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/sched/clock.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/elevator.h>
  14 #include <linux/ktime.h>
  15 #include <linux/rbtree.h>
  16 #include <linux/ioprio.h>
  17 #include <linux/blktrace_api.h>
  18 #include <linux/blk-cgroup.h>
  19 #include "blk.h"
  20 #include "blk-wbt.h"
  21
  22 /*
  23  * tunables
  24  */
  25 /* max queue in one round of service */
  26 static const int cfq_quantum = 8;
  27 static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
  28 /* maximum backwards seek, in KiB */
  29 static const int cfq_back_max = 16 * 1024;
  30 /* penalty of a backwards seek */
  31 static const int cfq_back_penalty = 2;
  32 static const u64 cfq_slice_sync = NSEC_PER_SEC / 10;
  33 static u64 cfq_slice_async = NSEC_PER_SEC / 25;
  34 static const int cfq_slice_async_rq = 2;
  35 static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
  36 static u64 cfq_group_idle = NSEC_PER_SEC / 125;
  37 static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
  38 static const int cfq_hist_divisor = 4;
  39
  40 /*
  41  * offset from end of queue service tree for idle class
  42  */
  43 #define CFQ_IDLE_DELAY          (NSEC_PER_SEC / 5)
  44 /* offset from end of group service tree under time slice mode */
  45 #define CFQ_SLICE_MODE_GROUP_DELAY (NSEC_PER_SEC / 5)
  46 /* offset from end of group service under IOPS mode */
  47 #define CFQ_IOPS_MODE_GROUP_DELAY (HZ / 5)
  48
  49 /*
  50  * below this threshold, we consider thinktime immediate
  51  */
  52 #define CFQ_MIN_TT              (2 * NSEC_PER_SEC / HZ)
  53
  54 #define CFQ_SLICE_SCALE         (5)
  55 #define CFQ_HW_QUEUE_MIN        (5)
  56 #define CFQ_SERVICE_SHIFT       12
  57
  58 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  59 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  60 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  61 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  62
  63 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  64 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  65 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  66
  67 static struct kmem_cache *cfq_pool;
  68
  69 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  70 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  71 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  72
  73 #define sample_valid(samples)   ((samples) > 80)
  74 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  75
  76 /* blkio-related constants */
  77 #define CFQ_WEIGHT_LEGACY_MIN   10
  78 #define CFQ_WEIGHT_LEGACY_DFL   500
  79 #define CFQ_WEIGHT_LEGACY_MAX   1000
  80
  81 struct cfq_ttime {
  82         u64 last_end_request;
  83
  84         u64 ttime_total;
  85         u64 ttime_mean;
  86         unsigned long ttime_samples;
  87 };
  88
  89 /*
  90  * Most of our rbtree usage is for sorting with min extraction, so
  91  * if we cache the leftmost node we don't have to walk down the tree
  92  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  93  * move this into the elevator for the rq sorting as well.
  94  */
  95 struct cfq_rb_root {
  96         struct rb_root rb;
  97         struct rb_node *left;
  98         unsigned count;
  99         u64 min_vdisktime;
 100         struct cfq_ttime ttime;
 101 };
 102 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
 103                         .ttime = {.last_end_request = ktime_get_ns(),},}
 104
 105 /*
 106  * Per process-grouping structure
 107  */
 108 struct cfq_queue {
 109         /* reference count */
 110         int ref;
 111         /* various state flags, see below */
 112         unsigned int flags;
 113         /* parent cfq_data */
 114         struct cfq_data *cfqd;
 115         /* service_tree member */
 116         struct rb_node rb_node;
 117         /* service_tree key */
 118         u64 rb_key;
 119         /* prio tree member */
 120         struct rb_node p_node;
 121         /* prio tree root we belong to, if any */
 122         struct rb_root *p_root;
 123         /* sorted list of pending requests */
 124         struct rb_root sort_list;
 125         /* if fifo isn't expired, next request to serve */
 126         struct request *next_rq;
 127         /* requests queued in sort_list */
 128         int queued[2];
 129         /* currently allocated requests */
 130         int allocated[2];
 131         /* fifo list of requests in sort_list */
 132         struct list_head fifo;
 133
 134         /* time when queue got scheduled in to dispatch first request. */
 135         u64 dispatch_start;
 136         u64 allocated_slice;
 137         u64 slice_dispatch;
 138         /* time when first request from queue completed and slice started. */
 139         u64 slice_start;
 140         u64 slice_end;
 141         s64 slice_resid;
 142
 143         /* pending priority requests */
 144         int prio_pending;
 145         /* number of requests that are on the dispatch list or inside driver */
 146         int dispatched;
 147
 148         /* io prio of this group */
 149         unsigned short ioprio, org_ioprio;
 150         unsigned short ioprio_class, org_ioprio_class;
 151
 152         pid_t pid;
 153
 154         u32 seek_history;
 155         sector_t last_request_pos;
 156
 157         struct cfq_rb_root *service_tree;
 158         struct cfq_queue *new_cfqq;
 159         struct cfq_group *cfqg;
 160         /* Number of sectors dispatched from queue in single dispatch round */
 161         unsigned long nr_sectors;
 162 };
 163
 164 /*
 165  * First index in the service_trees.
 166  * IDLE is handled separately, so it has negative index
 167  */
 168 enum wl_class_t {
 169         BE_WORKLOAD = 0,
 170         RT_WORKLOAD = 1,
 171         IDLE_WORKLOAD = 2,
 172         CFQ_PRIO_NR,
 173 };
 174
 175 /*
 176  * Second index in the service_trees.
