLAG_FNS(waiting)
 486 CFQG_FLAG_FNS(idling)
 487 CFQG_FLAG_FNS(empty)
 488 #undef CFQG_FLAG_FNS
 489
 490 /* This should be called with the queue_lock held. */
 491 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 492 {
 493         unsigned long long now;
 494
 495         if (!cfqg_stats_waiting(stats))
 496                 return;
 497
 498         now = sched_clock();
 499         if (time_after64(now, stats->start_group_wait_time))
 500                 blkg_stat_add(&stats->group_wait_time,
 501                               now - stats->start_group_wait_time);
 502         cfqg_stats_clear_waiting(stats);
 503 }
 504
 505 /* This should be called with the queue_lock held. */
 506 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 507                                                  struct cfq_group *curr_cfqg)
 508 {
 509         struct cfqg_stats *stats = &cfqg->stats;
 510
 511         if (cfqg_stats_waiting(stats))
 512                 return;
 513         if (cfqg == curr_cfqg)
 514                 return;
 515         stats->start_group_wait_time = sched_clock();
 516         cfqg_stats_mark_waiting(stats);
 517 }
 518
 519 /* This should be called with the queue_lock held. */
 520 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 521 {
 522         unsigned long long now;
 523
 524         if (!cfqg_stats_empty(stats))
 525                 return;
 526
 527         now = sched_clock();
 528         if (time_after64(now, stats->start_empty_time))
 529                 blkg_stat_add(&stats->empty_time,
 530                               now - stats->start_empty_time);
 531         cfqg_stats_clear_empty(stats);
 532 }
 533
 534 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 535 {
 536         blkg_stat_add(&cfqg->stats.dequeue, 1);
 537 }
 538
 539 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 540 {
 541         struct cfqg_stats *stats = &cfqg->stats;
 542
 543         if (blkg_rwstat_total(&stats->queued))
 544                 return;
 545
 546         /*
 547          * group is already marked empty. This can happen if cfqq got new
 548          * request in parent group and moved to this group while being added
 549          * to service tree. Just ignore the event and move on.
 550          */
 551         if (cfqg_stats_empty(stats))
 552                 return;
 553
 554         stats->start_empty_time = sched_clock();
 555         cfqg_stats_mark_empty(stats);
 556 }
 557
 558 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 559 {
 560         struct cfqg_stats *stats = &cfqg->stats;
 561
 562         if (cfqg_stats_idling(stats)) {
 563                 unsigned long long now = sched_clock();
 564
 565                 if (time_after64(now, stats->start_idle_time))
 566                         blkg_stat_add(&stats->idle_time,
 567                                       now - stats->start_idle_time);
 568                 cfqg_stats_clear_idling(stats);
 569         }
 570 }
 571
 572 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 573 {
 574         struct cfqg_stats *stats = &cfqg->stats;
 575
 576         BUG_ON(cfqg_stats_idling(stats));
 577
 578         stats->start_idle_time = sched_clock();
 579         cfqg_stats_mark_idling(stats);
 580 }
 581
 582 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 583 {
 584         struct cfqg_stats *stats = &cfqg->stats;
 585
 586         blkg_stat_add(&stats->avg_queue_size_sum,
 587                       blkg_rwstat_total(&stats->queued));
 588         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 589         cfqg_stats_update_group_wait_time(stats);
 590 }
 591
 592 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 593
 594 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 595 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 596 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 597 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 598 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 599 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 600 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 601
 602 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 603
 604 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 605
 606 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 607 {
 608         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 609 }
 610
 611 static struct cfq_group_data
 612 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 613 {
 614         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 615 }
 616
 617 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 618 {
 619         return pd_to_blkg(&cfqg->pd);
 620 }
 621
 622 static struct blkcg_policy blkcg_policy_cfq;
 623
 624 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 625 {
 626         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 627 }
 628
 629 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 630 {
 631         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 632 }
 633
 634 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 635 {
 636         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 637
 638         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 639 }
 640
 641 static inline void cfqg_get(struct cfq_group *cfqg)
 642 {
 643         return blkg_get(cfqg_to_blkg(cfqg));
 644 }
 645
 646 static inline void cfqg_put(struct cfq_group *cfqg)
 647 {
 648         return blkg_put(cfqg_to_blkg(cfqg));
 649 }
 650
 651 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 652         char __pbuf[128];                                               \
 653                                                                         \
 654         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 655         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 656                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 657                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 658                           __pbuf, ##args);                              \
 659 } while (0)
 660
 661 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 662         char __pbuf[128];                                               \
 663                                                                         \
 664         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 665         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 666 } while (0)
 667
 668 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 669                                             struct cfq_group *curr_cfqg, int rw)
 670 {
 671         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 672         cfqg_stats_end_empty_time(&cfqg->stats);
 673         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 674 }
 675
 676 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 677                         unsigned long time, unsigned long unaccounted_time)
 678 {
 679         blkg_stat_add(&cfqg->stats.time, time);
 680 #ifdef CONFIG_DEBUG_BLK_CGROUP
 681         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 682 #endif
 683 }
 684
 685 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 686 {
 687         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 688 }
 689
 690 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 691 {
 692         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 693 }
 694
 695 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 696                                               uint64_t bytes, int rw)
 697 {
 698         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
 699         blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
 700         blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
 701 }
 702
 703 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 704                         uint64_t start_time, uint64_t io_start_time, int rw)
 705 {
 706         struct cfqg_stats *stats = &cfqg->stats;
 707         unsigned long long now = sched_clock();
 708
 709         if (time_after64(now, io_start_time))
 710                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 711         if (time_after64(io_start_time, start_time))
 712                 blkg_rwstat_add(&stats->wait_time, rw,
 713                                 io_start_time - start_time);
 714 }
 715
 716 /* @stats = 0 */
 717 static void cfqg_stats_reset(struct cfqg_stats *stats)
 718 {
 719         /* queued stats shouldn't be cleared */
 720         blkg_rwstat_reset(&stats->service_bytes);
 721         blkg_rwstat_reset(&stats->serviced);
 722         blkg_rwstat_reset(&stats->merged);
 723         blkg_rwstat_reset(&stats->service_time);
 724         blkg_rwstat_reset(&stats->wait_time);
 725         blkg_stat_reset(&stats->time);
 726 #ifdef CONFIG_DEBUG_BLK_CGROUP
 727         blkg_stat_reset(&stats->unaccounted_time);
 728         blkg_stat_reset(&stats->avg_queue_size_sum);
 729         blkg_stat_reset(&stats->avg_queue_size_samples);
 730         blkg_stat_reset(&stats->dequeue);
 731         blkg_stat_reset(&stats->group_wait_time);
 732         blkg_stat_reset(&stats->idle_time);
 733         blkg_stat_reset(&stats->empty_time);
 734 #endif
 735 }
 736
 737 /* @to += @from */
 738 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 739 {
 740         /* queued stats shouldn't be cleared */
 741         blkg_rwstat_add_aux(&to->service_bytes, &from->service_bytes);
 742         blkg_rwstat_add_aux(&to->serviced, &from->serviced);
 743         blkg_rwstat_add_aux(&to->merged, &from->merged);
 744         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 745         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 746         blkg_stat_add_aux(&from->time, &from->time);
 747 #ifdef CONFIG_DEBUG_BLK_CGROUP
 748         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 749         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 750         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 751         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 752         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 753         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 754         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 755 #endif
 756 }
 757
 758 /*
 759  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 760  * recursive stats can still account for the amount used by this cfqg after
 761  * it's gone.
 762  */
 763 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 764 {
 765         struct cfq_group *parent = cfqg_parent(cfqg);
 766
 767         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 768
 769         if (unlikely(!parent))
 770                 return;
 771
 772         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 773         cfqg_stats_reset(&cfqg->stats);
 774 }
 775
 776 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 777
 778 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 779 static inline void cfqg_get(struct cfq_group *cfqg) { }
 780 static inline void cfqg_put(struct cfq_group *cfqg) { }
 781
 782 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 783         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 784                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 785                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 786                                 ##args)
 787 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 788
 789 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 790                         struct cfq_group *curr_cfqg, int rw) { }
 791 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 792                         unsigned long time, unsigned long unaccounted_time) { }
 793 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 794 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 795 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 796                                               uint64_t bytes, int rw) { }
 797 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 798                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 799
 800 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 801
 802 #define cfq_log(cfqd, fmt, args...)     \
 803         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 804
 805 /* Traverses through cfq group service trees */
 806 #define for_each_cfqg_st(cfqg, i, j, st) \
 807         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 808                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 809                         : &cfqg->service_tree_idle; \
 810                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 811                         (i == IDLE_WORKLOAD && j == 0); \
 812                         j++, st = i < IDLE_WORKLOAD ? \
 813                         &cfqg->service_trees[i][j]: NULL) \
 814
 815 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 816         struct cfq_ttime *ttime, bool group_idle)
 817 {
 818         unsigned long slice;
 819         if (!sample_valid(ttime->ttime_samples))
 820                 return false;
 821         if (group_idle)
 822                 slice = cfqd->cfq_group_idle;
 823         else
 824                 slice = cfqd->cfq_slice_idle;
 825         return ttime->ttime_mean > slice;
 826 }
 827
 828 static inline bool iops_mode(struct cfq_data *cfqd)
 829 {
 830         /*
 831          * If we are not idling on queues and it is a NCQ drive, parallel
 832          * execution of requests is on and measuring time is not possible
 833          * in most of the cases until and unless we drive shallower queue
 834          * depths and that becomes a performance bottleneck. In such cases
 835          * switch to start providing fairness in terms of number of IOs.
 836          */
 837         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 838                 return true;
 839         else
 840                 return false;
 841 }
 842
 843 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 844 {
 845         if (cfq_class_idle(cfqq))
 846                 return IDLE_WORKLOAD;
 847         if (cfq_class_rt(cfqq))
 848                 return RT_WORKLOAD;
 849         return BE_WORKLOAD;
 850 }
 851
 852
 853 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 854 {
 855         if (!cfq_cfqq_sync(cfqq))
 856                 return ASYNC_WORKLOAD;
 857         if (!cfq_cfqq_idle_window(cfqq))
 858                 return SYNC_NOIDLE_WORKLOAD;
 859         return SYNC_WORKLOAD;
 860 }
 861
 862 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 863                                         struct cfq_data *cfqd,
 864                                         struct cfq_group *cfqg)
 865 {
 866         if (wl_class == IDLE_WORKLOAD)
 867                 return cfqg->service_tree_idle.count;
 868
 869         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 870                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 871                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 872 }
 873
 874 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 875                                         struct cfq_group *cfqg)
 876 {
 877         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 878                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 879 }
 880
 881 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 882 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 883                                        struct cfq_io_cq *cic, struct bio *bio);
 884
 885 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 886 {
 887         /* cic->icq is the first member, %NULL will convert to %NULL */
 888         return container_of(icq, struct cfq_io_cq, icq);
 889 }
 890
 891 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 892                                                struct io_context *ioc)
 893 {
 894         if (ioc)
 895                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 896         return NULL;
 897 }
 898
 899 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 900 {
 901         return cic->cfqq[is_sync];
 902 }
 903
 904 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 905                                 bool is_sync)
 906 {
 907         cic->cfqq[is_sync] = cfqq;
 908 }
 909
 910 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 911 {
 912         return cic->icq.q->elevator->elevator_data;
 913 }
 914
 915 /*
 916  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 917  * set (in which case it could also be direct WRITE).
 918  */
 919 static inline bool cfq_bio_sync(struct bio *bio)
 920 {
 921         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 922 }
 923
 924 /*
 925  * scheduler run of queue, if there are requests pending and no one in the
 926  * driver that will restart queueing
 927  */
 928 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 929 {
 930         if (cfqd->busy_queues) {
 931                 cfq_log(cfqd, "schedule dispatch");
 932                 kblockd_schedule_work(&cfqd->unplug_work);
 933         }
 934 }
 935
 936 /*
 937  * Scale schedule slice based on io priority. Use the sync time slice only
 938  * if a queue is marked sync and has sync io queued. A sync queue with async
 939  * io only, should not get full sync slice length.
 940  */
 941 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 942                                  unsigned short prio)
 943 {
 944         const int base_slice = cfqd->cfq_slice[sync];
 945
 946         WARN_ON(prio >= IOPRIO_BE_NR);
 947
 948         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 949 }
 950
 951 static inline int
 952 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 953 {
 954         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 955 }
 956
 957 /**
 958  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 959  * @charge: disk time being charged
 960  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 961  *
 962  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 963  * scaling is inversely proportional.
 964  *
 965  * scaled = charge / vfraction
 966  *
 967  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 968  */
 969 static inline u64 cfqg_scale_charge(unsigned long charge,
 970                                     unsigned int vfraction)
 971 {
 972         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 973
 974         /* charge / vfraction */
 975         c <<= CFQ_SERVICE_SHIFT;
 976         do_div(c, vfraction);
 977         return c;
 978 }
 979
 980 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 981 {
 982         s64 delta = (s64)(vdisktime - min_vdisktime);
 983         if (delta > 0)
 984                 min_vdisktime = vdisktime;
 985
 986         return min_vdisktime;
 987 }
 988
 989 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 990 {
 991         s64 delta = (s64)(vdisktime - min_vdisktime);
 992         if (delta < 0)
 993                 min_vdisktime = vdisktime;
 994
 995         return min_vdisktime;
 996 }
 997
 998 static void update_min_vdisktime(struct cfq_rb_root *st)
 999 {
1000         struct cfq_group *cfqg;
1001
1002         if (st->left) {
1003                 cfqg = rb_entry_cfqg(st->left);
1004                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1005                                                   cfqg->vdisktime);
1006         }
1007 }
1008
1009 /*
1010  * get averaged number of queues of RT/BE priority.
1011  * average is updated, with a formula that gives more weight to higher numbers,
1012  * to quickly follows sudden increases and decrease slowly
1013  */
1014
1015 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1016                                         struct cfq_group *cfqg, bool rt)
1017 {
1018         unsigned min_q, max_q;
1019         unsigned mult  = cfq_hist_divisor - 1;
1020         unsigned round = cfq_hist_divisor / 2;
1021         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1022
1023         min_q = min(cfqg->busy_queues_avg[rt], busy);
1024         max_q = max(cfqg->busy_queues_avg[rt], busy);
1025         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1026                 cfq_hist_divisor;
1027         return cfqg->busy_queues_avg[rt];
1028 }
1029
1030 static inline unsigned
1031 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1032 {
1033         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1034 }
1035
1036 static inline unsigned
1037 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1038 {
1039         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1040         if (cfqd->cfq_latency) {
1041                 /*
1042                  * interested queues (we consider only the ones with the same
1043                  * priority class in the cfq group)
1044                  */
1045                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1046                                                 cfq_class_rt(cfqq));
1047                 unsigned sync_slice = cfqd->cfq_slice[1];
1048                 unsigned expect_latency = sync_slice * iq;
1049                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1050
1051                 if (expect_latency > group_slice) {
1052                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1053                         /* scale low_slice according to IO priority
1054                          * and sync vs async */
1055                         unsigned low_slice =
1056                                 min(slice, base_low_slice * slice / sync_slice);
1057                         /* the adapted slice value is scaled to fit all iqs
1058                          * into the target latency */
1059                         slice = max(slice * group_slice / expect_latency,
1060                                     low_slice);
1061                 }
1062         }
1063         return slice;
1064 }
1065
1066 static inline void
1067 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1068 {
1069         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1070
1071         cfqq->slice_start = jiffies;
1072         cfqq->slice_end = jiffies + slice;
1073         cfqq->allocated_slice = slice;
1074         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1075 }
1076
1077 /*
1078  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1079  * isn't valid until the first request from the dispatch is activated
1080  * and the slice time set.
1081  */
1082 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1083 {
1084         if (cfq_cfqq_slice_new(cfqq))
1085                 return false;
1086         if (time_before(jiffies, cfqq->slice_end))
1087                 return false;
1088
1089         return true;
1090 }
1091
1092 /*
1093  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1094  * We choose the request that is closest to the head right now. Distance
1095  * behind the head is penalized and only allowed to a certain extent.
1096  */
1097 static struct request *
1098 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1099 {
1100         sector_t s1, s2, d1 = 0, d2 = 0;
1101         unsigned long back_max;
1102 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1103 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1104         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1105
1106         if (rq1 == NULL || rq1 == rq2)
1107                 return rq2;
1108         if (rq2 == NULL)
1109                 return rq1;
1110
1111         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1112                 return rq_is_sync(rq1) ? rq1 : rq2;
1113
1114         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1115                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1116
1117         s1 = blk_rq_pos(rq1);
1118         s2 = blk_rq_pos(rq2);
1119
1120         /*
1121          * by definition, 1KiB is 2 sectors
1122          */
1123         back_max = cfqd->cfq_back_max * 2;
1124
1125         /*
1126          * Strict one way elevator _except_ in the case where we allow
1127          * short backward seeks which are biased as twice the cost of a
1128          * similar forward seek.
1129          */
1130         if (s1 >= last)
1131                 d1 = s1 - last;
1132         else if (s1 + back_max >= last)
1133                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1134         else
1135                 wrap |= CFQ_RQ1_WRAP;
1136
1137         if (s2 >= last)
1138                 d2 = s2 - last;
1139         else if (s2 + back_max >= last)
1140                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1141         else
1142                 wrap |= CFQ_RQ2_WRAP;
1143
1144         /* Found required data */
1145
1146         /*
1147          * By doing switch() on the bit mask "wrap" we avoid having to
1148          * check two variables for all permutations: --> faster!
1149          */
1150         switch (wrap) {
1151         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1152                 if (d1 < d2)
1153                         return rq1;
1154                 else if (d2 < d1)
1155                         return rq2;
1156                 else {
1157                         if (s1 >= s2)
1158                                 return rq1;
1159                         else
1160                                 return rq2;
1161                 }
1162
1163         case CFQ_RQ2_WRAP:
1164                 return rq1;
1165         case CFQ_RQ1_WRAP:
1166                 return rq2;
1167         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1168         default:
1169                 /*
1170                  * Since both rqs are wrapped,
1171                  * start with the one that's further behind head
1172                  * (--> only *one* back seek required),
1173                  * since back seek takes more time than forward.
1174                  */
1175                 if (s1 <= s2)
1176                         return rq1;
1177                 else
1178                         return rq2;
1179         }
1180 }
1181
1182 /*
1183  * The below is leftmost cache rbtree addon
1184  */
1185 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1186 {
1187         /* Service tree is empty */
1188         if (!root->count)
1189                 return NULL;
1190
1191         if (!root->left)
1192                 root->left = rb_first(&root->rb);
1193
1194         if (root->left)
1195                 return rb_entry(root->left, struct cfq_queue, rb_node);
1196
1197         return NULL;
1198 }
1199
1200 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1201 {
1202         if (!root->left)
1203                 root->left = rb_first(&root->rb);
1204
1205         if (root->left)
1206                 return rb_entry_cfqg(root->left);
1207
1208         return NULL;
1209 }
1210
1211 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1212 {
1213         rb_erase(n, root);
1214         RB_CLEAR_NODE(n);
1215 }
1216
1217 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1218 {
1219         if (root->left == n)
1220                 root->left = NULL;
1221         rb_erase_init(n, &root->rb);
1222         --root->count;
1223 }
1224
1225 /*
1226  * would be nice to take fifo expire time into account as well
1227  */
1228 static struct request *
1229 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1230                   struct request *last)
1231 {
1232         struct rb_node *rbnext = rb_next(&last->rb_node);
1233         struct rb_node *rbprev = rb_prev(&last->rb_node);
1234         struct request *next = NULL, *prev = NULL;
1235
1236         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1237
1238         if (rbprev)
1239                 prev = rb_entry_rq(rbprev);
1240
1241         if (rbnext)
1242                 next = rb_entry_rq(rbnext);
1243         else {
1244                 rbnext = rb_first(&cfqq->sort_list);
1245                 if (rbnext && rbnext != &last->rb_node)
1246                         next = rb_entry_rq(rbnext);
1247         }
1248
1249         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1250 }
1251
1252 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1253                                       struct cfq_queue *cfqq)
1254 {
1255         /*
1256          * just an approximation, should be ok.
1257          */
1258         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1259                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1260 }
1261
1262 static inline s64
1263 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1264 {
1265         return cfqg->vdisktime - st->min_vdisktime;
1266 }
1267
1268 static void
1269 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1270 {
1271         struct rb_node **node = &st->rb.rb_node;
1272         struct rb_node *parent = NULL;
1273         struct cfq_group *__cfqg;
1274         s64 key = cfqg_key(st, cfqg);
1275         int left = 1;
1276
1277         while (*node != NULL) {
1278                 parent = *node;
1279                 __cfqg = rb_entry_cfqg(parent);
1280
1281                 if (key < cfqg_key(st, __cfqg))
1282                         node = &parent->rb_left;
1283                 else {
1284                         node = &parent->rb_right;
1285                         left = 0;
1286                 }
1287         }
1288
1289         if (left)
1290                 st->left = &cfqg->rb_node;
1291
1292         rb_link_node(&cfqg->rb_node, parent, node);
1293         rb_insert_color(&cfqg->rb_node, &st->rb);
1294 }
1295
1296 /*
1297  * This has to be called only on activation of cfqg
1298  */
1299 static void
1300 cfq_update_group_weight(struct cfq_group *cfqg)
1301 {
1302         if (cfqg->new_weight) {
1303                 cfqg->weight = cfqg->new_weight;
1304                 cfqg->new_weight = 0;
1305         }
1306 }
1307
1308 static void
1309 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1310 {
1311         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1312
1313         if (cfqg->new_leaf_weight) {
1314                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1315                 cfqg->new_leaf_weight = 0;
1316         }
1317 }
1318
1319 static void
1320 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1321 {
1322         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1323         struct cfq_group *pos = cfqg;
1324         struct cfq_group *parent;
1325         bool propagate;
1326
1327         /* add to the service tree */
1328         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1329
1330         /*
1331          * Update leaf_weight.  We cannot update weight at this point
1332          * because cfqg might already have been activated and is
1333          * contributing its current weight to the parent's child_weight.
1334          */
1335         cfq_update_group_leaf_weight(cfqg);
1336         __cfq_group_service_tree_add(st, cfqg);
1337
1338         /*
1339          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1340          * entitled to.  vfraction is calculated by walking the tree
1341          * towards the root calculating the fraction it has at each level.
1342          * The compounded ratio is how much vfraction @cfqg owns.
1343          *
1344          * Start with the proportion tasks in this cfqg has against active
1345          * children cfqgs - its leaf_weight against children_weight.
1346          */
1347         propagate = !pos->nr_active++;
1348         pos->children_weight += pos->leaf_weight;
1349         vfr = vfr * pos->leaf_weight / pos->children_weight;
1350
1351         /*
1352          * Compound ->weight walking up the tree.  Both activation and
1353          * vfraction calculation are done in the same loop.  Propagation
1354          * stops once an already activated node is met.  vfraction
1355          * calculation should always continue to the root.
1356          */
1357         while ((parent = cfqg_parent(pos))) {
1358                 if (propagate) {
1359                         cfq_update_group_weight(pos);
1360                         propagate = !parent->nr_active++;
1361                         parent->children_weight += pos->weight;
1362                 }
1363                 vfr = vfr * pos->weight / parent->children_weight;
1364                 pos = parent;
1365         }
1366
1367         cfqg->vfraction = max_t(unsigned, vfr, 1);
1368 }
1369
1370 static void
1371 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1372 {
1373         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1374         struct cfq_group *__cfqg;
1375         struct rb_node *n;
1376
1377         cfqg->nr_cfqq++;
1378         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1379                 return;
1380
1381         /*
1382          * Currently put the group at the end. Later implement something
1383          * so that groups get lesser vtime based on their weights, so that
1384          * if group does not loose all if it was not continuously backlogged.
1385          */
1386         n = rb_last(&st->rb);
1387         if (n) {
1388                 __cfqg = rb_entry_cfqg(n);
1389                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1390         } else
1391                 cfqg->vdisktime = st->min_vdisktime;
1392         cfq_group_service_tree_add(st, cfqg);
1393 }
1394
1395 static void
1396 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1397 {
1398         struct cfq_group *pos = cfqg;
1399         bool propagate;
1400
1401         /*
1402          * Undo activation from cfq_group_service_tree_add().  Deactivate
1403          * @cfqg and propagate deactivation upwards.
1404          */
1405         propagate = !--pos->nr_active;
1406         pos->children_weight -= pos->leaf_weight;
1407
1408         while (propagate) {
1409                 struct cfq_group *parent = cfqg_parent(pos);
1410
1411                 /* @pos has 0 nr_active at this point */
1412                 WARN_ON_ONCE(pos->children_weight);
1413                 pos->vfraction = 0;
1414
1415                 if (!parent)
1416                         break;
1417
1418                 propagate = !--parent->nr_active;
1419                 parent->children_weight -= pos->weight;
1420                 pos = parent;
1421         }
1422
1423         /* remove from the service tree */
1424         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1425                 cfq_rb_erase(&cfqg->rb_node, st);
1426 }
1427
1428 static void
1429 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1430 {
1431         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1432
1433         BUG_ON(cfqg->nr_cfqq < 1);
1434         cfqg->nr_cfqq--;
1435
1436         /* If there are other cfq queues under this group, don't delete it */
1437         if (cfqg->nr_cfqq)
1438                 return;
1439
1440         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1441         cfq_group_service_tree_del(st, cfqg);
1442         cfqg->saved_wl_slice = 0;
1443         cfqg_stats_update_dequeue(cfqg);
1444 }
1445
1446 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1447                                                 unsigned int *unaccounted_time)
1448 {
1449         unsigned int slice_used;
1450
1451         /*
1452          * Queue got expired before even a single request completed or
1453          * got expired immediately after first request completion.
1454          */
1455         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1456                 /*
1457                  * Also charge the seek time incurred to the group, otherwise
1458                  * if there are mutiple queues in the group, each can dispatch
1459                  * a single request on seeky media and cause lots of seek time
1460                  * and group will never know it.
1461                  */
1462                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1463                                         1);
1464         } else {
1465                 slice_used = jiffies - cfqq->slice_start;
1466                 if (slice_used > cfqq->allocated_slice) {
1467                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1468                         slice_used = cfqq->allocated_slice;
1469                 }
1470                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1471                         *unaccounted_time += cfqq->slice_start -
1472                                         cfqq->dispatch_start;
1473         }
1474
1475         return slice_used;
1476 }
1477
1478 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1479                                 struct cfq_queue *cfqq)
1480 {
1481         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1482         unsigned int used_sl, charge, unaccounted_sl = 0;
1483         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1484                         - cfqg->service_tree_idle.count;
1485         unsigned int vfr;
1486
1487         BUG_ON(nr_sync < 0);
1488         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1489
1490         if (iops_mode(cfqd))
1491                 charge = cfqq->slice_dispatch;
1492         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1493                 charge = cfqq->allocated_slice;
1494
1495         /*
1496          * Can't update vdisktime while on service tree and cfqg->vfraction
1497          * is valid only while on it.  Cache vfr, leave the service tree,
1498          * update vdisktime and go back on.  The re-addition to the tree
1499          * will also update the weights as necessary.
1500          */
1501         vfr = cfqg->vfraction;
1502         cfq_group_service_tree_del(st, cfqg);
1503         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1504         cfq_group_service_tree_add(st, cfqg);
1505
1506         /* This group is being expired. Save the context */
1507         if (time_after(cfqd->workload_expires, jiffies)) {
1508                 cfqg->saved_wl_slice = cfqd->workload_expires
1509                                                 - jiffies;
1510                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1511                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1512         } else
1513                 cfqg->saved_wl_slice = 0;
1514
1515         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1516                                         st->min_vdisktime);
1517         cfq_log_cfqq(cfqq->cfqd, cfqq,
1518                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1519                      used_sl, cfqq->slice_dispatch, charge,
1520                      iops_mode(cfqd), cfqq->nr_sectors);
1521         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1522         cfqg_stats_set_start_empty_time(cfqg);
1523 }
1524
1525 /**
1526  * cfq_init_cfqg_base - initialize base part of a cfq_group
1527  * @cfqg: cfq_group to initialize
1528  *
1529  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1530  * is enabled or not.