 177  */
 178 enum wl_type_t {
 179         ASYNC_WORKLOAD = 0,
 180         SYNC_NOIDLE_WORKLOAD = 1,
 181         SYNC_WORKLOAD = 2
 182 };
 183
 184 struct cfqg_stats {
 185 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 186         /* number of ios merged */
 187         struct blkg_rwstat              merged;
 188         /* total time spent on device in ns, may not be accurate w/ queueing */
 189         struct blkg_rwstat              service_time;
 190         /* total time spent waiting in scheduler queue in ns */
 191         struct blkg_rwstat              wait_time;
 192         /* number of IOs queued up */
 193         struct blkg_rwstat              queued;
 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         u64 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         u64 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 hrtimer 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_back_penalty;
 384         unsigned int cfq_back_max;
 385         unsigned int cfq_slice_async_rq;
 386         unsigned int cfq_latency;
 387         u64 cfq_fifo_expire[2];
 388         u64 cfq_slice[2];
 389         u64 cfq_slice_idle;
 390         u64 cfq_group_idle;
 391         u64 cfq_target_latency;
 392
 393         /*
 394          * Fallback dummy cfqq for extreme OOM conditions
 395          */
 396         struct cfq_queue oom_cfqq;
 397
 398         u64 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 bool cfqg_is_descendant(struct cfq_group *cfqg,
 642                                       struct cfq_group *ancestor)
 643 {
 644         return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
 645                                     cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
 646 }
 647
 648 static inline void cfqg_get(struct cfq_group *cfqg)
 649 {
 650         return blkg_get(cfqg_to_blkg(cfqg));
 651 }
 652
 653 static inline void cfqg_put(struct cfq_group *cfqg)
 654 {
 655         return blkg_put(cfqg_to_blkg(cfqg));
 656 }
 657
 658 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 659         char __pbuf[128];                                               \
 660                                                                         \
 661         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 662         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 663                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 664                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 665                           __pbuf, ##args);                              \
 666 } while (0)
 667
 668 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 669         char __pbuf[128];                                               \
 670                                                                         \
 671         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 672         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 673 } while (0)
 674
 675 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 676                                             struct cfq_group *curr_cfqg,
 677                                             unsigned int op)
 678 {
 679         blkg_rwstat_add(&cfqg->stats.queued, op, 1);
 680         cfqg_stats_end_empty_time(&cfqg->stats);
 681         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 682 }
 683
 684 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 685                         uint64_t time, unsigned long unaccounted_time)
 686 {
 687         blkg_stat_add(&cfqg->stats.time, time);
 688 #ifdef CONFIG_DEBUG_BLK_CGROUP
 689         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 690 #endif
 691 }
 692
 693 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
 694                                                unsigned int op)
 695 {
 696         blkg_rwstat_add(&cfqg->stats.queued, op, -1);
 697 }
 698
 699 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
 700                                                unsigned int op)
 701 {
 702         blkg_rwstat_add(&cfqg->stats.merged, op, 1);
 703 }
 704
 705 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 706                         uint64_t start_time, uint64_t io_start_time,
 707                         unsigned int op)
 708 {
 709         struct cfqg_stats *stats = &cfqg->stats;
 710         unsigned long long now = sched_clock();
 711
 712         if (time_after64(now, io_start_time))
 713                 blkg_rwstat_add(&stats->service_time, op, now - io_start_time);
 714         if (time_after64(io_start_time, start_time))
 715                 blkg_rwstat_add(&stats->wait_time, op,
 716                                 io_start_time - start_time);
 717 }
 718
 719 /* @stats = 0 */
 720 static void cfqg_stats_reset(struct cfqg_stats *stats)
 721 {
 722         /* queued stats shouldn't be cleared */
 723         blkg_rwstat_reset(&stats->merged);
 724         blkg_rwstat_reset(&stats->service_time);
 725         blkg_rwstat_reset(&stats->wait_time);
 726         blkg_stat_reset(&stats->time);
 727 #ifdef CONFIG_DEBUG_BLK_CGROUP
 728         blkg_stat_reset(&stats->unaccounted_time);
 729         blkg_stat_reset(&stats->avg_queue_size_sum);
 730         blkg_stat_reset(&stats->avg_queue_size_samples);
 731         blkg_stat_reset(&stats->dequeue);
 732         blkg_stat_reset(&stats->group_wait_time);
 733         blkg_stat_reset(&stats->idle_time);
 734         blkg_stat_reset(&stats->empty_time);
 735 #endif
 736 }
 737
 738 /* @to += @from */
 739 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 740 {
 741         /* queued stats shouldn't be cleared */
 742         blkg_rwstat_add_aux(&to->merged, &from->merged);
 743         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 744         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 745         blkg_stat_add_aux(&from->time, &from->time);
 746 #ifdef CONFIG_DEBUG_BLK_CGROUP
 747         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 748         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 749         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 750         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 751         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 752         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 753         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 754 #endif
 755 }
 756
 757 /*
 758  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 759  * recursive stats can still account for the amount used by this cfqg after
 760  * it's gone.
 761  */
 762 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 763 {
 764         struct cfq_group *parent = cfqg_parent(cfqg);
 765
 766         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 767
 768         if (unlikely(!parent))
 769                 return;
 770
 771         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 772         cfqg_stats_reset(&cfqg->stats);
 773 }
 774
 775 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 776
 777 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 778 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 779                                       struct cfq_group *ancestor)
 780 {
 781         return true;
 782 }
 783 static inline void cfqg_get(struct cfq_group *cfqg) { }
 784 static inline void cfqg_put(struct cfq_group *cfqg) { }
 785
 786 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 787         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 788                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 789                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 790                                 ##args)
 791 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 792
 793 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 794                         struct cfq_group *curr_cfqg, unsigned int op) { }
 795 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 796                         uint64_t time, unsigned long unaccounted_time) { }
 797 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
 798                         unsigned int op) { }
 799 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
 800                         unsigned int op) { }
 801 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 802                         uint64_t start_time, uint64_t io_start_time,
 803                         unsigned int op) { }
 804
 805 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 806
 807 #define cfq_log(cfqd, fmt, args...)     \
 808         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 809
 810 /* Traverses through cfq group service trees */
 811 #define for_each_cfqg_st(cfqg, i, j, st) \
 812         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 813                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 814                         : &cfqg->service_tree_idle; \
 815                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 816                         (i == IDLE_WORKLOAD && j == 0); \
 817                         j++, st = i < IDLE_WORKLOAD ? \
 818                         &cfqg->service_trees[i][j]: NULL) \
 819
 820 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 821         struct cfq_ttime *ttime, bool group_idle)
 822 {
 823         u64 slice;
 824         if (!sample_valid(ttime->ttime_samples))
 825                 return false;
 826         if (group_idle)
 827                 slice = cfqd->cfq_group_idle;
 828         else
 829                 slice = cfqd->cfq_slice_idle;
 830         return ttime->ttime_mean > slice;
 831 }
 832
 833 static inline bool iops_mode(struct cfq_data *cfqd)
 834 {
 835         /*
 836          * If we are not idling on queues and it is a NCQ drive, parallel
 837          * execution of requests is on and measuring time is not possible
 838          * in most of the cases until and unless we drive shallower queue
 839          * depths and that becomes a performance bottleneck. In such cases
 840          * switch to start providing fairness in terms of number of IOs.
 841          */
 842         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 843                 return true;
 844         else
 845                 return false;
 846 }
 847
 848 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 849 {
 850         if (cfq_class_idle(cfqq))
 851                 return IDLE_WORKLOAD;
 852         if (cfq_class_rt(cfqq))
 853                 return RT_WORKLOAD;
 854         return BE_WORKLOAD;
 855 }
 856
 857
 858 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 859 {
 860         if (!cfq_cfqq_sync(cfqq))
 861                 return ASYNC_WORKLOAD;
 862         if (!cfq_cfqq_idle_window(cfqq))
 863                 return SYNC_NOIDLE_WORKLOAD;
 864         return SYNC_WORKLOAD;
 865 }
 866
 867 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 868                                         struct cfq_data *cfqd,
 869                                         struct cfq_group *cfqg)
 870 {
 871         if (wl_class == IDLE_WORKLOAD)
 872                 return cfqg->service_tree_idle.count;
 873
 874         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 875                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 876                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 877 }
 878
 879 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 880                                         struct cfq_group *cfqg)
 881 {
 882         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 883                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 884 }
 885
 886 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 887 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 888                                        struct cfq_io_cq *cic, struct bio *bio);
 889
 890 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 891 {
 892         /* cic->icq is the first member, %NULL will convert to %NULL */
 893         return container_of(icq, struct cfq_io_cq, icq);
 894 }
 895
 896 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 897                                                struct io_context *ioc)
 898 {
 899         if (ioc)
 900                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 901         return NULL;
 902 }
 903
 904 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 905 {
 906         return cic->cfqq[is_sync];
 907 }
 908
 909 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 910                                 bool is_sync)
 911 {
 912         cic->cfqq[is_sync] = cfqq;
 913 }
 914
 915 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 916 {
 917         return cic->icq.q->elevator->elevator_data;
 918 }
 919
 920 /*
 921  * scheduler run of queue, if there are requests pending and no one in the
 922  * driver that will restart queueing
 923  */
 924 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 925 {
 926         if (cfqd->busy_queues) {
 927                 cfq_log(cfqd, "schedule dispatch");
 928                 kblockd_schedule_work(&cfqd->unplug_work);
 929         }
 930 }
 931
 932 /*
 933  * Scale schedule slice based on io priority. Use the sync time slice only
 934  * if a queue is marked sync and has sync io queued. A sync queue with async
 935  * io only, should not get full sync slice length.
 936  */
 937 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 938                                  unsigned short prio)
 939 {
 940         u64 base_slice = cfqd->cfq_slice[sync];
 941         u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
 942
 943         WARN_ON(prio >= IOPRIO_BE_NR);
 944
 945         return base_slice + (slice * (4 - prio));
 946 }
 947
 948 static inline u64
 949 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 950 {
 951         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 952 }
 953
 954 /**
 955  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 956  * @charge: disk time being charged
 957  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 958  *
 959  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 960  * scaling is inversely proportional.
 961  *
 962  * scaled = charge / vfraction
 963  *
 964  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 965  */
 966 static inline u64 cfqg_scale_charge(u64 charge,
 967                                     unsigned int vfraction)
 968 {
 969         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 970
 971         /* charge / vfraction */
 972         c <<= CFQ_SERVICE_SHIFT;
 973         return div_u64(c, vfraction);
 974 }
 975
 976 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 977 {
 978         s64 delta = (s64)(vdisktime - min_vdisktime);
 979         if (delta > 0)
 980                 min_vdisktime = vdisktime;
 981
 982         return min_vdisktime;
 983 }
 984
 985 static void update_min_vdisktime(struct cfq_rb_root *st)
 986 {
 987         struct cfq_group *cfqg;
 988
 989         if (st->left) {
 990                 cfqg = rb_entry_cfqg(st->left);
 991                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 992                                                   cfqg->vdisktime);
 993         }
 994 }
 995
 996 /*
 997  * get averaged number of queues of RT/BE priority.
 998  * average is updated, with a formula that gives more weight to higher numbers,
 999  * to quickly follows sudden increases and decrease slowly
1000  */
1001
1002 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1003                                         struct cfq_group *cfqg, bool rt)
1004 {
1005         unsigned min_q, max_q;
1006         unsigned mult  = cfq_hist_divisor - 1;
1007         unsigned round = cfq_hist_divisor / 2;
1008         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1009
1010         min_q = min(cfqg->busy_queues_avg[rt], busy);
1011         max_q = max(cfqg->busy_queues_avg[rt], busy);
1012         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1013                 cfq_hist_divisor;
1014         return cfqg->busy_queues_avg[rt];
1015 }
1016
1017 static inline u64
1018 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1019 {
1020         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1021 }
1022
1023 static inline u64
1024 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1025 {
1026         u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1027         if (cfqd->cfq_latency) {
1028                 /*
1029                  * interested queues (we consider only the ones with the same
1030                  * priority class in the cfq group)
1031                  */
1032                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1033                                                 cfq_class_rt(cfqq));
1034                 u64 sync_slice = cfqd->cfq_slice[1];
1035                 u64 expect_latency = sync_slice * iq;
1036                 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1037
1038                 if (expect_latency > group_slice) {
1039                         u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1040                         u64 low_slice;
1041
1042                         /* scale low_slice according to IO priority
1043                          * and sync vs async */
1044                         low_slice = div64_u64(base_low_slice*slice, sync_slice);
1045                         low_slice = min(slice, low_slice);
1046                         /* the adapted slice value is scaled to fit all iqs
1047                          * into the target latency */
1048                         slice = div64_u64(slice*group_slice, expect_latency);
1049                         slice = max(slice, low_slice);
1050                 }
1051         }
1052         return slice;
1053 }
1054
1055 static inline void
1056 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1057 {
1058         u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1059         u64 now = ktime_get_ns();
1060
1061         cfqq->slice_start = now;
1062         cfqq->slice_end = now + slice;
1063         cfqq->allocated_slice = slice;
1064         cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1065 }
1066
1067 /*
1068  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1069  * isn't valid until the first request from the dispatch is activated
1070  * and the slice time set.
1071  */
1072 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1073 {
1074         if (cfq_cfqq_slice_new(cfqq))
1075                 return false;
1076         if (ktime_get_ns() < cfqq->slice_end)
1077                 return false;
1078
1079         return true;
1080 }
1081
1082 /*
1083  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1084  * We choose the request that is closest to the head right now. Distance
1085  * behind the head is penalized and only allowed to a certain extent.
1086  */
1087 static struct request *
1088 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1089 {
1090         sector_t s1, s2, d1 = 0, d2 = 0;
1091         unsigned long back_max;
1092 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1093 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1094         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1095
1096         if (rq1 == NULL || rq1 == rq2)
1097                 return rq2;
1098         if (rq2 == NULL)
1099                 return rq1;
1100
1101         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1102                 return rq_is_sync(rq1) ? rq1 : rq2;
1103
1104         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1105                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1106
1107         s1 = blk_rq_pos(rq1);
1108         s2 = blk_rq_pos(rq2);
1109
1110         /*
1111          * by definition, 1KiB is 2 sectors
1112          */
1113         back_max = cfqd->cfq_back_max * 2;
1114
1115         /*
1116          * Strict one way elevator _except_ in the case where we allow
1117          * short backward seeks which are biased as twice the cost of a
1118          * similar forward seek.
1119          */
1120         if (s1 >= last)
1121                 d1 = s1 - last;
1122         else if (s1 + back_max >= last)
1123                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1124         else
1125                 wrap |= CFQ_RQ1_WRAP;
1126
1127         if (s2 >= last)
1128                 d2 = s2 - last;
1129         else if (s2 + back_max >= last)
1130                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1131         else
1132                 wrap |= CFQ_RQ2_WRAP;
1133
1134         /* Found required data */
1135
1136         /*
1137          * By doing switch() on the bit mask "wrap" we avoid having to
1138          * check two variables for all permutations: --> faster!
1139          */
1140         switch (wrap) {
1141         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1142                 if (d1 < d2)
1143                         return rq1;
1144                 else if (d2 < d1)
1145                         return rq2;
1146                 else {
1147                         if (s1 >= s2)
1148                                 return rq1;
1149                         else
1150                                 return rq2;
1151                 }
1152
1153         case CFQ_RQ2_WRAP:
1154                 return rq1;
1155         case CFQ_RQ1_WRAP:
1156                 return rq2;
1157         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1158         default:
1159                 /*
1160                  * Since both rqs are wrapped,
1161                  * start with the one that's further behind head
1162                  * (--> only *one* back seek required),
1163                  * since back seek takes more time than forward.