1531  */
1532 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1533 {
1534         struct cfq_rb_root *st;
1535         int i, j;
1536
1537         for_each_cfqg_st(cfqg, i, j, st)
1538                 *st = CFQ_RB_ROOT;
1539         RB_CLEAR_NODE(&cfqg->rb_node);
1540
1541         cfqg->ttime.last_end_request = jiffies;
1542 }
1543
1544 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1545 static void cfqg_stats_init(struct cfqg_stats *stats)
1546 {
1547         blkg_rwstat_init(&stats->service_bytes);
1548         blkg_rwstat_init(&stats->serviced);
1549         blkg_rwstat_init(&stats->merged);
1550         blkg_rwstat_init(&stats->service_time);
1551         blkg_rwstat_init(&stats->wait_time);
1552         blkg_rwstat_init(&stats->queued);
1553
1554         blkg_stat_init(&stats->sectors);
1555         blkg_stat_init(&stats->time);
1556
1557 #ifdef CONFIG_DEBUG_BLK_CGROUP
1558         blkg_stat_init(&stats->unaccounted_time);
1559         blkg_stat_init(&stats->avg_queue_size_sum);
1560         blkg_stat_init(&stats->avg_queue_size_samples);
1561         blkg_stat_init(&stats->dequeue);
1562         blkg_stat_init(&stats->group_wait_time);
1563         blkg_stat_init(&stats->idle_time);
1564         blkg_stat_init(&stats->empty_time);
1565 #endif
1566 }
1567
1568 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1569 {
1570         struct cfq_group_data *cgd;
1571
1572         cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1573         if (!cgd)
1574                 return NULL;
1575         return &cgd->cpd;
1576 }
1577
1578 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1579 {
1580         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1581
1582         if (cpd_to_blkcg(cpd) == &blkcg_root) {
1583                 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1584                 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1585         } else {
1586                 cgd->weight = CFQ_WEIGHT_DEFAULT;
1587                 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1588         }
1589 }
1590
1591 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1592 {
1593         kfree(cpd_to_cfqgd(cpd));
1594 }
1595
1596 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1597 {
1598         struct cfq_group *cfqg;
1599
1600         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1601         if (!cfqg)
1602                 return NULL;
1603
1604         cfq_init_cfqg_base(cfqg);
1605         cfqg_stats_init(&cfqg->stats);
1606
1607         return &cfqg->pd;
1608 }
1609
1610 static void cfq_pd_init(struct blkg_policy_data *pd)
1611 {
1612         struct cfq_group *cfqg = pd_to_cfqg(pd);
1613         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1614
1615         cfqg->weight = cgd->weight;
1616         cfqg->leaf_weight = cgd->leaf_weight;
1617 }
1618
1619 static void cfq_pd_offline(struct blkg_policy_data *pd)
1620 {
1621         struct cfq_group *cfqg = pd_to_cfqg(pd);
1622         int i;
1623
1624         for (i = 0; i < IOPRIO_BE_NR; i++) {
1625                 if (cfqg->async_cfqq[0][i])
1626                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1627                 if (cfqg->async_cfqq[1][i])
1628                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1629         }
1630
1631         if (cfqg->async_idle_cfqq)
1632                 cfq_put_queue(cfqg->async_idle_cfqq);
1633
1634         /*
1635          * @blkg is going offline and will be ignored by
1636          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1637          * that they don't get lost.  If IOs complete after this point, the
1638          * stats for them will be lost.  Oh well...
1639          */
1640         cfqg_stats_xfer_dead(cfqg);
1641 }
1642
1643 static void cfq_pd_free(struct blkg_policy_data *pd)
1644 {
1645         return kfree(pd);
1646 }
1647
1648 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1649 {
1650         struct cfq_group *cfqg = pd_to_cfqg(pd);
1651
1652         cfqg_stats_reset(&cfqg->stats);
1653 }
1654
1655 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1656                                          struct blkcg *blkcg)
1657 {
1658         struct blkcg_gq *blkg;
1659
1660         blkg = blkg_lookup(blkcg, cfqd->queue);
1661         if (likely(blkg))
1662                 return blkg_to_cfqg(blkg);
1663         return NULL;
1664 }
1665
1666 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1667 {
1668         cfqq->cfqg = cfqg;
1669         /* cfqq reference on cfqg */
1670         cfqg_get(cfqg);
1671 }
1672
1673 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1674                                      struct blkg_policy_data *pd, int off)
1675 {
1676         struct cfq_group *cfqg = pd_to_cfqg(pd);
1677
1678         if (!cfqg->dev_weight)
1679                 return 0;
1680         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1681 }
1682
1683 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1684 {
1685         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1686                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1687                           0, false);
1688         return 0;
1689 }
1690
1691 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1692                                           struct blkg_policy_data *pd, int off)
1693 {
1694         struct cfq_group *cfqg = pd_to_cfqg(pd);
1695
1696         if (!cfqg->dev_leaf_weight)
1697                 return 0;
1698         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1699 }
1700
1701 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1702 {
1703         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1704                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1705                           0, false);
1706         return 0;
1707 }
1708
1709 static int cfq_print_weight(struct seq_file *sf, void *v)
1710 {
1711         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1712         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1713         unsigned int val = 0;
1714
1715         if (cgd)
1716                 val = cgd->weight;
1717
1718         seq_printf(sf, "%u\n", val);
1719         return 0;
1720 }
1721
1722 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1723 {
1724         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1725         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1726         unsigned int val = 0;
1727
1728         if (cgd)
1729                 val = cgd->leaf_weight;
1730
1731         seq_printf(sf, "%u\n", val);
1732         return 0;
1733 }
1734
1735 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1736                                         char *buf, size_t nbytes, loff_t off,
1737                                         bool is_leaf_weight)
1738 {
1739         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1740         struct blkg_conf_ctx ctx;
1741         struct cfq_group *cfqg;
1742         struct cfq_group_data *cfqgd;
1743         int ret;
1744
1745         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1746         if (ret)
1747                 return ret;
1748
1749         ret = -EINVAL;
1750         cfqg = blkg_to_cfqg(ctx.blkg);
1751         cfqgd = blkcg_to_cfqgd(blkcg);
1752         if (!cfqg || !cfqgd)
1753                 goto err;
1754
1755         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1756                 if (!is_leaf_weight) {
1757                         cfqg->dev_weight = ctx.v;
1758                         cfqg->new_weight = ctx.v ?: cfqgd->weight;
1759                 } else {
1760                         cfqg->dev_leaf_weight = ctx.v;
1761                         cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1762                 }
1763                 ret = 0;
1764         }
1765
1766 err:
1767         blkg_conf_finish(&ctx);
1768         return ret ?: nbytes;
1769 }
1770
1771 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1772                                       char *buf, size_t nbytes, loff_t off)
1773 {
1774         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1775 }
1776
1777 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1778                                            char *buf, size_t nbytes, loff_t off)
1779 {
1780         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1781 }
1782
1783 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1784                             u64 val, bool is_leaf_weight)
1785 {
1786         struct blkcg *blkcg = css_to_blkcg(css);
1787         struct blkcg_gq *blkg;
1788         struct cfq_group_data *cfqgd;
1789         int ret = 0;
1790
1791         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1792                 return -EINVAL;
1793
1794         spin_lock_irq(&blkcg->lock);
1795         cfqgd = blkcg_to_cfqgd(blkcg);
1796         if (!cfqgd) {
1797                 ret = -EINVAL;
1798                 goto out;
1799         }
1800
1801         if (!is_leaf_weight)
1802                 cfqgd->weight = val;
1803         else
1804                 cfqgd->leaf_weight = val;
1805
1806         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1807                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1808
1809                 if (!cfqg)
1810                         continue;
1811
1812                 if (!is_leaf_weight) {
1813                         if (!cfqg->dev_weight)
1814                                 cfqg->new_weight = cfqgd->weight;
1815                 } else {
1816                         if (!cfqg->dev_leaf_weight)
1817                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1818                 }
1819         }
1820
1821 out:
1822         spin_unlock_irq(&blkcg->lock);
1823         return ret;
1824 }
1825
1826 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1827                           u64 val)
1828 {
1829         return __cfq_set_weight(css, cft, val, false);
1830 }
1831
1832 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1833                                struct cftype *cft, u64 val)
1834 {
1835         return __cfq_set_weight(css, cft, val, true);
1836 }
1837
1838 static int cfqg_print_stat(struct seq_file *sf, void *v)
1839 {
1840         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1841                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1842         return 0;
1843 }
1844
1845 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1846 {
1847         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1848                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1849         return 0;
1850 }
1851
1852 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1853                                       struct blkg_policy_data *pd, int off)
1854 {
1855         u64 sum = blkg_stat_recursive_sum(pd, off);
1856
1857         return __blkg_prfill_u64(sf, pd, sum);
1858 }
1859
1860 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1861                                         struct blkg_policy_data *pd, int off)
1862 {
1863         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd, off);
1864
1865         return __blkg_prfill_rwstat(sf, pd, &sum);
1866 }
1867
1868 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1869 {
1870         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1871                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1872                           seq_cft(sf)->private, false);
1873         return 0;
1874 }
1875
1876 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1877 {
1878         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1879                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1880                           seq_cft(sf)->private, true);
1881         return 0;
1882 }
1883
1884 #ifdef CONFIG_DEBUG_BLK_CGROUP
1885 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1886                                       struct blkg_policy_data *pd, int off)
1887 {
1888         struct cfq_group *cfqg = pd_to_cfqg(pd);
1889         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1890         u64 v = 0;
1891
1892         if (samples) {
1893                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1894                 v = div64_u64(v, samples);
1895         }
1896         __blkg_prfill_u64(sf, pd, v);
1897         return 0;
1898 }
1899
1900 /* print avg_queue_size */
1901 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1902 {
1903         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1904                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1905                           0, false);
1906         return 0;
1907 }
1908 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1909
1910 static struct cftype cfq_blkcg_files[] = {
1911         /* on root, weight is mapped to leaf_weight */
1912         {
1913                 .name = "weight_device",
1914                 .flags = CFTYPE_ONLY_ON_ROOT,
1915                 .seq_show = cfqg_print_leaf_weight_device,
1916                 .write = cfqg_set_leaf_weight_device,
1917         },
1918         {
1919                 .name = "weight",
1920                 .flags = CFTYPE_ONLY_ON_ROOT,
1921                 .seq_show = cfq_print_leaf_weight,
1922                 .write_u64 = cfq_set_leaf_weight,
1923         },
1924
1925         /* no such mapping necessary for !roots */
1926         {
1927                 .name = "weight_device",
1928                 .flags = CFTYPE_NOT_ON_ROOT,
1929                 .seq_show = cfqg_print_weight_device,
1930                 .write = cfqg_set_weight_device,
1931         },
1932         {
1933                 .name = "weight",
1934                 .flags = CFTYPE_NOT_ON_ROOT,
1935                 .seq_show = cfq_print_weight,
1936                 .write_u64 = cfq_set_weight,
1937         },
1938
1939         {
1940                 .name = "leaf_weight_device",
1941                 .seq_show = cfqg_print_leaf_weight_device,
1942                 .write = cfqg_set_leaf_weight_device,
1943         },
1944         {
1945                 .name = "leaf_weight",
1946                 .seq_show = cfq_print_leaf_weight,
1947                 .write_u64 = cfq_set_leaf_weight,
1948         },
1949
1950         /* statistics, covers only the tasks in the cfqg */
1951         {
1952                 .name = "time",
1953                 .private = offsetof(struct cfq_group, stats.time),
1954                 .seq_show = cfqg_print_stat,
1955         },
1956         {
1957                 .name = "sectors",
1958                 .private = offsetof(struct cfq_group, stats.sectors),
1959                 .seq_show = cfqg_print_stat,
1960         },
1961         {
1962                 .name = "io_service_bytes",
1963                 .private = offsetof(struct cfq_group, stats.service_bytes),
1964                 .seq_show = cfqg_print_rwstat,
1965         },
1966         {
1967                 .name = "io_serviced",
1968                 .private = offsetof(struct cfq_group, stats.serviced),
1969                 .seq_show = cfqg_print_rwstat,
1970         },
1971         {
1972                 .name = "io_service_time",
1973                 .private = offsetof(struct cfq_group, stats.service_time),
1974                 .seq_show = cfqg_print_rwstat,
1975         },
1976         {
1977                 .name = "io_wait_time",
1978                 .private = offsetof(struct cfq_group, stats.wait_time),
1979                 .seq_show = cfqg_print_rwstat,
1980         },
1981         {
1982                 .name = "io_merged",
1983                 .private = offsetof(struct cfq_group, stats.merged),
1984                 .seq_show = cfqg_print_rwstat,
1985         },
1986         {
1987                 .name = "io_queued",
1988                 .private = offsetof(struct cfq_group, stats.queued),
1989                 .seq_show = cfqg_print_rwstat,
1990         },
1991
1992         /* the same statictics which cover the cfqg and its descendants */
1993         {
1994                 .name = "time_recursive",
1995                 .private = offsetof(struct cfq_group, stats.time),
1996                 .seq_show = cfqg_print_stat_recursive,
1997         },
1998         {
1999                 .name = "sectors_recursive",
2000                 .private = offsetof(struct cfq_group, stats.sectors),
2001                 .seq_show = cfqg_print_stat_recursive,
2002         },
2003         {
2004                 .name = "io_service_bytes_recursive",
2005                 .private = offsetof(struct cfq_group, stats.service_bytes),
2006                 .seq_show = cfqg_print_rwstat_recursive,
2007         },
2008         {
2009                 .name = "io_serviced_recursive",
2010                 .private = offsetof(struct cfq_group, stats.serviced),
2011                 .seq_show = cfqg_print_rwstat_recursive,
2012         },
2013         {
2014                 .name = "io_service_time_recursive",
2015                 .private = offsetof(struct cfq_group, stats.service_time),
2016                 .seq_show = cfqg_print_rwstat_recursive,
2017         },
2018         {
2019                 .name = "io_wait_time_recursive",
2020                 .private = offsetof(struct cfq_group, stats.wait_time),
2021                 .seq_show = cfqg_print_rwstat_recursive,
2022         },
2023         {
2024                 .name = "io_merged_recursive",
2025                 .private = offsetof(struct cfq_group, stats.merged),
2026                 .seq_show = cfqg_print_rwstat_recursive,
2027         },
2028         {
2029                 .name = "io_queued_recursive",
2030                 .private = offsetof(struct cfq_group, stats.queued),
2031                 .seq_show = cfqg_print_rwstat_recursive,
2032         },
2033 #ifdef CONFIG_DEBUG_BLK_CGROUP
2034         {
2035                 .name = "avg_queue_size",
2036                 .seq_show = cfqg_print_avg_queue_size,
2037         },
2038         {
2039                 .name = "group_wait_time",
2040                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2041                 .seq_show = cfqg_print_stat,
2042         },
2043         {
2044                 .name = "idle_time",
2045                 .private = offsetof(struct cfq_group, stats.idle_time),
2046                 .seq_show = cfqg_print_stat,
2047         },
2048         {
2049                 .name = "empty_time",
2050                 .private = offsetof(struct cfq_group, stats.empty_time),
2051                 .seq_show = cfqg_print_stat,
2052         },
2053         {
2054                 .name = "dequeue",
2055                 .private = offsetof(struct cfq_group, stats.dequeue),
2056                 .seq_show = cfqg_print_stat,
2057         },
2058         {
2059                 .name = "unaccounted_time",
2060                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2061                 .seq_show = cfqg_print_stat,
2062         },
2063 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2064         { }     /* terminate */
2065 };
2066 #else /* GROUP_IOSCHED */
2067 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2068                                          struct blkcg *blkcg)
2069 {
2070         return cfqd->root_group;
2071 }
2072
2073 static inline void
2074 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2075         cfqq->cfqg = cfqg;
2076 }
2077
2078 #endif /* GROUP_IOSCHED */
2079
2080 /*
2081  * The cfqd->service_trees holds all pending cfq_queue's that have
2082  * requests waiting to be processed. It is sorted in the order that
2083  * we will service the queues.