1164                  */
1165                 if (s1 <= s2)
1166                         return rq1;
1167                 else
1168                         return rq2;
1169         }
1170 }
1171
1172 /*
1173  * The below is leftmost cache rbtree addon
1174  */
1175 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1176 {
1177         /* Service tree is empty */
1178         if (!root->count)
1179                 return NULL;
1180
1181         if (!root->left)
1182                 root->left = rb_first(&root->rb);
1183
1184         if (root->left)
1185                 return rb_entry(root->left, struct cfq_queue, rb_node);
1186
1187         return NULL;
1188 }
1189
1190 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1191 {
1192         if (!root->left)
1193                 root->left = rb_first(&root->rb);
1194
1195         if (root->left)
1196                 return rb_entry_cfqg(root->left);
1197
1198         return NULL;
1199 }
1200
1201 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1202 {
1203         rb_erase(n, root);
1204         RB_CLEAR_NODE(n);
1205 }
1206
1207 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1208 {
1209         if (root->left == n)
1210                 root->left = NULL;
1211         rb_erase_init(n, &root->rb);
1212         --root->count;
1213 }
1214
1215 /*
1216  * would be nice to take fifo expire time into account as well
1217  */
1218 static struct request *
1219 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1220                   struct request *last)
1221 {
1222         struct rb_node *rbnext = rb_next(&last->rb_node);
1223         struct rb_node *rbprev = rb_prev(&last->rb_node);
1224         struct request *next = NULL, *prev = NULL;
1225
1226         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1227
1228         if (rbprev)
1229                 prev = rb_entry_rq(rbprev);
1230
1231         if (rbnext)
1232                 next = rb_entry_rq(rbnext);
1233         else {
1234                 rbnext = rb_first(&cfqq->sort_list);
1235                 if (rbnext && rbnext != &last->rb_node)
1236                         next = rb_entry_rq(rbnext);
1237         }
1238
1239         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1240 }
1241
1242 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1243                             struct cfq_queue *cfqq)
1244 {
1245         /*
1246          * just an approximation, should be ok.
1247          */
1248         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1249                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1250 }
1251
1252 static inline s64
1253 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1254 {
1255         return cfqg->vdisktime - st->min_vdisktime;
1256 }
1257
1258 static void
1259 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1260 {
1261         struct rb_node **node = &st->rb.rb_node;
1262         struct rb_node *parent = NULL;
1263         struct cfq_group *__cfqg;
1264         s64 key = cfqg_key(st, cfqg);
1265         int left = 1;
1266
1267         while (*node != NULL) {
1268                 parent = *node;
1269                 __cfqg = rb_entry_cfqg(parent);
1270
1271                 if (key < cfqg_key(st, __cfqg))
1272                         node = &parent->rb_left;
1273                 else {
1274                         node = &parent->rb_right;
1275                         left = 0;
1276                 }
1277         }
1278
1279         if (left)
1280                 st->left = &cfqg->rb_node;
1281
1282         rb_link_node(&cfqg->rb_node, parent, node);
1283         rb_insert_color(&cfqg->rb_node, &st->rb);
1284 }
1285
1286 /*
1287  * This has to be called only on activation of cfqg
1288  */
1289 static void
1290 cfq_update_group_weight(struct cfq_group *cfqg)
1291 {
1292         if (cfqg->new_weight) {
1293                 cfqg->weight = cfqg->new_weight;
1294                 cfqg->new_weight = 0;
1295         }
1296 }
1297
1298 static void
1299 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1300 {
1301         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1302
1303         if (cfqg->new_leaf_weight) {
1304                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1305                 cfqg->new_leaf_weight = 0;
1306         }
1307 }
1308
1309 static void
1310 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1311 {
1312         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1313         struct cfq_group *pos = cfqg;
1314         struct cfq_group *parent;
1315         bool propagate;
1316
1317         /* add to the service tree */
1318         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1319
1320         /*
1321          * Update leaf_weight.  We cannot update weight at this point
1322          * because cfqg might already have been activated and is
1323          * contributing its current weight to the parent's child_weight.
1324          */
1325         cfq_update_group_leaf_weight(cfqg);
1326         __cfq_group_service_tree_add(st, cfqg);
1327
1328         /*
1329          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1330          * entitled to.  vfraction is calculated by walking the tree
1331          * towards the root calculating the fraction it has at each level.
1332          * The compounded ratio is how much vfraction @cfqg owns.
1333          *
1334          * Start with the proportion tasks in this cfqg has against active
1335          * children cfqgs - its leaf_weight against children_weight.
1336          */
1337         propagate = !pos->nr_active++;
1338         pos->children_weight += pos->leaf_weight;
1339         vfr = vfr * pos->leaf_weight / pos->children_weight;
1340
1341         /*
1342          * Compound ->weight walking up the tree.  Both activation and
1343          * vfraction calculation are done in the same loop.  Propagation
1344          * stops once an already activated node is met.  vfraction
1345          * calculation should always continue to the root.
1346          */
1347         while ((parent = cfqg_parent(pos))) {
1348                 if (propagate) {
1349                         cfq_update_group_weight(pos);
1350                         propagate = !parent->nr_active++;
1351                         parent->children_weight += pos->weight;
1352                 }
1353                 vfr = vfr * pos->weight / parent->children_weight;
1354                 pos = parent;
1355         }
1356
1357         cfqg->vfraction = max_t(unsigned, vfr, 1);
1358 }
1359
1360 static inline u64 cfq_get_cfqg_vdisktime_delay(struct cfq_data *cfqd)
1361 {
1362         if (!iops_mode(cfqd))
1363                 return CFQ_SLICE_MODE_GROUP_DELAY;
1364         else
1365                 return CFQ_IOPS_MODE_GROUP_DELAY;
1366 }
1367
1368 static void
1369 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1370 {
1371         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1372         struct cfq_group *__cfqg;
1373         struct rb_node *n;
1374
1375         cfqg->nr_cfqq++;
1376         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1377                 return;
1378
1379         /*
1380          * Currently put the group at the end. Later implement something
1381          * so that groups get lesser vtime based on their weights, so that
1382          * if group does not loose all if it was not continuously backlogged.
1383          */
1384         n = rb_last(&st->rb);
1385         if (n) {
1386                 __cfqg = rb_entry_cfqg(n);
1387                 cfqg->vdisktime = __cfqg->vdisktime +
1388                         cfq_get_cfqg_vdisktime_delay(cfqd);
1389         } else
1390                 cfqg->vdisktime = st->min_vdisktime;
1391         cfq_group_service_tree_add(st, cfqg);
1392 }
1393
1394 static void
1395 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1396 {
1397         struct cfq_group *pos = cfqg;
1398         bool propagate;
1399
1400         /*
1401          * Undo activation from cfq_group_service_tree_add().  Deactivate
1402          * @cfqg and propagate deactivation upwards.
1403          */
1404         propagate = !--pos->nr_active;
1405         pos->children_weight -= pos->leaf_weight;
1406
1407         while (propagate) {
1408                 struct cfq_group *parent = cfqg_parent(pos);
1409
1410                 /* @pos has 0 nr_active at this point */
1411                 WARN_ON_ONCE(pos->children_weight);
1412                 pos->vfraction = 0;
1413
1414                 if (!parent)
1415                         break;
1416
1417                 propagate = !--parent->nr_active;
1418                 parent->children_weight -= pos->weight;
1419                 pos = parent;
1420         }
1421
1422         /* remove from the service tree */
1423         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1424                 cfq_rb_erase(&cfqg->rb_node, st);
1425 }
1426
1427 static void
1428 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1429 {
1430         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1431
1432         BUG_ON(cfqg->nr_cfqq < 1);
1433         cfqg->nr_cfqq--;
1434
1435         /* If there are other cfq queues under this group, don't delete it */
1436         if (cfqg->nr_cfqq)
1437                 return;
1438
1439         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1440         cfq_group_service_tree_del(st, cfqg);
1441         cfqg->saved_wl_slice = 0;
1442         cfqg_stats_update_dequeue(cfqg);
1443 }
1444
1445 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1446                                        u64 *unaccounted_time)
1447 {
1448         u64 slice_used;
1449         u64 now = ktime_get_ns();
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 == now) {
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(u64, (now - cfqq->dispatch_start),
1463                                         jiffies_to_nsecs(1));
1464         } else {
1465                 slice_used = now - 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 (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         u64 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         u64 now = ktime_get_ns();
1487
1488         BUG_ON(nr_sync < 0);
1489         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1490
1491         if (iops_mode(cfqd))
1492                 charge = cfqq->slice_dispatch;
1493         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1494                 charge = cfqq->allocated_slice;
1495
1496         /*
1497          * Can't update vdisktime while on service tree and cfqg->vfraction
1498          * is valid only while on it.  Cache vfr, leave the service tree,
1499          * update vdisktime and go back on.  The re-addition to the tree
1500          * will also update the weights as necessary.
1501          */
1502         vfr = cfqg->vfraction;
1503         cfq_group_service_tree_del(st, cfqg);
1504         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1505         cfq_group_service_tree_add(st, cfqg);
1506
1507         /* This group is being expired. Save the context */
1508         if (cfqd->workload_expires > now) {
1509                 cfqg->saved_wl_slice = cfqd->workload_expires - now;
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=%llu disp=%llu charge=%llu 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 = ktime_get_ns();
1542 }
1543
1544 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1545 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1546                             bool on_dfl, bool reset_dev, bool is_leaf_weight);
1547
1548 static void cfqg_stats_exit(struct cfqg_stats *stats)
1549 {
1550         blkg_rwstat_exit(&stats->merged);
1551         blkg_rwstat_exit(&stats->service_time);
1552         blkg_rwstat_exit(&stats->wait_time);
1553         blkg_rwstat_exit(&stats->queued);
1554         blkg_stat_exit(&stats->time);
1555 #ifdef CONFIG_DEBUG_BLK_CGROUP
1556         blkg_stat_exit(&stats->unaccounted_time);
1557         blkg_stat_exit(&stats->avg_queue_size_sum);
1558         blkg_stat_exit(&stats->avg_queue_size_samples);
1559         blkg_stat_exit(&stats->dequeue);
1560         blkg_stat_exit(&stats->group_wait_time);
1561         blkg_stat_exit(&stats->idle_time);
1562         blkg_stat_exit(&stats->empty_time);
1563 #endif
1564 }
1565
1566 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1567 {
1568         if (blkg_rwstat_init(&stats->merged, gfp) ||
1569             blkg_rwstat_init(&stats->service_time, gfp) ||
1570             blkg_rwstat_init(&stats->wait_time, gfp) ||
1571             blkg_rwstat_init(&stats->queued, gfp) ||
1572             blkg_stat_init(&stats->time, gfp))
1573                 goto err;
1574
1575 #ifdef CONFIG_DEBUG_BLK_CGROUP
1576         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1577             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1578             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1579             blkg_stat_init(&stats->dequeue, gfp) ||
1580             blkg_stat_init(&stats->group_wait_time, gfp) ||
1581             blkg_stat_init(&stats->idle_time, gfp) ||
1582             blkg_stat_init(&stats->empty_time, gfp))
1583                 goto err;
1584 #endif
1585         return 0;
1586 err:
1587         cfqg_stats_exit(stats);
1588         return -ENOMEM;
1589 }
1590
1591 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1592 {
1593         struct cfq_group_data *cgd;
1594
1595         cgd = kzalloc(sizeof(*cgd), gfp);
1596         if (!cgd)
1597                 return NULL;
1598         return &cgd->cpd;
1599 }
1600
1601 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1602 {
1603         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1604         unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1605                               CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1606
1607         if (cpd_to_blkcg(cpd) == &blkcg_root)
1608                 weight *= 2;
1609
1610         cgd->weight = weight;
1611         cgd->leaf_weight = weight;
1612 }
1613
1614 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1615 {
1616         kfree(cpd_to_cfqgd(cpd));
1617 }
1618
1619 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1620 {
1621         struct blkcg *blkcg = cpd_to_blkcg(cpd);
1622         bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1623         unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1624
1625         if (blkcg == &blkcg_root)
1626                 weight *= 2;
1627
1628         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1629         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1630 }
1631
1632 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1633 {
1634         struct cfq_group *cfqg;
1635
1636         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1637         if (!cfqg)
1638                 return NULL;
1639
1640         cfq_init_cfqg_base(cfqg);
1641         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1642                 kfree(cfqg);
1643                 return NULL;
1644         }
1645
1646         return &cfqg->pd;
1647 }
1648
1649 static void cfq_pd_init(struct blkg_policy_data *pd)
1650 {
1651         struct cfq_group *cfqg = pd_to_cfqg(pd);
1652         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1653
1654         cfqg->weight = cgd->weight;
1655         cfqg->leaf_weight = cgd->leaf_weight;
1656 }
1657
1658 static void cfq_pd_offline(struct blkg_policy_data *pd)
1659 {
1660         struct cfq_group *cfqg = pd_to_cfqg(pd);
1661         int i;
1662
1663         for (i = 0; i < IOPRIO_BE_NR; i++) {
1664                 if (cfqg->async_cfqq[0][i])
1665                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1666                 if (cfqg->async_cfqq[1][i])
1667                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1668         }
1669
1670         if (cfqg->async_idle_cfqq)
1671                 cfq_put_queue(cfqg->async_idle_cfqq);
1672
1673         /*
1674          * @blkg is going offline and will be ignored by
1675          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1676          * that they don't get lost.  If IOs complete after this point, the
1677          * stats for them will be lost.  Oh well...