2084  */
2085 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2086                                  bool add_front)
2087 {
2088         struct rb_node **p, *parent;
2089         struct cfq_queue *__cfqq;
2090         unsigned long rb_key;
2091         struct cfq_rb_root *st;
2092         int left;
2093         int new_cfqq = 1;
2094
2095         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2096         if (cfq_class_idle(cfqq)) {
2097                 rb_key = CFQ_IDLE_DELAY;
2098                 parent = rb_last(&st->rb);
2099                 if (parent && parent != &cfqq->rb_node) {
2100                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2101                         rb_key += __cfqq->rb_key;
2102                 } else
2103                         rb_key += jiffies;
2104         } else if (!add_front) {
2105                 /*
2106                  * Get our rb key offset. Subtract any residual slice
2107                  * value carried from last service. A negative resid
2108                  * count indicates slice overrun, and this should position
2109                  * the next service time further away in the tree.
2110                  */
2111                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2112                 rb_key -= cfqq->slice_resid;
2113                 cfqq->slice_resid = 0;
2114         } else {
2115                 rb_key = -HZ;
2116                 __cfqq = cfq_rb_first(st);
2117                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2118         }
2119
2120         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2121                 new_cfqq = 0;
2122                 /*
2123                  * same position, nothing more to do
2124                  */
2125                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2126                         return;
2127
2128                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2129                 cfqq->service_tree = NULL;
2130         }
2131
2132         left = 1;
2133         parent = NULL;
2134         cfqq->service_tree = st;
2135         p = &st->rb.rb_node;
2136         while (*p) {
2137                 parent = *p;
2138                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2139
2140                 /*
2141                  * sort by key, that represents service time.
2142                  */
2143                 if (time_before(rb_key, __cfqq->rb_key))
2144                         p = &parent->rb_left;
2145                 else {
2146                         p = &parent->rb_right;
2147                         left = 0;
2148                 }
2149         }
2150
2151         if (left)
2152                 st->left = &cfqq->rb_node;
2153
2154         cfqq->rb_key = rb_key;
2155         rb_link_node(&cfqq->rb_node, parent, p);
2156         rb_insert_color(&cfqq->rb_node, &st->rb);
2157         st->count++;
2158         if (add_front || !new_cfqq)
2159                 return;
2160         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2161 }
2162
2163 static struct cfq_queue *
2164 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2165                      sector_t sector, struct rb_node **ret_parent,
2166                      struct rb_node ***rb_link)
2167 {
2168         struct rb_node **p, *parent;
2169         struct cfq_queue *cfqq = NULL;
2170
2171         parent = NULL;
2172         p = &root->rb_node;
2173         while (*p) {
2174                 struct rb_node **n;
2175
2176                 parent = *p;
2177                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2178
2179                 /*
2180                  * Sort strictly based on sector.  Smallest to the left,
2181                  * largest to the right.
2182                  */
2183                 if (sector > blk_rq_pos(cfqq->next_rq))
2184                         n = &(*p)->rb_right;
2185                 else if (sector < blk_rq_pos(cfqq->next_rq))
2186                         n = &(*p)->rb_left;
2187                 else
2188                         break;
2189                 p = n;
2190                 cfqq = NULL;
2191         }
2192
2193         *ret_parent = parent;
2194         if (rb_link)
2195                 *rb_link = p;
2196         return cfqq;
2197 }
2198
2199 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2200 {
2201         struct rb_node **p, *parent;
2202         struct cfq_queue *__cfqq;
2203
2204         if (cfqq->p_root) {
2205                 rb_erase(&cfqq->p_node, cfqq->p_root);
2206                 cfqq->p_root = NULL;
2207         }
2208
2209         if (cfq_class_idle(cfqq))
2210                 return;
2211         if (!cfqq->next_rq)
2212                 return;
2213
2214         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2215         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2216                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2217         if (!__cfqq) {
2218                 rb_link_node(&cfqq->p_node, parent, p);
2219                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2220         } else
2221                 cfqq->p_root = NULL;
2222 }
2223
2224 /*
2225  * Update cfqq's position in the service tree.
2226  */
2227 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2228 {
2229         /*
2230          * Resorting requires the cfqq to be on the RR list already.
2231          */
2232         if (cfq_cfqq_on_rr(cfqq)) {
2233                 cfq_service_tree_add(cfqd, cfqq, 0);
2234                 cfq_prio_tree_add(cfqd, cfqq);
2235         }
2236 }
2237
2238 /*
2239  * add to busy list of queues for service, trying to be fair in ordering
2240  * the pending list according to last request service
2241  */
2242 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2243 {
2244         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2245         BUG_ON(cfq_cfqq_on_rr(cfqq));
2246         cfq_mark_cfqq_on_rr(cfqq);
2247         cfqd->busy_queues++;
2248         if (cfq_cfqq_sync(cfqq))
2249                 cfqd->busy_sync_queues++;
2250
2251         cfq_resort_rr_list(cfqd, cfqq);
2252 }
2253
2254 /*
2255  * Called when the cfqq no longer has requests pending, remove it from
2256  * the service tree.
2257  */
2258 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2259 {
2260         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2261         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2262         cfq_clear_cfqq_on_rr(cfqq);
2263
2264         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2265                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2266                 cfqq->service_tree = NULL;
2267         }
2268         if (cfqq->p_root) {
2269                 rb_erase(&cfqq->p_node, cfqq->p_root);
2270                 cfqq->p_root = NULL;
2271         }
2272
2273         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2274         BUG_ON(!cfqd->busy_queues);
2275         cfqd->busy_queues--;
2276         if (cfq_cfqq_sync(cfqq))
2277                 cfqd->busy_sync_queues--;
2278 }
2279
2280 /*
2281  * rb tree support functions
2282  */
2283 static void cfq_del_rq_rb(struct request *rq)
2284 {
2285         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2286         const int sync = rq_is_sync(rq);
2287
2288         BUG_ON(!cfqq->queued[sync]);
2289         cfqq->queued[sync]--;
2290
2291         elv_rb_del(&cfqq->sort_list, rq);
2292
2293         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2294                 /*
2295                  * Queue will be deleted from service tree when we actually
2296                  * expire it later. Right now just remove it from prio tree
2297                  * as it is empty.
2298                  */
2299                 if (cfqq->p_root) {
2300                         rb_erase(&cfqq->p_node, cfqq->p_root);
2301                         cfqq->p_root = NULL;
2302                 }
2303         }
2304 }
2305
2306 static void cfq_add_rq_rb(struct request *rq)
2307 {
2308         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2309         struct cfq_data *cfqd = cfqq->cfqd;
2310         struct request *prev;
2311
2312         cfqq->queued[rq_is_sync(rq)]++;
2313
2314         elv_rb_add(&cfqq->sort_list, rq);
2315
2316         if (!cfq_cfqq_on_rr(cfqq))
2317                 cfq_add_cfqq_rr(cfqd, cfqq);
2318
2319         /*
2320          * check if this request is a better next-serve candidate
2321          */
2322         prev = cfqq->next_rq;
2323         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2324
2325         /*
2326          * adjust priority tree position, if ->next_rq changes
2327          */
2328         if (prev != cfqq->next_rq)
2329                 cfq_prio_tree_add(cfqd, cfqq);
2330
2331         BUG_ON(!cfqq->next_rq);
2332 }
2333
2334 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2335 {
2336         elv_rb_del(&cfqq->sort_list, rq);
2337         cfqq->queued[rq_is_sync(rq)]--;
2338         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2339         cfq_add_rq_rb(rq);
2340         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2341                                  rq->cmd_flags);
2342 }
2343
2344 static struct request *
2345 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2346 {
2347         struct task_struct *tsk = current;
2348         struct cfq_io_cq *cic;
2349         struct cfq_queue *cfqq;
2350
2351         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2352         if (!cic)
2353                 return NULL;
2354
2355         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2356         if (cfqq)
2357                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2358
2359         return NULL;
2360 }
2361
2362 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2363 {
2364         struct cfq_data *cfqd = q->elevator->elevator_data;
2365
2366         cfqd->rq_in_driver++;
2367         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2368                                                 cfqd->rq_in_driver);
2369
2370         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2371 }
2372
2373 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2374 {
2375         struct cfq_data *cfqd = q->elevator->elevator_data;
2376
2377         WARN_ON(!cfqd->rq_in_driver);
2378         cfqd->rq_in_driver--;
2379         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2380                                                 cfqd->rq_in_driver);
2381 }
2382
2383 static void cfq_remove_request(struct request *rq)
2384 {
2385         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2386
2387         if (cfqq->next_rq == rq)
2388                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2389
2390         list_del_init(&rq->queuelist);
2391         cfq_del_rq_rb(rq);
2392
2393         cfqq->cfqd->rq_queued--;
2394         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2395         if (rq->cmd_flags & REQ_PRIO) {
2396                 WARN_ON(!cfqq->prio_pending);
2397                 cfqq->prio_pending--;
2398         }
2399 }
2400
2401 static int cfq_merge(struct request_queue *q, struct request **req,
2402                      struct bio *bio)
2403 {
2404         struct cfq_data *cfqd = q->elevator->elevator_data;
2405         struct request *__rq;
2406
2407         __rq = cfq_find_rq_fmerge(cfqd, bio);
2408         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2409                 *req = __rq;
2410                 return ELEVATOR_FRONT_MERGE;
2411         }
2412
2413         return ELEVATOR_NO_MERGE;
2414 }
2415
2416 static void cfq_merged_request(struct request_queue *q, struct request *req,
2417                                int type)
2418 {
2419         if (type == ELEVATOR_FRONT_MERGE) {
2420                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2421
2422                 cfq_reposition_rq_rb(cfqq, req);
2423         }
2424 }
2425
2426 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2427                                 struct bio *bio)
2428 {
2429         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2430 }
2431
2432 static void
2433 cfq_merged_requests(struct request_queue *q, struct request *rq,
2434                     struct request *next)
2435 {
2436         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2437         struct cfq_data *cfqd = q->elevator->elevator_data;
2438
2439         /*
2440          * reposition in fifo if next is older than rq
2441          */
2442         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2443             time_before(next->fifo_time, rq->fifo_time) &&
2444             cfqq == RQ_CFQQ(next)) {
2445                 list_move(&rq->queuelist, &next->queuelist);
2446                 rq->fifo_time = next->fifo_time;
2447         }
2448
2449         if (cfqq->next_rq == next)
2450                 cfqq->next_rq = rq;
2451         cfq_remove_request(next);
2452         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2453
2454         cfqq = RQ_CFQQ(next);
2455         /*
2456          * all requests of this queue are merged to other queues, delete it
2457          * from the service tree. If it's the active_queue,
2458          * cfq_dispatch_requests() will choose to expire it or do idle
2459          */
2460         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2461             cfqq != cfqd->active_queue)
2462                 cfq_del_cfqq_rr(cfqd, cfqq);
2463 }
2464
2465 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2466                            struct bio *bio)
2467 {
2468         struct cfq_data *cfqd = q->elevator->elevator_data;
2469         struct cfq_io_cq *cic;
2470         struct cfq_queue *cfqq;
2471
2472         /*
2473          * Disallow merge of a sync bio into an async request.
2474          */
2475         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2476                 return false;
2477
2478         /*
2479          * Lookup the cfqq that this bio will be queued with and allow
2480          * merge only if rq is queued there.
2481          */
2482         cic = cfq_cic_lookup(cfqd, current->io_context);
2483         if (!cic)
2484                 return false;
2485
2486         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2487         return cfqq == RQ_CFQQ(rq);
2488 }
2489
2490 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2491 {
2492         del_timer(&cfqd->idle_slice_timer);
2493         cfqg_stats_update_idle_time(cfqq->cfqg);
2494 }
2495
2496 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2497                                    struct cfq_queue *cfqq)
2498 {
2499         if (cfqq) {
2500                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2501                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2502                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2503                 cfqq->slice_start = 0;
2504                 cfqq->dispatch_start = jiffies;
2505                 cfqq->allocated_slice = 0;
2506                 cfqq->slice_end = 0;
2507                 cfqq->slice_dispatch = 0;
2508                 cfqq->nr_sectors = 0;
2509
2510                 cfq_clear_cfqq_wait_request(cfqq);
2511                 cfq_clear_cfqq_must_dispatch(cfqq);
2512                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2513                 cfq_clear_cfqq_fifo_expire(cfqq);
2514                 cfq_mark_cfqq_slice_new(cfqq);
2515
2516                 cfq_del_timer(cfqd, cfqq);
2517         }
2518
2519         cfqd->active_queue = cfqq;
2520 }
2521
2522 /*
2523  * current cfqq expired its slice (or was too idle), select new one
2524  */
2525 static void
2526 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2527                     bool timed_out)
2528 {
2529         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2530
2531         if (cfq_cfqq_wait_request(cfqq))
2532                 cfq_del_timer(cfqd, cfqq);
2533
2534         cfq_clear_cfqq_wait_request(cfqq);
2535         cfq_clear_cfqq_wait_busy(cfqq);
2536
2537         /*
2538          * If this cfqq is shared between multiple processes, check to
2539          * make sure that those processes are still issuing I/Os within
2540          * the mean seek distance.  If not, it may be time to break the
2541          * queues apart again.
2542          */
2543         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2544                 cfq_mark_cfqq_split_coop(cfqq);
2545
2546         /*
2547          * store what was left of this slice, if the queue idled/timed out
2548          */
2549         if (timed_out) {
2550                 if (cfq_cfqq_slice_new(cfqq))
2551                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2552                 else
2553                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2554                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2555         }
2556
2557         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2558
2559         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2560                 cfq_del_cfqq_rr(cfqd, cfqq);
2561
2562         cfq_resort_rr_list(cfqd, cfqq);
2563
2564         if (cfqq == cfqd->active_queue)
2565                 cfqd->active_queue = NULL;
2566
2567         if (cfqd->active_cic) {
2568                 put_io_context(cfqd->active_cic->icq.ioc);
2569                 cfqd->active_cic = NULL;
2570         }
2571 }
2572
2573 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2574 {
2575         struct cfq_queue *cfqq = cfqd->active_queue;
2576
2577         if (cfqq)
2578                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2579 }
2580
2581 /*
2582  * Get next queue for service. Unless we have a queue preemption,
2583  * we'll simply select the first cfqq in the service tree.
2584  */
2585 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2586 {
2587         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2588                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2589
2590         if (!cfqd->rq_queued)
2591                 return NULL;
2592
2593         /* There is nothing to dispatch */
2594         if (!st)
2595                 return NULL;
2596         if (RB_EMPTY_ROOT(&st->rb))
2597                 return NULL;
2598         return cfq_rb_first(st);
2599 }
2600
2601 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2602 {
2603         struct cfq_group *cfqg;
2604         struct cfq_queue *cfqq;
2605         int i, j;
2606         struct cfq_rb_root *st;
2607
2608         if (!cfqd->rq_queued)
2609                 return NULL;
2610
2611         cfqg = cfq_get_next_cfqg(cfqd);
2612         if (!cfqg)
2613                 return NULL;
2614
2615         for_each_cfqg_st(cfqg, i, j, st)
2616                 if ((cfqq = cfq_rb_first(st)) != NULL)
2617                         return cfqq;
2618         return NULL;
2619 }
2620
2621 /*
2622  * Get and set a new active queue for service.
2623  */
2624 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2625                                               struct cfq_queue *cfqq)
2626 {
2627         if (!cfqq)
2628                 cfqq = cfq_get_next_queue(cfqd);
2629
2630         __cfq_set_active_queue(cfqd, cfqq);
2631         return cfqq;
2632 }
2633
2634 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2635                                           struct request *rq)
2636 {
2637         if (blk_rq_pos(rq) >= cfqd->last_position)
2638                 return blk_rq_pos(rq) - cfqd->last_position;
2639         else
2640                 return cfqd->last_position - blk_rq_pos(rq);
2641 }
2642
2643 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2644                                struct request *rq)
2645 {
2646         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2647 }
2648
2649 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2650                                     struct cfq_queue *cur_cfqq)
2651 {
2652         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2653         struct rb_node *parent, *node;
2654         struct cfq_queue *__cfqq;
2655         sector_t sector = cfqd->last_position;
2656
2657         if (RB_EMPTY_ROOT(root))
2658                 return NULL;
2659
2660         /*
2661          * First, if we find a request starting at the end of the last
2662          * request, choose it.
2663          */
2664         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2665         if (__cfqq)
2666                 return __cfqq;
2667
2668         /*
2669          * If the exact sector wasn't found, the parent of the NULL leaf
2670          * will contain the closest sector.
2671          */
2672         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2673         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2674                 return __cfqq;
2675
2676         if (blk_rq_pos(__cfqq->next_rq) < sector)
2677                 node = rb_next(&__cfqq->p_node);
2678         else
2679                 node = rb_prev(&__cfqq->p_node);
2680         if (!node)
2681                 return NULL;
2682
2683         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2684         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2685                 return __cfqq;
2686
2687         return NULL;
2688 }
2689
2690 /*
2691  * cfqd - obvious
2692  * cur_cfqq - passed in so that we don't decide that the current queue is
2693  *            closely cooperating with itself.
2694  *
2695  * So, basically we're assuming that that cur_cfqq has dispatched at least
2696  * one request, and that cfqd->last_position reflects a position on the disk
2697  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2698  * assumption.
2699  */
2700 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2701                                               struct cfq_queue *cur_cfqq)
2702 {
2703         struct cfq_queue *cfqq;
2704
2705         if (cfq_class_idle(cur_cfqq))
2706                 return NULL;
2707         if (!cfq_cfqq_sync(cur_cfqq))
2708                 return NULL;
2709         if (CFQQ_SEEKY(cur_cfqq))
2710                 return NULL;
2711
2712         /*
2713          * Don't search priority tree if it's the only queue in the group.
2714          */
2715         if (cur_cfqq->cfqg->nr_cfqq == 1)
2716                 return NULL;
2717
2718         /*
2719          * We should notice if some of the queues are cooperating, eg
2720          * working closely on the same area of the disk. In that case,
2721          * we can group them together and don't waste time idling.
2722          */
2723         cfqq = cfqq_close(cfqd, cur_cfqq);
2724         if (!cfqq)
2725                 return NULL;
2726
2727         /* If new queue belongs to different cfq_group, don't choose it */
2728         if (cur_cfqq->cfqg != cfqq->cfqg)
2729                 return NULL;
2730
2731         /*
2732          * It only makes sense to merge sync queues.
2733          */
2734         if (!cfq_cfqq_sync(cfqq))
2735                 return NULL;
2736         if (CFQQ_SEEKY(cfqq))
2737                 return NULL;
2738
2739         /*
2740          * Do not merge queues of different priority classes
2741          */
2742         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2743                 return NULL;
2744
2745         return cfqq;
2746 }
2747
2748 /*
2749  * Determine whether we should enforce idle window for this queue.
2750  */
2751
2752 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2753 {
2754         enum wl_class_t wl_class = cfqq_class(cfqq);
2755         struct cfq_rb_root *st = cfqq->service_tree;
2756
2757         BUG_ON(!st);
2758         BUG_ON(!st->count);
2759
2760         if (!cfqd->cfq_slice_idle)
2761                 return false;
2762
2763         /* We never do for idle class queues. */
2764         if (wl_class == IDLE_WORKLOAD)
2765                 return false;
2766
2767         /* We do for queues that were marked with idle window flag. */
2768         if (cfq_cfqq_idle_window(cfqq) &&
2769            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2770                 return true;
2771
2772         /*
2773          * Otherwise, we do only if they are the last ones
2774          * in their service tree.
2775          */
2776         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2777            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2778                 return true;
2779         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2780         return false;
2781 }
2782
2783 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2784 {
2785         struct cfq_queue *cfqq = cfqd->active_queue;
2786         struct cfq_io_cq *cic;
2787         unsigned long sl, group_idle = 0;
2788
2789         /*
2790          * SSD device without seek penalty, disable idling. But only do so
2791          * for devices that support queuing, otherwise we still have a problem
2792          * with sync vs async workloads.
2793          */
2794         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2795                 return;
2796
2797         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2798         WARN_ON(cfq_cfqq_slice_new(cfqq));
2799
2800         /*
2801          * idle is disabled, either manually or by past process history
2802          */
2803         if (!cfq_should_idle(cfqd, cfqq)) {
2804                 /* no queue idling. Check for group idling */
2805                 if (cfqd->cfq_group_idle)
2806                         group_idle = cfqd->cfq_group_idle;
2807                 else
2808                         return;
2809         }
2810
2811         /*
2812          * still active requests from this queue, don't idle
2813          */
2814         if (cfqq->dispatched)
2815                 return;
2816
2817         /*
2818          * task has exited, don't wait
2819          */
2820         cic = cfqd->active_cic;
2821         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2822                 return;
2823
2824         /*
2825          * If our average think time is larger than the remaining time
2826          * slice, then don't idle. This avoids overrunning the allotted
2827          * time slice.
2828          */
2829         if (sample_valid(cic->ttime.ttime_samples) &&
2830             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2831                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2832                              cic->ttime.ttime_mean);
2833                 return;
2834         }
2835
2836         /* There are other queues in the group, don't do group idle */
2837         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2838                 return;
2839
2840         cfq_mark_cfqq_wait_request(cfqq);
2841
2842         if (group_idle)
2843                 sl = cfqd->cfq_group_idle;
2844         else
2845                 sl = cfqd->cfq_slice_idle;
2846
2847         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2848         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2849         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2850                         group_idle ? 1 : 0);
2851 }
2852
2853 /*
2854  * Move request from internal lists to the request queue dispatch list.
2855  */
2856 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2857 {
2858         struct cfq_data *cfqd = q->elevator->elevator_data;
2859         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2860
2861         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2862
2863         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2864         cfq_remove_request(rq);
2865         cfqq->dispatched++;
2866         (RQ_CFQG(rq))->dispatched++;
2867         elv_dispatch_sort(q, rq);
2868
2869         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2870         cfqq->nr_sectors += blk_rq_sectors(rq);
2871         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2872 }
2873
2874 /*
2875  * return expired entry, or NULL to just start from scratch in rbtree
2876  */
2877 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2878 {
2879         struct request *rq = NULL;
2880
2881         if (cfq_cfqq_fifo_expire(cfqq))
2882                 return NULL;
2883
2884         cfq_mark_cfqq_fifo_expire(cfqq);
2885
2886         if (list_empty(&cfqq->fifo))
2887                 return NULL;
2888
2889         rq = rq_entry_fifo(cfqq->fifo.next);
2890         if (time_before(jiffies, rq->fifo_time))
2891                 rq = NULL;
2892
2893         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2894         return rq;
2895 }
2896
2897 static inline int
2898 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2899 {
2900         const int base_rq = cfqd->cfq_slice_async_rq;
2901
2902         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2903
2904         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2905 }
2906
2907 /*
2908  * Must be called with the queue_lock held.
2909  */
2910 static int cfqq_process_refs(struct cfq_queue *cfqq)
2911 {
2912         int process_refs, io_refs;
2913
2914         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2915         process_refs = cfqq->ref - io_refs;
2916         BUG_ON(process_refs < 0);
2917         return process_refs;
2918 }
2919
2920 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2921 {
2922         int process_refs, new_process_refs;
2923         struct cfq_queue *__cfqq;
2924
2925         /*
2926          * If there are no process references on the new_cfqq, then it is
2927          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2928          * chain may have dropped their last reference (not just their
2929          * last process reference).
2930          */
2931         if (!cfqq_process_refs(new_cfqq))
2932                 return;
2933
2934         /* Avoid a circular list and skip interim queue merges */
2935         while ((__cfqq = new_cfqq->new_cfqq)) {
2936                 if (__cfqq == cfqq)
2937                         return;
2938                 new_cfqq = __cfqq;
2939         }
2940
2941         process_refs = cfqq_process_refs(cfqq);
2942         new_process_refs = cfqq_process_refs(new_cfqq);
2943         /*
2944          * If the process for the cfqq has gone away, there is no
2945          * sense in merging the queues.
2946          */
2947         if (process_refs == 0 || new_process_refs == 0)
2948                 return;
2949
2950         /*
2951          * Merge in the direction of the lesser amount of work.
2952          */
2953         if (new_process_refs >= process_refs) {
2954                 cfqq->new_cfqq = new_cfqq;
2955                 new_cfqq->ref += process_refs;
2956         } else {
2957                 new_cfqq->new_cfqq = cfqq;
2958                 cfqq->ref += new_process_refs;
2959         }
2960 }
2961
2962 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2963                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2964 {
2965         struct cfq_queue *queue;
2966         int i;
2967         bool key_valid = false;
2968         unsigned long lowest_key = 0;
2969         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2970
2971         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2972                 /* select the one with lowest rb_key */
2973                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2974                 if (queue &&
2975                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2976                         lowest_key = queue->rb_key;
2977                         cur_best = i;
2978                         key_valid = true;
2979                 }
2980         }
2981
2982         return cur_best;
2983 }
2984
2985 static void
2986 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2987 {
2988         unsigned slice;
2989         unsigned count;
2990         struct cfq_rb_root *st;
2991         unsigned group_slice;
2992         enum wl_class_t original_class = cfqd->serving_wl_class;
2993
2994         /* Choose next priority. RT > BE > IDLE */
2995         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2996                 cfqd->serving_wl_class = RT_WORKLOAD;
2997         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2998                 cfqd->serving_wl_class = BE_WORKLOAD;
2999         else {
3000                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3001                 cfqd->workload_expires = jiffies + 1;
3002                 return;
3003         }
3004
3005         if (original_class != cfqd->serving_wl_class)
3006                 goto new_workload;
3007
3008         /*
3009          * For RT and BE, we have to choose also the type
3010          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3011          * expiration time
3012          */
3013         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3014         count = st->count;
3015
3016         /*
3017          * check workload expiration, and that we still have other queues ready
3018          */
3019         if (count && !time_after(jiffies, cfqd->workload_expires))
3020                 return;
3021
3022 new_workload:
3023         /* otherwise select new workload type */
3024         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3025                                         cfqd->serving_wl_class);
3026         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3027         count = st->count;
3028
3029         /*
3030          * the workload slice is computed as a fraction of target latency
3031          * proportional to the number of queues in that workload, over
3032          * all the queues in the same priority class
3033          */
3034         group_slice = cfq_group_slice(cfqd, cfqg);
3035
3036         slice = group_slice * count /
3037                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3038                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3039                                         cfqg));
3040
3041         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3042                 unsigned int tmp;
3043
3044                 /*
3045                  * Async queues are currently system wide. Just taking
3046                  * proportion of queues with-in same group will lead to higher
3047                  * async ratio system wide as generally root group is going
3048                  * to have higher weight. A more accurate thing would be to
3049                  * calculate system wide asnc/sync ratio.
3050                  */
3051                 tmp = cfqd->cfq_target_latency *
3052                         cfqg_busy_async_queues(cfqd, cfqg);
3053                 tmp = tmp/cfqd->busy_queues;
3054                 slice = min_t(unsigned, slice, tmp);
3055
3056                 /* async workload slice is scaled down according to
3057                  * the sync/async slice ratio. */
3058                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3059         } else
3060                 /* sync workload slice is at least 2 * cfq_slice_idle */
3061                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3062
3063         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3064         cfq_log(cfqd, "workload slice:%d", slice);
3065         cfqd->workload_expires = jiffies + slice;
3066 }
3067
3068 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3069 {
3070         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3071         struct cfq_group *cfqg;
3072
3073         if (RB_EMPTY_ROOT(&st->rb))
3074                 return NULL;
3075         cfqg = cfq_rb_first_group(st);
3076         update_min_vdisktime(st);
3077         return cfqg;
3078 }
3079
3080 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3081 {
3082         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3083
3084         cfqd->serving_group = cfqg;
3085
3086         /* Restore the workload type data */
3087         if (cfqg->saved_wl_slice) {
3088                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3089                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3090                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3091         } else
3092                 cfqd->workload_expires = jiffies - 1;
3093
3094         choose_wl_class_and_type(cfqd, cfqg);
3095 }
3096
3097 /*
3098  * Select a queue for service. If we have a current active queue,
3099  * check whether to continue servicing it, or retrieve and set a new one.
3100  */
3101 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3102 {
3103         struct cfq_queue *cfqq, *new_cfqq = NULL;
3104
3105         cfqq = cfqd->active_queue;
3106         if (!cfqq)
3107                 goto new_queue;
3108
3109         if (!cfqd->rq_queued)
3110                 return NULL;
3111
3112         /*
3113          * We were waiting for group to get backlogged. Expire the queue
3114          */
3115         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3116                 goto expire;
3117
3118         /*
3119          * The active queue has run out of time, expire it and select new.
3120          */
3121         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3122                 /*
3123                  * If slice had not expired at the completion of last request
3124                  * we might not have turned on wait_busy flag. Don't expire
3125                  * the queue yet. Allow the group to get backlogged.
3126                  *
3127                  * The very fact that we have used the slice, that means we
3128                  * have been idling all along on this queue and it should be
3129                  * ok to wait for this request to complete.
3130                  */
3131                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3132                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3133                         cfqq = NULL;
3134                         goto keep_queue;
3135                 } else
3136                         goto check_group_idle;
3137         }
3138
3139         /*
3140          * The active queue has requests and isn't expired, allow it to
3141          * dispatch.
3142          */
3143         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3144                 goto keep_queue;
3145
3146         /*
3147          * If another queue has a request waiting within our mean seek
3148          * distance, let it run.  The expire code will check for close
3149          * cooperators and put the close queue at the front of the service
3150          * tree.  If possible, merge the expiring queue with the new cfqq.
3151          */
3152         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3153         if (new_cfqq) {
3154                 if (!cfqq->new_cfqq)
3155                         cfq_setup_merge(cfqq, new_cfqq);
3156                 goto expire;
3157         }
3158
3159         /*
3160          * No requests pending. If the active queue still has requests in
3161          * flight or is idling for a new request, allow either of these
3162          * conditions to happen (or time out) before selecting a new queue.