1678          */
1679         cfqg_stats_xfer_dead(cfqg);
1680 }
1681
1682 static void cfq_pd_free(struct blkg_policy_data *pd)
1683 {
1684         struct cfq_group *cfqg = pd_to_cfqg(pd);
1685
1686         cfqg_stats_exit(&cfqg->stats);
1687         return kfree(cfqg);
1688 }
1689
1690 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1691 {
1692         struct cfq_group *cfqg = pd_to_cfqg(pd);
1693
1694         cfqg_stats_reset(&cfqg->stats);
1695 }
1696
1697 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1698                                          struct blkcg *blkcg)
1699 {
1700         struct blkcg_gq *blkg;
1701
1702         blkg = blkg_lookup(blkcg, cfqd->queue);
1703         if (likely(blkg))
1704                 return blkg_to_cfqg(blkg);
1705         return NULL;
1706 }
1707
1708 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1709 {
1710         cfqq->cfqg = cfqg;
1711         /* cfqq reference on cfqg */
1712         cfqg_get(cfqg);
1713 }
1714
1715 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1716                                      struct blkg_policy_data *pd, int off)
1717 {
1718         struct cfq_group *cfqg = pd_to_cfqg(pd);
1719
1720         if (!cfqg->dev_weight)
1721                 return 0;
1722         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1723 }
1724
1725 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1726 {
1727         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1728                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1729                           0, false);
1730         return 0;
1731 }
1732
1733 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1734                                           struct blkg_policy_data *pd, int off)
1735 {
1736         struct cfq_group *cfqg = pd_to_cfqg(pd);
1737
1738         if (!cfqg->dev_leaf_weight)
1739                 return 0;
1740         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1741 }
1742
1743 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1744 {
1745         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1746                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1747                           0, false);
1748         return 0;
1749 }
1750
1751 static int cfq_print_weight(struct seq_file *sf, void *v)
1752 {
1753         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1754         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1755         unsigned int val = 0;
1756
1757         if (cgd)
1758                 val = cgd->weight;
1759
1760         seq_printf(sf, "%u\n", val);
1761         return 0;
1762 }
1763
1764 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1765 {
1766         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1767         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1768         unsigned int val = 0;
1769
1770         if (cgd)
1771                 val = cgd->leaf_weight;
1772
1773         seq_printf(sf, "%u\n", val);
1774         return 0;
1775 }
1776
1777 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1778                                         char *buf, size_t nbytes, loff_t off,
1779                                         bool on_dfl, bool is_leaf_weight)
1780 {
1781         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1782         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1783         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1784         struct blkg_conf_ctx ctx;
1785         struct cfq_group *cfqg;
1786         struct cfq_group_data *cfqgd;
1787         int ret;
1788         u64 v;
1789
1790         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1791         if (ret)
1792                 return ret;
1793
1794         if (sscanf(ctx.body, "%llu", &v) == 1) {
1795                 /* require "default" on dfl */
1796                 ret = -ERANGE;
1797                 if (!v && on_dfl)
1798                         goto out_finish;
1799         } else if (!strcmp(strim(ctx.body), "default")) {
1800                 v = 0;
1801         } else {
1802                 ret = -EINVAL;
1803                 goto out_finish;
1804         }
1805
1806         cfqg = blkg_to_cfqg(ctx.blkg);
1807         cfqgd = blkcg_to_cfqgd(blkcg);
1808
1809         ret = -ERANGE;
1810         if (!v || (v >= min && v <= max)) {
1811                 if (!is_leaf_weight) {
1812                         cfqg->dev_weight = v;
1813                         cfqg->new_weight = v ?: cfqgd->weight;
1814                 } else {
1815                         cfqg->dev_leaf_weight = v;
1816                         cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1817                 }
1818                 ret = 0;
1819         }
1820 out_finish:
1821         blkg_conf_finish(&ctx);
1822         return ret ?: nbytes;
1823 }
1824
1825 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1826                                       char *buf, size_t nbytes, loff_t off)
1827 {
1828         return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1829 }
1830
1831 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1832                                            char *buf, size_t nbytes, loff_t off)
1833 {
1834         return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1835 }
1836
1837 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1838                             bool on_dfl, bool reset_dev, bool is_leaf_weight)
1839 {
1840         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1841         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1842         struct blkcg *blkcg = css_to_blkcg(css);
1843         struct blkcg_gq *blkg;
1844         struct cfq_group_data *cfqgd;
1845         int ret = 0;
1846
1847         if (val < min || val > max)
1848                 return -ERANGE;
1849
1850         spin_lock_irq(&blkcg->lock);
1851         cfqgd = blkcg_to_cfqgd(blkcg);
1852         if (!cfqgd) {
1853                 ret = -EINVAL;
1854                 goto out;
1855         }
1856
1857         if (!is_leaf_weight)
1858                 cfqgd->weight = val;
1859         else
1860                 cfqgd->leaf_weight = val;
1861
1862         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1863                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1864
1865                 if (!cfqg)
1866                         continue;
1867
1868                 if (!is_leaf_weight) {
1869                         if (reset_dev)
1870                                 cfqg->dev_weight = 0;
1871                         if (!cfqg->dev_weight)
1872                                 cfqg->new_weight = cfqgd->weight;
1873                 } else {
1874                         if (reset_dev)
1875                                 cfqg->dev_leaf_weight = 0;
1876                         if (!cfqg->dev_leaf_weight)
1877                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1878                 }
1879         }
1880
1881 out:
1882         spin_unlock_irq(&blkcg->lock);
1883         return ret;
1884 }
1885
1886 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1887                           u64 val)
1888 {
1889         return __cfq_set_weight(css, val, false, false, false);
1890 }
1891
1892 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1893                                struct cftype *cft, u64 val)
1894 {
1895         return __cfq_set_weight(css, val, false, false, true);
1896 }
1897
1898 static int cfqg_print_stat(struct seq_file *sf, void *v)
1899 {
1900         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1901                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1902         return 0;
1903 }
1904
1905 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1906 {
1907         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1908                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1909         return 0;
1910 }
1911
1912 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1913                                       struct blkg_policy_data *pd, int off)
1914 {
1915         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1916                                           &blkcg_policy_cfq, off);
1917         return __blkg_prfill_u64(sf, pd, sum);
1918 }
1919
1920 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1921                                         struct blkg_policy_data *pd, int off)
1922 {
1923         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1924                                                         &blkcg_policy_cfq, off);
1925         return __blkg_prfill_rwstat(sf, pd, &sum);
1926 }
1927
1928 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1929 {
1930         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1931                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1932                           seq_cft(sf)->private, false);
1933         return 0;
1934 }
1935
1936 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1937 {
1938         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1939                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1940                           seq_cft(sf)->private, true);
1941         return 0;
1942 }
1943
1944 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1945                                int off)
1946 {
1947         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1948
1949         return __blkg_prfill_u64(sf, pd, sum >> 9);
1950 }
1951
1952 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1953 {
1954         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1955                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1956         return 0;
1957 }
1958
1959 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1960                                          struct blkg_policy_data *pd, int off)
1961 {
1962         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1963                                         offsetof(struct blkcg_gq, stat_bytes));
1964         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1965                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1966
1967         return __blkg_prfill_u64(sf, pd, sum >> 9);
1968 }
1969
1970 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1971 {
1972         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1973                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1974                           false);