3163          */
3164         if (timer_pending(&cfqd->idle_slice_timer)) {
3165                 cfqq = NULL;
3166                 goto keep_queue;
3167         }
3168
3169         /*
3170          * This is a deep seek queue, but the device is much faster than
3171          * the queue can deliver, don't idle
3172          **/
3173         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3174             (cfq_cfqq_slice_new(cfqq) ||
3175             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3176                 cfq_clear_cfqq_deep(cfqq);
3177                 cfq_clear_cfqq_idle_window(cfqq);
3178         }
3179
3180         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3181                 cfqq = NULL;
3182                 goto keep_queue;
3183         }
3184
3185         /*
3186          * If group idle is enabled and there are requests dispatched from
3187          * this group, wait for requests to complete.
3188          */
3189 check_group_idle:
3190         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3191             cfqq->cfqg->dispatched &&
3192             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3193                 cfqq = NULL;
3194                 goto keep_queue;
3195         }
3196
3197 expire:
3198         cfq_slice_expired(cfqd, 0);
3199 new_queue:
3200         /*
3201          * Current queue expired. Check if we have to switch to a new
3202          * service tree
3203          */
3204         if (!new_cfqq)
3205                 cfq_choose_cfqg(cfqd);
3206
3207         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3208 keep_queue:
3209         return cfqq;
3210 }
3211
3212 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3213 {
3214         int dispatched = 0;
3215
3216         while (cfqq->next_rq) {
3217                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3218                 dispatched++;
3219         }
3220
3221         BUG_ON(!list_empty(&cfqq->fifo));
3222
3223         /* By default cfqq is not expired if it is empty. Do it explicitly */
3224         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3225         return dispatched;
3226 }
3227
3228 /*
3229  * Drain our current requests. Used for barriers and when switching
3230  * io schedulers on-the-fly.
3231  */
3232 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3233 {
3234         struct cfq_queue *cfqq;
3235         int dispatched = 0;
3236
3237         /* Expire the timeslice of the current active queue first */
3238         cfq_slice_expired(cfqd, 0);
3239         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3240                 __cfq_set_active_queue(cfqd, cfqq);
3241                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3242         }
3243
3244         BUG_ON(cfqd->busy_queues);
3245
3246         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3247         return dispatched;
3248 }
3249
3250 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3251         struct cfq_queue *cfqq)
3252 {
3253         /* the queue hasn't finished any request, can't estimate */
3254         if (cfq_cfqq_slice_new(cfqq))
3255                 return true;
3256         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3257                 cfqq->slice_end))
3258                 return true;
3259
3260         return false;
3261 }
3262
3263 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3264 {
3265         unsigned int max_dispatch;
3266
3267         /*
3268          * Drain async requests before we start sync IO
3269          */
3270         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3271                 return false;
3272
3273         /*
3274          * If this is an async queue and we have sync IO in flight, let it wait
3275          */
3276         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3277                 return false;
3278
3279         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3280         if (cfq_class_idle(cfqq))
3281                 max_dispatch = 1;
3282
3283         /*
3284          * Does this cfqq already have too much IO in flight?
3285          */
3286         if (cfqq->dispatched >= max_dispatch) {
3287                 bool promote_sync = false;
3288                 /*
3289                  * idle queue must always only have a single IO in flight
3290                  */
3291                 if (cfq_class_idle(cfqq))
3292                         return false;
3293
3294                 /*
3295                  * If there is only one sync queue
3296                  * we can ignore async queue here and give the sync
3297                  * queue no dispatch limit. The reason is a sync queue can
3298                  * preempt async queue, limiting the sync queue doesn't make
3299                  * sense. This is useful for aiostress test.
3300                  */
3301                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3302                         promote_sync = true;
3303
3304                 /*
3305                  * We have other queues, don't allow more IO from this one
3306                  */
3307                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3308                                 !promote_sync)
3309                         return false;
3310
3311                 /*
3312                  * Sole queue user, no limit
3313                  */
3314                 if (cfqd->busy_queues == 1 || promote_sync)
3315                         max_dispatch = -1;
3316                 else
3317                         /*
3318                          * Normally we start throttling cfqq when cfq_quantum/2
3319                          * requests have been dispatched. But we can drive
3320                          * deeper queue depths at the beginning of slice
3321                          * subjected to upper limit of cfq_quantum.
3322                          * */
3323                         max_dispatch = cfqd->cfq_quantum;
3324         }
3325
3326         /*
3327          * Async queues must wait a bit before being allowed dispatch.
3328          * We also ramp up the dispatch depth gradually for async IO,
3329          * based on the last sync IO we serviced
3330          */
3331         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3332                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3333                 unsigned int depth;
3334
3335                 depth = last_sync / cfqd->cfq_slice[1];
3336                 if (!depth && !cfqq->dispatched)
3337                         depth = 1;
3338                 if (depth < max_dispatch)
3339                         max_dispatch = depth;
3340         }
3341
3342         /*
3343          * If we're below the current max, allow a dispatch
3344          */
3345         return cfqq->dispatched < max_dispatch;
3346 }
3347
3348 /*
3349  * Dispatch a request from cfqq, moving them to the request queue
3350  * dispatch list.
3351  */
3352 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3353 {
3354         struct request *rq;
3355
3356         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3357
3358         if (!cfq_may_dispatch(cfqd, cfqq))
3359                 return false;
3360
3361         /*
3362          * follow expired path, else get first next available
3363          */
3364         rq = cfq_check_fifo(cfqq);
3365         if (!rq)
3366                 rq = cfqq->next_rq;
3367
3368         /*
3369          * insert request into driver dispatch list
3370          */
3371         cfq_dispatch_insert(cfqd->queue, rq);
3372
3373         if (!cfqd->active_cic) {
3374                 struct cfq_io_cq *cic = RQ_CIC(rq);
3375
3376                 atomic_long_inc(&cic->icq.ioc->refcount);
3377                 cfqd->active_cic = cic;
3378         }
3379
3380         return true;
3381 }
3382
3383 /*
3384  * Find the cfqq that we need to service and move a request from that to the
3385  * dispatch list
3386  */
3387 static int cfq_dispatch_requests(struct request_queue *q, int force)
3388 {
3389         struct cfq_data *cfqd = q->elevator->elevator_data;
3390         struct cfq_queue *cfqq;
3391
3392         if (!cfqd->busy_queues)
3393                 return 0;
3394
3395         if (unlikely(force))
3396                 return cfq_forced_dispatch(cfqd);
3397
3398         cfqq = cfq_select_queue(cfqd);
3399         if (!cfqq)
3400                 return 0;
3401
3402         /*
3403          * Dispatch a request from this cfqq, if it is allowed
3404          */
3405         if (!cfq_dispatch_request(cfqd, cfqq))
3406                 return 0;
3407
3408         cfqq->slice_dispatch++;
3409         cfq_clear_cfqq_must_dispatch(cfqq);
3410
3411         /*
3412          * expire an async queue immediately if it has used up its slice. idle
3413          * queue always expire after 1 dispatch round.
3414          */
3415         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3416             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3417             cfq_class_idle(cfqq))) {
3418                 cfqq->slice_end = jiffies + 1;
3419                 cfq_slice_expired(cfqd, 0);
3420         }
3421
3422         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3423         return 1;
3424 }
3425
3426 /*
3427  * task holds one reference to the queue, dropped when task exits. each rq
3428  * in-flight on this queue also holds a reference, dropped when rq is freed.
3429  *
3430  * Each cfq queue took a reference on the parent group. Drop it now.
3431  * queue lock must be held here.
3432  */
3433 static void cfq_put_queue(struct cfq_queue *cfqq)
3434 {
3435         struct cfq_data *cfqd = cfqq->cfqd;
3436         struct cfq_group *cfqg;
3437
3438         BUG_ON(cfqq->ref <= 0);
3439
3440         cfqq->ref--;
3441         if (cfqq->ref)
3442                 return;
3443
3444         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3445         BUG_ON(rb_first(&cfqq->sort_list));
3446         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3447         cfqg = cfqq->cfqg;
3448
3449         if (unlikely(cfqd->active_queue == cfqq)) {
3450                 __cfq_slice_expired(cfqd, cfqq, 0);
3451                 cfq_schedule_dispatch(cfqd);
3452         }
3453
3454         BUG_ON(cfq_cfqq_on_rr(cfqq));
3455         kmem_cache_free(cfq_pool, cfqq);
3456         cfqg_put(cfqg);
3457 }
3458
3459 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3460 {
3461         struct cfq_queue *__cfqq, *next;
3462
3463         /*
3464          * If this queue was scheduled to merge with another queue, be
3465          * sure to drop the reference taken on that queue (and others in
3466          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3467          */
3468         __cfqq = cfqq->new_cfqq;
3469         while (__cfqq) {
3470                 if (__cfqq == cfqq) {
3471                         WARN(1, "cfqq->new_cfqq loop detected\n");
3472                         break;
3473                 }
3474                 next = __cfqq->new_cfqq;
3475                 cfq_put_queue(__cfqq);
3476                 __cfqq = next;
3477         }
3478 }
3479
3480 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3481 {
3482         if (unlikely(cfqq == cfqd->active_queue)) {
3483                 __cfq_slice_expired(cfqd, cfqq, 0);
3484                 cfq_schedule_dispatch(cfqd);
3485         }
3486
3487         cfq_put_cooperator(cfqq);
3488
3489         cfq_put_queue(cfqq);
3490 }
3491
3492 static void cfq_init_icq(struct io_cq *icq)
3493 {
3494         struct cfq_io_cq *cic = icq_to_cic(icq);
3495
3496         cic->ttime.last_end_request = jiffies;
3497 }
3498
3499 static void cfq_exit_icq(struct io_cq *icq)
3500 {
3501         struct cfq_io_cq *cic = icq_to_cic(icq);
3502         struct cfq_data *cfqd = cic_to_cfqd(cic);
3503
3504         if (cic_to_cfqq(cic, false)) {
3505                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3506                 cic_set_cfqq(cic, NULL, false);
3507         }
3508
3509         if (cic_to_cfqq(cic, true)) {
3510                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3511                 cic_set_cfqq(cic, NULL, true);
3512         }
3513 }
3514
3515 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3516 {
3517         struct task_struct *tsk = current;
3518         int ioprio_class;
3519
3520         if (!cfq_cfqq_prio_changed(cfqq))
3521                 return;
3522
3523         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3524         switch (ioprio_class) {
3525         default:
3526                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3527         case IOPRIO_CLASS_NONE:
3528                 /*
3529                  * no prio set, inherit CPU scheduling settings
3530                  */
3531                 cfqq->ioprio = task_nice_ioprio(tsk);
3532                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3533                 break;
3534         case IOPRIO_CLASS_RT:
3535                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3536                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3537                 break;
3538         case IOPRIO_CLASS_BE:
3539                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3540                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3541                 break;
3542         case IOPRIO_CLASS_IDLE:
3543                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3544                 cfqq->ioprio = 7;
3545                 cfq_clear_cfqq_idle_window(cfqq);
3546                 break;
3547         }
3548
3549         /*
3550          * keep track of original prio settings in case we have to temporarily
3551          * elevate the priority of this queue
3552          */
3553         cfqq->org_ioprio = cfqq->ioprio;
3554         cfq_clear_cfqq_prio_changed(cfqq);
3555 }
3556
3557 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3558 {
3559         int ioprio = cic->icq.ioc->ioprio;
3560         struct cfq_data *cfqd = cic_to_cfqd(cic);
3561         struct cfq_queue *cfqq;
3562
3563         /*
3564          * Check whether ioprio has changed.  The condition may trigger
3565          * spuriously on a newly created cic but there's no harm.
3566          */
3567         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3568                 return;
3569
3570         cfqq = cic_to_cfqq(cic, false);
3571         if (cfqq) {
3572                 cfq_put_queue(cfqq);
3573                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3574                 cic_set_cfqq(cic, cfqq, false);
3575         }
3576
3577         cfqq = cic_to_cfqq(cic, true);
3578         if (cfqq)
3579                 cfq_mark_cfqq_prio_changed(cfqq);
3580
3581         cic->ioprio = ioprio;
3582 }
3583
3584 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3585                           pid_t pid, bool is_sync)
3586 {
3587         RB_CLEAR_NODE(&cfqq->rb_node);
3588         RB_CLEAR_NODE(&cfqq->p_node);
3589         INIT_LIST_HEAD(&cfqq->fifo);
3590
3591         cfqq->ref = 0;
3592         cfqq->cfqd = cfqd;
3593
3594         cfq_mark_cfqq_prio_changed(cfqq);
3595
3596         if (is_sync) {
3597                 if (!cfq_class_idle(cfqq))
3598                         cfq_mark_cfqq_idle_window(cfqq);
3599                 cfq_mark_cfqq_sync(cfqq);
3600         }
3601         cfqq->pid = pid;
3602 }
3603
3604 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3605 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3606 {
3607         struct cfq_data *cfqd = cic_to_cfqd(cic);
3608         struct cfq_queue *cfqq;
3609         uint64_t serial_nr;
3610
3611         rcu_read_lock();
3612         serial_nr = bio_blkcg(bio)->css.serial_nr;
3613         rcu_read_unlock();
3614
3615         /*
3616          * Check whether blkcg has changed.  The condition may trigger
3617          * spuriously on a newly created cic but there's no harm.
3618          */
3619         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3620                 return;
3621
3622         /*
3623          * Drop reference to queues.  New queues will be assigned in new
3624          * group upon arrival of fresh requests.
3625          */
3626         cfqq = cic_to_cfqq(cic, false);
3627         if (cfqq) {
3628                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3629                 cic_set_cfqq(cic, NULL, false);
3630                 cfq_put_queue(cfqq);
3631         }
3632
3633         cfqq = cic_to_cfqq(cic, true);
3634         if (cfqq) {
3635                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3636                 cic_set_cfqq(cic, NULL, true);
3637                 cfq_put_queue(cfqq);
3638         }
3639
3640         cic->blkcg_serial_nr = serial_nr;
3641 }
3642 #else
3643 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3644 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3645
3646 static struct cfq_queue **
3647 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3648 {
3649         switch (ioprio_class) {
3650         case IOPRIO_CLASS_RT:
3651                 return &cfqg->async_cfqq[0][ioprio];
3652         case IOPRIO_CLASS_NONE:
3653                 ioprio = IOPRIO_NORM;
3654                 /* fall through */
3655         case IOPRIO_CLASS_BE:
3656                 return &cfqg->async_cfqq[1][ioprio];
3657         case IOPRIO_CLASS_IDLE:
3658                 return &cfqg->async_idle_cfqq;
3659         default:
3660                 BUG();
3661         }
3662 }
3663
3664 static struct cfq_queue *
3665 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3666               struct bio *bio)
3667 {
3668         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3669         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3670         struct cfq_queue **async_cfqq = NULL;
3671         struct cfq_queue *cfqq;
3672         struct cfq_group *cfqg;
3673
3674         rcu_read_lock();
3675         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3676         if (!cfqg) {
3677                 cfqq = &cfqd->oom_cfqq;
3678                 goto out;
3679         }
3680
3681         if (!is_sync) {
3682                 if (!ioprio_valid(cic->ioprio)) {
3683                         struct task_struct *tsk = current;
3684                         ioprio = task_nice_ioprio(tsk);
3685                         ioprio_class = task_nice_ioclass(tsk);
3686                 }
3687                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3688                 cfqq = *async_cfqq;
3689                 if (cfqq)
3690                         goto out;
3691         }
3692
3693         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3694                                      cfqd->queue->node);
3695         if (!cfqq) {
3696                 cfqq = &cfqd->oom_cfqq;
3697                 goto out;
3698         }
3699
3700         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3701         cfq_init_prio_data(cfqq, cic);
3702         cfq_link_cfqq_cfqg(cfqq, cfqg);
3703         cfq_log_cfqq(cfqd, cfqq, "alloced");
3704
3705         if (async_cfqq) {
3706                 /* a new async queue is created, pin and remember */
3707                 cfqq->ref++;
3708                 *async_cfqq = cfqq;
3709         }
3710 out:
3711         cfqq->ref++;
3712         rcu_read_unlock();
3713         return cfqq;
3714 }
3715
3716 static void
3717 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3718 {
3719         unsigned long elapsed = jiffies - ttime->last_end_request;
3720         elapsed = min(elapsed, 2UL * slice_idle);
3721
3722         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3723         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3724         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3725 }
3726
3727 static void
3728 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3729                         struct cfq_io_cq *cic)
3730 {
3731         if (cfq_cfqq_sync(cfqq)) {
3732                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3733                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3734                         cfqd->cfq_slice_idle);
3735         }
3736 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3737         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3738 #endif
3739 }
3740
3741 static void
3742 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3743                        struct request *rq)
3744 {
3745         sector_t sdist = 0;
3746         sector_t n_sec = blk_rq_sectors(rq);
3747         if (cfqq->last_request_pos) {
3748                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3749                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3750                 else
3751                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3752         }
3753
3754         cfqq->seek_history <<= 1;
3755         if (blk_queue_nonrot(cfqd->queue))
3756                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3757         else
3758                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3759 }
3760
3761 /*
3762  * Disable idle window if the process thinks too long or seeks so much that
3763  * it doesn't matter
3764  */
3765 static void
3766 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3767                        struct cfq_io_cq *cic)
3768 {
3769         int old_idle, enable_idle;
3770
3771         /*
3772          * Don't idle for async or idle io prio class
3773          */
3774         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3775                 return;
3776
3777         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3778
3779         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3780                 cfq_mark_cfqq_deep(cfqq);
3781
3782         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3783                 enable_idle = 0;
3784         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3785                  !cfqd->cfq_slice_idle ||
3786                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3787                 enable_idle = 0;
3788         else if (sample_valid(cic->ttime.ttime_samples)) {
3789                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3790                         enable_idle = 0;
3791                 else
3792                         enable_idle = 1;
3793         }
3794
3795         if (old_idle != enable_idle) {
3796                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3797                 if (enable_idle)
3798                         cfq_mark_cfqq_idle_window(cfqq);
3799                 else
3800                         cfq_clear_cfqq_idle_window(cfqq);
3801         }
3802 }
3803
3804 /*
3805  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3806  * no or if we aren't sure, a 1 will cause a preempt.
3807  */
3808 static bool
3809 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3810                    struct request *rq)
3811 {
3812         struct cfq_queue *cfqq;
3813
3814         cfqq = cfqd->active_queue;
3815         if (!cfqq)
3816                 return false;
3817
3818         if (cfq_class_idle(new_cfqq))
3819                 return false;
3820
3821         if (cfq_class_idle(cfqq))
3822                 return true;
3823
3824         /*
3825          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3826          */
3827         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3828                 return false;
3829
3830         /*
3831          * if the new request is sync, but the currently running queue is
3832          * not, let the sync request have priority.
3833          */
3834         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3835                 return true;
3836
3837         if (new_cfqq->cfqg != cfqq->cfqg)
3838                 return false;
3839
3840         if (cfq_slice_used(cfqq))
3841                 return true;
3842
3843         /* Allow preemption only if we are idling on sync-noidle tree */
3844         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3845             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3846             new_cfqq->service_tree->count == 2 &&
3847             RB_EMPTY_ROOT(&cfqq->sort_list))
3848                 return true;
3849
3850         /*
3851          * So both queues are sync. Let the new request get disk time if
3852          * it's a metadata request and the current queue is doing regular IO.
3853          */
3854         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3855                 return true;
3856
3857         /*
3858          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3859          */
3860         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3861                 return true;
3862
3863         /* An idle queue should not be idle now for some reason */
3864         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3865                 return true;
3866
3867         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3868                 return false;
3869
3870         /*
3871          * if this request is as-good as one we would expect from the
3872          * current cfqq, let it preempt
3873          */
3874         if (cfq_rq_close(cfqd, cfqq, rq))
3875                 return true;
3876
3877         return false;
3878 }
3879
3880 /*
3881  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3882  * let it have half of its nominal slice.
3883  */
3884 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3885 {
3886         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3887
3888         cfq_log_cfqq(cfqd, cfqq, "preempt");
3889         cfq_slice_expired(cfqd, 1);
3890
3891         /*
3892          * workload type is changed, don't save slice, otherwise preempt
3893          * doesn't happen
3894          */
3895         if (old_type != cfqq_type(cfqq))
3896                 cfqq->cfqg->saved_wl_slice = 0;
3897
3898         /*
3899          * Put the new queue at the front of the of the current list,
3900          * so we know that it will be selected next.
3901          */
3902         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3903
3904         cfq_service_tree_add(cfqd, cfqq, 1);
3905
3906         cfqq->slice_end = 0;
3907         cfq_mark_cfqq_slice_new(cfqq);
3908 }
3909
3910 /*
3911  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3912  * something we should do about it
3913  */
3914 static void
3915 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3916                 struct request *rq)
3917 {
3918         struct cfq_io_cq *cic = RQ_CIC(rq);
3919
3920         cfqd->rq_queued++;
3921         if (rq->cmd_flags & REQ_PRIO)
3922                 cfqq->prio_pending++;
3923
3924         cfq_update_io_thinktime(cfqd, cfqq, cic);
3925         cfq_update_io_seektime(cfqd, cfqq, rq);
3926         cfq_update_idle_window(cfqd, cfqq, cic);
3927
3928         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3929
3930         if (cfqq == cfqd->active_queue) {
3931                 /*
3932                  * Remember that we saw a request from this process, but
3933                  * don't start queuing just yet. Otherwise we risk seeing lots
3934                  * of tiny requests, because we disrupt the normal plugging
3935                  * and merging. If the request is already larger than a single
3936                  * page, let it rip immediately. For that case we assume that
3937                  * merging is already done. Ditto for a busy system that
3938                  * has other work pending, don't risk delaying until the
3939                  * idle timer unplug to continue working.
3940                  */
3941                 if (cfq_cfqq_wait_request(cfqq)) {
3942                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3943                             cfqd->busy_queues > 1) {
3944                                 cfq_del_timer(cfqd, cfqq);
3945                                 cfq_clear_cfqq_wait_request(cfqq);
3946                                 __blk_run_queue(cfqd->queue);
3947                         } else {
3948                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3949                                 cfq_mark_cfqq_must_dispatch(cfqq);
3950                         }
3951                 }
3952         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3953                 /*
3954                  * not the active queue - expire current slice if it is
3955                  * idle and has expired it's mean thinktime or this new queue
3956                  * has some old slice time left and is of higher priority or
3957                  * this new queue is RT and the current one is BE
3958                  */
3959                 cfq_preempt_queue(cfqd, cfqq);
3960                 __blk_run_queue(cfqd->queue);
3961         }
3962 }
3963
3964 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3965 {
3966         struct cfq_data *cfqd = q->elevator->elevator_data;
3967         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3968
3969         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3970         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3971
3972         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
3973         list_add_tail(&rq->queuelist, &cfqq->fifo);
3974         cfq_add_rq_rb(rq);
3975         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3976                                  rq->cmd_flags);
3977         cfq_rq_enqueued(cfqd, cfqq, rq);
3978 }
3979
3980 /*
3981  * Update hw_tag based on peak queue depth over 50 samples under
3982  * sufficient load.
3983  */
3984 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3985 {
3986         struct cfq_queue *cfqq = cfqd->active_queue;
3987
3988         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3989                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3990
3991         if (cfqd->hw_tag == 1)
3992                 return;
3993
3994         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3995             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3996                 return;
3997
3998         /*
3999          * If active queue hasn't enough requests and can idle, cfq might not
4000          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4001          * case
4002          */
4003         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4004             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4005             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4006                 return;
4007
4008         if (cfqd->hw_tag_samples++ < 50)
4009                 return;
4010
4011         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4012                 cfqd->hw_tag = 1;
4013         else
4014                 cfqd->hw_tag = 0;
4015 }
4016
4017 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4018 {
4019         struct cfq_io_cq *cic = cfqd->active_cic;
4020
4021         /* If the queue already has requests, don't wait */
4022         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4023                 return false;
4024
4025         /* If there are other queues in the group, don't wait */
4026         if (cfqq->cfqg->nr_cfqq > 1)
4027                 return false;
4028
4029         /* the only queue in the group, but think time is big */
4030         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4031                 return false;
4032
4033         if (cfq_slice_used(cfqq))
4034                 return true;
4035
4036         /* if slice left is less than think time, wait busy */
4037         if (cic && sample_valid(cic->ttime.ttime_samples)
4038             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4039                 return true;
4040
4041         /*
4042          * If think times is less than a jiffy than ttime_mean=0 and above
4043          * will not be true. It might happen that slice has not expired yet
4044          * but will expire soon (4-5 ns) during select_queue(). To cover the
4045          * case where think time is less than a jiffy, mark the queue wait
4046          * busy if only 1 jiffy is left in the slice.
4047          */
4048         if (cfqq->slice_end - jiffies == 1)
4049                 return true;
4050
4051         return false;
4052 }
4053
4054 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4055 {
4056         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4057         struct cfq_data *cfqd = cfqq->cfqd;
4058         const int sync = rq_is_sync(rq);
4059         unsigned long now;
4060
4061         now = jiffies;
4062         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4063                      !!(rq->cmd_flags & REQ_NOIDLE));
4064
4065         cfq_update_hw_tag(cfqd);
4066
4067         WARN_ON(!cfqd->rq_in_driver);
4068         WARN_ON(!cfqq->dispatched);
4069         cfqd->rq_in_driver--;
4070         cfqq->dispatched--;
4071         (RQ_CFQG(rq))->dispatched--;
4072         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4073                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4074
4075         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4076
4077         if (sync) {
4078                 struct cfq_rb_root *st;
4079
4080                 RQ_CIC(rq)->ttime.last_end_request = now;
4081
4082                 if (cfq_cfqq_on_rr(cfqq))
4083                         st = cfqq->service_tree;
4084                 else
4085                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4086                                         cfqq_type(cfqq));
4087
4088                 st->ttime.last_end_request = now;
4089                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4090                         cfqd->last_delayed_sync = now;
4091         }
4092
4093 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4094         cfqq->cfqg->ttime.last_end_request = now;
4095 #endif
4096
4097         /*
4098          * If this is the active queue, check if it needs to be expired,
4099          * or if we want to idle in case it has no pending requests.
4100          */
4101         if (cfqd->active_queue == cfqq) {
4102                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4103
4104                 if (cfq_cfqq_slice_new(cfqq)) {
4105                         cfq_set_prio_slice(cfqd, cfqq);
4106                         cfq_clear_cfqq_slice_new(cfqq);
4107                 }
4108
4109                 /*
4110                  * Should we wait for next request to come in before we expire
4111                  * the queue.
4112                  */
4113                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4114                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4115                         if (!cfqd->cfq_slice_idle)
4116                                 extend_sl = cfqd->cfq_group_idle;
4117                         cfqq->slice_end = jiffies + extend_sl;
4118                         cfq_mark_cfqq_wait_busy(cfqq);
4119                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4120                 }
4121
4122                 /*
4123                  * Idling is not enabled on:
4124                  * - expired queues
4125                  * - idle-priority queues
4126                  * - async queues
4127                  * - queues with still some requests queued
4128                  * - when there is a close cooperator
4129                  */
4130                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4131                         cfq_slice_expired(cfqd, 1);
4132                 else if (sync && cfqq_empty &&
4133                          !cfq_close_cooperator(cfqd, cfqq)) {
4134                         cfq_arm_slice_timer(cfqd);
4135                 }
4136         }
4137
4138         if (!cfqd->rq_in_driver)
4139                 cfq_schedule_dispatch(cfqd);
4140 }
4141
4142 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4143 {
4144         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4145                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4146                 return ELV_MQUEUE_MUST;
4147         }
4148
4149         return ELV_MQUEUE_MAY;
4150 }
4151
4152 static int cfq_may_queue(struct request_queue *q, int rw)
4153 {
4154         struct cfq_data *cfqd = q->elevator->elevator_data;
4155         struct task_struct *tsk = current;
4156         struct cfq_io_cq *cic;
4157         struct cfq_queue *cfqq;
4158
4159         /*
4160          * don't force setup of a queue from here, as a call to may_queue
4161          * does not necessarily imply that a request actually will be queued.
4162          * so just lookup a possibly existing queue, or return 'may queue'
4163          * if that fails
4164          */
4165         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4166         if (!cic)
4167                 return ELV_MQUEUE_MAY;
4168
4169         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4170         if (cfqq) {
4171                 cfq_init_prio_data(cfqq, cic);
4172
4173                 return __cfq_may_queue(cfqq);
4174         }
4175
4176         return ELV_MQUEUE_MAY;
4177 }
4178
4179 /*
4180  * queue lock held here
4181  */
4182 static void cfq_put_request(struct request *rq)
4183 {
4184         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4185
4186         if (cfqq) {
4187                 const int rw = rq_data_dir(rq);
4188
4189                 BUG_ON(!cfqq->allocated[rw]);
4190                 cfqq->allocated[rw]--;
4191
4192                 /* Put down rq reference on cfqg */
4193                 cfqg_put(RQ_CFQG(rq));
4194                 rq->elv.priv[0] = NULL;
4195                 rq->elv.priv[1] = NULL;
4196
4197                 cfq_put_queue(cfqq);
4198         }
4199 }
4200
4201 static struct cfq_queue *
4202 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4203                 struct cfq_queue *cfqq)
4204 {
4205         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4206         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4207         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4208         cfq_put_queue(cfqq);
4209         return cic_to_cfqq(cic, 1);
4210 }
4211
4212 /*
4213  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4214  * was the last process referring to said cfqq.