1975         return 0;
1976 }
1977
1978 #ifdef CONFIG_DEBUG_BLK_CGROUP
1979 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1980                                       struct blkg_policy_data *pd, int off)
1981 {
1982         struct cfq_group *cfqg = pd_to_cfqg(pd);
1983         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1984         u64 v = 0;
1985
1986         if (samples) {
1987                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1988                 v = div64_u64(v, samples);
1989         }
1990         __blkg_prfill_u64(sf, pd, v);
1991         return 0;
1992 }
1993
1994 /* print avg_queue_size */
1995 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1996 {
1997         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1998                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1999                           0, false);
2000         return 0;
2001 }
2002 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2003
2004 static struct cftype cfq_blkcg_legacy_files[] = {
2005         /* on root, weight is mapped to leaf_weight */
2006         {
2007                 .name = "weight_device",
2008                 .flags = CFTYPE_ONLY_ON_ROOT,
2009                 .seq_show = cfqg_print_leaf_weight_device,
2010                 .write = cfqg_set_leaf_weight_device,
2011         },
2012         {
2013                 .name = "weight",
2014                 .flags = CFTYPE_ONLY_ON_ROOT,
2015                 .seq_show = cfq_print_leaf_weight,
2016                 .write_u64 = cfq_set_leaf_weight,
2017         },
2018
2019         /* no such mapping necessary for !roots */
2020         {
2021                 .name = "weight_device",
2022                 .flags = CFTYPE_NOT_ON_ROOT,
2023                 .seq_show = cfqg_print_weight_device,
2024                 .write = cfqg_set_weight_device,
2025         },
2026         {
2027                 .name = "weight",
2028                 .flags = CFTYPE_NOT_ON_ROOT,
2029                 .seq_show = cfq_print_weight,
2030                 .write_u64 = cfq_set_weight,
2031         },
2032
2033         {
2034                 .name = "leaf_weight_device",
2035                 .seq_show = cfqg_print_leaf_weight_device,
2036                 .write = cfqg_set_leaf_weight_device,
2037         },
2038         {
2039                 .name = "leaf_weight",
2040                 .seq_show = cfq_print_leaf_weight,
2041                 .write_u64 = cfq_set_leaf_weight,
2042         },
2043
2044         /* statistics, covers only the tasks in the cfqg */
2045         {
2046                 .name = "time",
2047                 .private = offsetof(struct cfq_group, stats.time),
2048                 .seq_show = cfqg_print_stat,
2049         },
2050         {
2051                 .name = "sectors",
2052                 .seq_show = cfqg_print_stat_sectors,
2053         },
2054         {
2055                 .name = "io_service_bytes",
2056                 .private = (unsigned long)&blkcg_policy_cfq,
2057                 .seq_show = blkg_print_stat_bytes,
2058         },
2059         {
2060                 .name = "io_serviced",
2061                 .private = (unsigned long)&blkcg_policy_cfq,
2062                 .seq_show = blkg_print_stat_ios,
2063         },
2064         {
2065                 .name = "io_service_time",
2066                 .private = offsetof(struct cfq_group, stats.service_time),
2067                 .seq_show = cfqg_print_rwstat,
2068         },
2069         {
2070                 .name = "io_wait_time",
2071                 .private = offsetof(struct cfq_group, stats.wait_time),
2072                 .seq_show = cfqg_print_rwstat,
2073         },
2074         {
2075                 .name = "io_merged",
2076                 .private = offsetof(struct cfq_group, stats.merged),
2077                 .seq_show = cfqg_print_rwstat,
2078         },
2079         {
2080                 .name = "io_queued",
2081                 .private = offsetof(struct cfq_group, stats.queued),
2082                 .seq_show = cfqg_print_rwstat,
2083         },
2084
2085         /* the same statictics which cover the cfqg and its descendants */
2086         {
2087                 .name = "time_recursive",
2088                 .private = offsetof(struct cfq_group, stats.time),
2089                 .seq_show = cfqg_print_stat_recursive,
2090         },
2091         {
2092                 .name = "sectors_recursive",
2093                 .seq_show = cfqg_print_stat_sectors_recursive,
2094         },
2095         {
2096                 .name = "io_service_bytes_recursive",
2097                 .private = (unsigned long)&blkcg_policy_cfq,
2098                 .seq_show = blkg_print_stat_bytes_recursive,
2099         },
2100         {
2101                 .name = "io_serviced_recursive",
2102                 .private = (unsigned long)&blkcg_policy_cfq,
2103                 .seq_show = blkg_print_stat_ios_recursive,
2104         },
2105         {
2106                 .name = "io_service_time_recursive",
2107                 .private = offsetof(struct cfq_group, stats.service_time),
2108                 .seq_show = cfqg_print_rwstat_recursive,
2109         },
2110         {
2111                 .name = "io_wait_time_recursive",
2112                 .private = offsetof(struct cfq_group, stats.wait_time),
2113                 .seq_show = cfqg_print_rwstat_recursive,
2114         },
2115         {
2116                 .name = "io_merged_recursive",
2117                 .private = offsetof(struct cfq_group, stats.merged),
2118                 .seq_show = cfqg_print_rwstat_recursive,
2119         },
2120         {
2121                 .name = "io_queued_recursive",
2122                 .private = offsetof(struct cfq_group, stats.queued),
2123                 .seq_show = cfqg_print_rwstat_recursive,
2124         },
2125 #ifdef CONFIG_DEBUG_BLK_CGROUP
2126         {
2127                 .name = "avg_queue_size",
2128                 .seq_show = cfqg_print_avg_queue_size,
2129         },
2130         {
2131                 .name = "group_wait_time",
2132                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2133                 .seq_show = cfqg_print_stat,
2134         },
2135         {
2136                 .name = "idle_time",
2137                 .private = offsetof(struct cfq_group, stats.idle_time),
2138                 .seq_show = cfqg_print_stat,
2139         },
2140         {
2141                 .name = "empty_time",
2142                 .private = offsetof(struct cfq_group, stats.empty_time),
2143                 .seq_show = cfqg_print_stat,
2144         },
2145         {
2146                 .name = "dequeue",
2147                 .private = offsetof(struct cfq_group, stats.dequeue),
2148                 .seq_show = cfqg_print_stat,
2149         },
2150         {
2151                 .name = "unaccounted_time",
2152                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2153                 .seq_show = cfqg_print_stat,
2154         },
2155 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2156         { }     /* terminate */
2157 };
2158
2159 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2160 {
2161         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2162         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2163
2164         seq_printf(sf, "default %u\n", cgd->weight);
2165         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2166                           &blkcg_policy_cfq, 0, false);
2167         return 0;
2168 }
2169
2170 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2171                                      char *buf, size_t nbytes, loff_t off)
2172 {
2173         char *endp;
2174         int ret;
2175         u64 v;
2176
2177         buf = strim(buf);
2178
2179         /* "WEIGHT" or "default WEIGHT" sets the default weight */
2180         v = simple_strtoull(buf, &endp, 0);
2181         if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2182                 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2183                 return ret ?: nbytes;
2184         }
2185
2186         /* "MAJ:MIN WEIGHT" */
2187         return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2188 }
2189
2190 static struct cftype cfq_blkcg_files[] = {
2191         {
2192                 .name = "weight",
2193                 .flags = CFTYPE_NOT_ON_ROOT,
2194                 .seq_show = cfq_print_weight_on_dfl,
2195                 .write = cfq_set_weight_on_dfl,
2196         },
2197         { }     /* terminate */
2198 };
2199
2200 #else /* GROUP_IOSCHED */
2201 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2202                                          struct blkcg *blkcg)
2203 {
2204         return cfqd->root_group;
2205 }
2206
2207 static inline void
2208 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2209         cfqq->cfqg = cfqg;
2210 }
2211
2212 #endif /* GROUP_IOSCHED */
2213
2214 /*
2215  * The cfqd->service_trees holds all pending cfq_queue's that have
2216  * requests waiting to be processed. It is sorted in the order that
2217  * we will service the queues.
2218  */
2219 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2220                                  bool add_front)
2221 {
2222         struct rb_node **p, *parent;
2223         struct cfq_queue *__cfqq;
2224         u64 rb_key;
2225         struct cfq_rb_root *st;
2226         int left;
2227         int new_cfqq = 1;
2228         u64 now = ktime_get_ns();
2229
2230         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2231         if (cfq_class_idle(cfqq)) {
2232                 rb_key = CFQ_IDLE_DELAY;
2233                 parent = rb_last(&st->rb);
2234                 if (parent && parent != &cfqq->rb_node) {
2235                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2236                         rb_key += __cfqq->rb_key;
2237                 } else
2238                         rb_key += now;
2239         } else if (!add_front) {
2240                 /*
2241                  * Get our rb key offset. Subtract any residual slice
2242                  * value carried from last service. A negative resid
2243                  * count indicates slice overrun, and this should position
2244                  * the next service time further away in the tree.