4215  */
4216 static struct cfq_queue *
4217 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4218 {
4219         if (cfqq_process_refs(cfqq) == 1) {
4220                 cfqq->pid = current->pid;
4221                 cfq_clear_cfqq_coop(cfqq);
4222                 cfq_clear_cfqq_split_coop(cfqq);
4223                 return cfqq;
4224         }
4225
4226         cic_set_cfqq(cic, NULL, 1);
4227
4228         cfq_put_cooperator(cfqq);
4229
4230         cfq_put_queue(cfqq);
4231         return NULL;
4232 }
4233 /*
4234  * Allocate cfq data structures associated with this request.
4235  */
4236 static int
4237 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4238                 gfp_t gfp_mask)
4239 {
4240         struct cfq_data *cfqd = q->elevator->elevator_data;
4241         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4242         const int rw = rq_data_dir(rq);
4243         const bool is_sync = rq_is_sync(rq);
4244         struct cfq_queue *cfqq;
4245
4246         spin_lock_irq(q->queue_lock);
4247
4248         check_ioprio_changed(cic, bio);
4249         check_blkcg_changed(cic, bio);
4250 new_queue:
4251         cfqq = cic_to_cfqq(cic, is_sync);
4252         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4253                 if (cfqq)
4254                         cfq_put_queue(cfqq);
4255                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4256                 cic_set_cfqq(cic, cfqq, is_sync);
4257         } else {
4258                 /*
4259                  * If the queue was seeky for too long, break it apart.
4260                  */
4261                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4262                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4263                         cfqq = split_cfqq(cic, cfqq);
4264                         if (!cfqq)
4265                                 goto new_queue;
4266                 }
4267
4268                 /*
4269                  * Check to see if this queue is scheduled to merge with
4270                  * another, closely cooperating queue.  The merging of
4271                  * queues happens here as it must be done in process context.
4272                  * The reference on new_cfqq was taken in merge_cfqqs.
4273                  */
4274                 if (cfqq->new_cfqq)
4275                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4276         }
4277
4278         cfqq->allocated[rw]++;
4279
4280         cfqq->ref++;
4281         cfqg_get(cfqq->cfqg);
4282         rq->elv.priv[0] = cfqq;
4283         rq->elv.priv[1] = cfqq->cfqg;
4284         spin_unlock_irq(q->queue_lock);
4285         return 0;
4286 }
4287
4288 static void cfq_kick_queue(struct work_struct *work)
4289 {
4290         struct cfq_data *cfqd =
4291                 container_of(work, struct cfq_data, unplug_work);
4292         struct request_queue *q = cfqd->queue;
4293
4294         spin_lock_irq(q->queue_lock);
4295         __blk_run_queue(cfqd->queue);
4296         spin_unlock_irq(q->queue_lock);
4297 }
4298
4299 /*
4300  * Timer running if the active_queue is currently idling inside its time slice
4301  */
4302 static void cfq_idle_slice_timer(unsigned long data)
4303 {
4304         struct cfq_data *cfqd = (struct cfq_data *) data;
4305         struct cfq_queue *cfqq;
4306         unsigned long flags;
4307         int timed_out = 1;
4308
4309         cfq_log(cfqd, "idle timer fired");
4310
4311         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4312
4313         cfqq = cfqd->active_queue;
4314         if (cfqq) {
4315                 timed_out = 0;
4316
4317                 /*
4318                  * We saw a request before the queue expired, let it through
4319                  */
4320                 if (cfq_cfqq_must_dispatch(cfqq))
4321                         goto out_kick;
4322
4323                 /*
4324                  * expired
4325                  */
4326                 if (cfq_slice_used(cfqq))
4327                         goto expire;
4328
4329                 /*
4330                  * only expire and reinvoke request handler, if there are
4331                  * other queues with pending requests
4332                  */
4333                 if (!cfqd->busy_queues)
4334                         goto out_cont;
4335
4336                 /*
4337                  * not expired and it has a request pending, let it dispatch
4338                  */
4339                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4340                         goto out_kick;
4341
4342                 /*
4343                  * Queue depth flag is reset only when the idle didn't succeed
4344                  */
4345                 cfq_clear_cfqq_deep(cfqq);
4346         }
4347 expire:
4348         cfq_slice_expired(cfqd, timed_out);
4349 out_kick:
4350         cfq_schedule_dispatch(cfqd);
4351 out_cont:
4352         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4353 }
4354
4355 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4356 {
4357         del_timer_sync(&cfqd->idle_slice_timer);
4358         cancel_work_sync(&cfqd->unplug_work);
4359 }
4360
4361 static void cfq_exit_queue(struct elevator_queue *e)
4362 {
4363         struct cfq_data *cfqd = e->elevator_data;
4364         struct request_queue *q = cfqd->queue;
4365
4366         cfq_shutdown_timer_wq(cfqd);
4367
4368         spin_lock_irq(q->queue_lock);
4369
4370         if (cfqd->active_queue)
4371                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4372
4373         spin_unlock_irq(q->queue_lock);
4374
4375         cfq_shutdown_timer_wq(cfqd);
4376
4377 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4378         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4379 #else
4380         kfree(cfqd->root_group);
4381 #endif
4382         kfree(cfqd);
4383 }
4384
4385 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4386 {
4387         struct cfq_data *cfqd;
4388         struct blkcg_gq *blkg __maybe_unused;
4389         int i, ret;
4390         struct elevator_queue *eq;
4391
4392         eq = elevator_alloc(q, e);
4393         if (!eq)
4394                 return -ENOMEM;
4395
4396         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4397         if (!cfqd) {
4398                 kobject_put(&eq->kobj);
4399                 return -ENOMEM;
4400         }
4401         eq->elevator_data = cfqd;
4402
4403         cfqd->queue = q;
4404         spin_lock_irq(q->queue_lock);
4405         q->elevator = eq;
4406         spin_unlock_irq(q->queue_lock);
4407
4408         /* Init root service tree */
4409         cfqd->grp_service_tree = CFQ_RB_ROOT;
4410
4411         /* Init root group and prefer root group over other groups by default */
4412 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4413         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4414         if (ret)
4415                 goto out_free;
4416
4417         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4418 #else
4419         ret = -ENOMEM;
4420         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4421                                         GFP_KERNEL, cfqd->queue->node);
4422         if (!cfqd->root_group)
4423                 goto out_free;
4424
4425         cfq_init_cfqg_base(cfqd->root_group);
4426 #endif
4427         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4428         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4429
4430         /*
4431          * Not strictly needed (since RB_ROOT just clears the node and we
4432          * zeroed cfqd on alloc), but better be safe in case someone decides
4433          * to add magic to the rb code
4434          */
4435         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4436                 cfqd->prio_trees[i] = RB_ROOT;
4437
4438         /*
4439          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4440          * Grab a permanent reference to it, so that the normal code flow
4441          * will not attempt to free it.  oom_cfqq is linked to root_group
4442          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4443          * the reference from linking right away.
4444          */
4445         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4446         cfqd->oom_cfqq.ref++;
4447
4448         spin_lock_irq(q->queue_lock);
4449         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4450         cfqg_put(cfqd->root_group);
4451         spin_unlock_irq(q->queue_lock);
4452
4453         init_timer(&cfqd->idle_slice_timer);
4454         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4455         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4456
4457         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4458
4459         cfqd->cfq_quantum = cfq_quantum;
4460         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4461         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4462         cfqd->cfq_back_max = cfq_back_max;
4463         cfqd->cfq_back_penalty = cfq_back_penalty;
4464         cfqd->cfq_slice[0] = cfq_slice_async;
4465         cfqd->cfq_slice[1] = cfq_slice_sync;
4466         cfqd->cfq_target_latency = cfq_target_latency;
4467         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4468         cfqd->cfq_slice_idle = cfq_slice_idle;
4469         cfqd->cfq_group_idle = cfq_group_idle;
4470         cfqd->cfq_latency = 1;
4471         cfqd->hw_tag = -1;
4472         /*
4473          * we optimistically start assuming sync ops weren't delayed in last
4474          * second, in order to have larger depth for async operations.
4475          */
4476         cfqd->last_delayed_sync = jiffies - HZ;
4477         return 0;
4478
4479 out_free:
4480         kfree(cfqd);
4481         kobject_put(&eq->kobj);
4482         return ret;
4483 }
4484
4485 static void cfq_registered_queue(struct request_queue *q)
4486 {
4487         struct elevator_queue *e = q->elevator;
4488         struct cfq_data *cfqd = e->elevator_data;
4489
4490         /*
4491          * Default to IOPS mode with no idling for SSDs
4492          */
4493         if (blk_queue_nonrot(q))
4494                 cfqd->cfq_slice_idle = 0;
4495 }
4496
4497 /*
4498  * sysfs parts below -->
4499  */
4500 static ssize_t
4501 cfq_var_show(unsigned int var, char *page)
4502 {
4503         return sprintf(page, "%u\n", var);
4504 }
4505
4506 static ssize_t
4507 cfq_var_store(unsigned int *var, const char *page, size_t count)
4508 {
4509         char *p = (char *) page;
4510
4511         *var = simple_strtoul(p, &p, 10);
4512         return count;
4513 }
4514
4515 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4516 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4517 {                                                                       \
4518         struct cfq_data *cfqd = e->elevator_data;                       \
4519         unsigned int __data = __VAR;                                    \
4520         if (__CONV)                                                     \
4521                 __data = jiffies_to_msecs(__data);                      \
4522         return cfq_var_show(__data, (page));                            \
4523 }
4524 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4525 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4526 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4527 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4528 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4529 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4530 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4531 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4532 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4533 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4534 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4535 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4536 #undef SHOW_FUNCTION
4537
4538 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4539 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4540 {                                                                       \
4541         struct cfq_data *cfqd = e->elevator_data;                       \
4542         unsigned int __data;                                            \
4543         int ret = cfq_var_store(&__data, (page), count);                \
4544         if (__data < (MIN))                                             \
4545                 __data = (MIN);                                         \
4546         else if (__data > (MAX))                                        \
4547                 __data = (MAX);                                         \
4548         if (__CONV)                                                     \
4549                 *(__PTR) = msecs_to_jiffies(__data);                    \
4550         else                                                            \
4551                 *(__PTR) = __data;                                      \
4552         return ret;                                                     \
4553 }
4554 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4555 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4556                 UINT_MAX, 1);
4557 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4558                 UINT_MAX, 1);
4559 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4560 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4561                 UINT_MAX, 0);
4562 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4563 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4564 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4565 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4566 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4567                 UINT_MAX, 0);
4568 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4569 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4570 #undef STORE_FUNCTION
4571
4572 #define CFQ_ATTR(name) \
4573         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4574
4575 static struct elv_fs_entry cfq_attrs[] = {
4576         CFQ_ATTR(quantum),
4577         CFQ_ATTR(fifo_expire_sync),
4578         CFQ_ATTR(fifo_expire_async),
4579         CFQ_ATTR(back_seek_max),
4580         CFQ_ATTR(back_seek_penalty),
4581         CFQ_ATTR(slice_sync),
4582         CFQ_ATTR(slice_async),
4583         CFQ_ATTR(slice_async_rq),
4584         CFQ_ATTR(slice_idle),
4585         CFQ_ATTR(group_idle),
4586         CFQ_ATTR(low_latency),
4587         CFQ_ATTR(target_latency),
4588         __ATTR_NULL
4589 };
4590
4591 static struct elevator_type iosched_cfq = {
4592         .ops = {
4593                 .elevator_merge_fn =            cfq_merge,
4594                 .elevator_merged_fn =           cfq_merged_request,
4595                 .elevator_merge_req_fn =        cfq_merged_requests,
4596                 .elevator_allow_merge_fn =      cfq_allow_merge,
4597                 .elevator_bio_merged_fn =       cfq_bio_merged,
4598                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4599                 .elevator_add_req_fn =          cfq_insert_request,
4600                 .elevator_activate_req_fn =     cfq_activate_request,
4601                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4602                 .elevator_completed_req_fn =    cfq_completed_request,
4603                 .elevator_former_req_fn =       elv_rb_former_request,
4604                 .elevator_latter_req_fn =       elv_rb_latter_request,
4605                 .elevator_init_icq_fn =         cfq_init_icq,
4606                 .elevator_exit_icq_fn =         cfq_exit_icq,
4607                 .elevator_set_req_fn =          cfq_set_request,
4608                 .elevator_put_req_fn =          cfq_put_request,
4609                 .elevator_may_queue_fn =        cfq_may_queue,
4610                 .elevator_init_fn =             cfq_init_queue,
4611                 .elevator_exit_fn =             cfq_exit_queue,
4612                 .elevator_registered_fn =       cfq_registered_queue,
4613         },
4614         .icq_size       =       sizeof(struct cfq_io_cq),
4615         .icq_align      =       __alignof__(struct cfq_io_cq),
4616         .elevator_attrs =       cfq_attrs,
4617         .elevator_name  =       "cfq",
4618         .elevator_owner =       THIS_MODULE,
4619 };
4620
4621 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4622 static struct blkcg_policy blkcg_policy_cfq = {
4623         .cftypes                = cfq_blkcg_files,
4624
4625         .cpd_alloc_fn           = cfq_cpd_alloc,
4626         .cpd_init_fn            = cfq_cpd_init,
4627         .cpd_free_fn            = cfq_cpd_free,
4628
4629         .pd_alloc_fn            = cfq_pd_alloc,
4630         .pd_init_fn             = cfq_pd_init,
4631         .pd_offline_fn          = cfq_pd_offline,
4632         .pd_free_fn             = cfq_pd_free,
4633         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4634 };
4635 #endif
4636
4637 static int __init cfq_init(void)
4638 {
4639         int ret;
4640
4641         /*
4642          * could be 0 on HZ < 1000 setups
4643          */
4644         if (!cfq_slice_async)
4645                 cfq_slice_async = 1;
4646         if (!cfq_slice_idle)
4647                 cfq_slice_idle = 1;
4648
4649 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4650         if (!cfq_group_idle)
4651                 cfq_group_idle = 1;
4652
4653         ret = blkcg_policy_register(&blkcg_policy_cfq);
4654         if (ret)
4655                 return ret;
4656 #else
4657         cfq_group_idle = 0;
4658 #endif
4659
4660         ret = -ENOMEM;
4661         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4662         if (!cfq_pool)
4663                 goto err_pol_unreg;
4664
4665         ret = elv_register(&iosched_cfq);
4666         if (ret)
4667                 goto err_free_pool;
4668
4669         return 0;
4670
4671 err_free_pool:
4672         kmem_cache_destroy(cfq_pool);
4673 err_pol_unreg:
4674 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4675         blkcg_policy_unregister(&blkcg_policy_cfq);
4676 #endif
4677         return ret;
4678 }
4679
4680 static void __exit cfq_exit(void)
4681 {
4682 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4683         blkcg_policy_unregister(&blkcg_policy_cfq);
4684 #endif
4685         elv_unregister(&iosched_cfq);
4686         kmem_cache_destroy(cfq_pool);
4687 }
4688
4689 module_init(cfq_init);
4690 module_exit(cfq_exit);
4691
4692 MODULE_AUTHOR("Jens Axboe");
4693 MODULE_LICENSE("GPL");
4694 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");