2245                  */
2246                 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2247                 rb_key -= cfqq->slice_resid;
2248                 cfqq->slice_resid = 0;
2249         } else {
2250                 rb_key = -NSEC_PER_SEC;
2251                 __cfqq = cfq_rb_first(st);
2252                 rb_key += __cfqq ? __cfqq->rb_key : now;
2253         }
2254
2255         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2256                 new_cfqq = 0;
2257                 /*
2258                  * same position, nothing more to do
2259                  */
2260                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2261                         return;
2262
2263                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2264                 cfqq->service_tree = NULL;
2265         }
2266
2267         left = 1;
2268         parent = NULL;
2269         cfqq->service_tree = st;
2270         p = &st->rb.rb_node;
2271         while (*p) {
2272                 parent = *p;
2273                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2274
2275                 /*
2276                  * sort by key, that represents service time.
2277                  */
2278                 if (rb_key < __cfqq->rb_key)
2279                         p = &parent->rb_left;
2280                 else {
2281                         p = &parent->rb_right;
2282                         left = 0;
2283                 }
2284         }
2285
2286         if (left)
2287                 st->left = &cfqq->rb_node;
2288
2289         cfqq->rb_key = rb_key;
2290         rb_link_node(&cfqq->rb_node, parent, p);
2291         rb_insert_color(&cfqq->rb_node, &st->rb);
2292         st->count++;
2293         if (add_front || !new_cfqq)
2294                 return;
2295         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2296 }
2297
2298 static struct cfq_queue *
2299 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2300                      sector_t sector, struct rb_node **ret_parent,
2301                      struct rb_node ***rb_link)
2302 {
2303         struct rb_node **p, *parent;
2304         struct cfq_queue *cfqq = NULL;
2305
2306         parent = NULL;
2307         p = &root->rb_node;
2308         while (*p) {
2309                 struct rb_node **n;
2310
2311                 parent = *p;
2312                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2313
2314                 /*
2315                  * Sort strictly based on sector.  Smallest to the left,
2316                  * largest to the right.
2317                  */
2318                 if (sector > blk_rq_pos(cfqq->next_rq))
2319                         n = &(*p)->rb_right;
2320                 else if (sector < blk_rq_pos(cfqq->next_rq))
2321                         n = &(*p)->rb_left;
2322                 else
2323                         break;
2324                 p = n;
2325                 cfqq = NULL;
2326         }
2327
2328         *ret_parent = parent;
2329         if (rb_link)
2330                 *rb_link = p;
2331         return cfqq;
2332 }
2333
2334 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2335 {
2336         struct rb_node **p, *parent;
2337         struct cfq_queue *__cfqq;
2338
2339         if (cfqq->p_root) {
2340                 rb_erase(&cfqq->p_node, cfqq->p_root);
2341                 cfqq->p_root = NULL;
2342         }
2343
2344         if (cfq_class_idle(cfqq))
2345                 return;
2346         if (!cfqq->next_rq)
2347                 return;
2348
2349         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2350         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2351                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2352         if (!__cfqq) {
2353                 rb_link_node(&cfqq->p_node, parent, p);
2354                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2355         } else
2356                 cfqq->p_root = NULL;
2357 }
2358
2359 /*
2360  * Update cfqq's position in the service tree.
2361  */
2362 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2363 {
2364         /*
2365          * Resorting requires the cfqq to be on the RR list already.
2366          */
2367         if (cfq_cfqq_on_rr(cfqq)) {
2368                 cfq_service_tree_add(cfqd, cfqq, 0);
2369                 cfq_prio_tree_add(cfqd, cfqq);
2370         }
2371 }
2372
2373 /*
2374  * add to busy list of queues for service, trying to be fair in ordering
2375  * the pending list according to last request service
2376  */
2377 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2378 {
2379         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2380         BUG_ON(cfq_cfqq_on_rr(cfqq));
2381         cfq_mark_cfqq_on_rr(cfqq);
2382         cfqd->busy_queues++;
2383         if (cfq_cfqq_sync(cfqq))
2384                 cfqd->busy_sync_queues++;
2385
2386         cfq_resort_rr_list(cfqd, cfqq);
2387 }
2388
2389 /*
2390  * Called when the cfqq no longer has requests pending, remove it from
2391  * the service tree.
2392  */
2393 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2394 {
2395         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2396         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2397         cfq_clear_cfqq_on_rr(cfqq);
2398
2399         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2400                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2401                 cfqq->service_tree = NULL;
2402         }
2403         if (cfqq->p_root) {
2404                 rb_erase(&cfqq->p_node, cfqq->p_root);
2405                 cfqq->p_root = NULL;
2406         }
2407
2408         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2409         BUG_ON(!cfqd->busy_queues);
2410         cfqd->busy_queues--;
2411         if (cfq_cfqq_sync(cfqq))
2412                 cfqd->busy_sync_queues--;
2413 }
2414
2415 /*
2416  * rb tree support functions
2417  */
2418 static void cfq_del_rq_rb(struct request *rq)
2419 {
2420         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2421         const int sync = rq_is_sync(rq);
2422
2423         BUG_ON(!cfqq->queued[sync]);
2424         cfqq->queued[sync]--;
2425
2426         elv_rb_del(&cfqq->sort_list, rq);
2427
2428         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2429                 /*
2430                  * Queue will be deleted from service tree when we actually
2431                  * expire it later. Right now just remove it from prio tree
2432                  * as it is empty.
2433                  */
2434                 if (cfqq->p_root) {
2435                         rb_erase(&cfqq->p_node, cfqq->p_root);
2436                         cfqq->p_root = NULL;
2437                 }
2438         }
2439 }
2440
2441 static void cfq_add_rq_rb(struct request *rq)
2442 {
2443         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2444         struct cfq_data *cfqd = cfqq->cfqd;
2445         struct request *prev;
2446
2447         cfqq->queued[rq_is_sync(rq)]++;
2448
2449         elv_rb_add(&cfqq->sort_list, rq);
2450
2451         if (!cfq_cfqq_on_rr(cfqq))
2452                 cfq_add_cfqq_rr(cfqd, cfqq);
2453
2454         /*
2455          * check if this request is a better next-serve candidate
2456          */
2457         prev = cfqq->next_rq;
2458         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2459
2460         /*
2461          * adjust priority tree position, if ->next_rq changes
2462          */
2463         if (prev != cfqq->next_rq)
2464                 cfq_prio_tree_add(cfqd, cfqq);
2465
2466         BUG_ON(!cfqq->next_rq);
2467 }
2468
2469 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2470 {
2471         elv_rb_del(&cfqq->sort_list, rq);
2472         cfqq->queued[rq_is_sync(rq)]--;
2473         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2474         cfq_add_rq_rb(rq);
2475         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2476                                  rq->cmd_flags);
2477 }
2478
2479 static struct request *
2480 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2481 {
2482         struct task_struct *tsk = current;
2483         struct cfq_io_cq *cic;
2484         struct cfq_queue *cfqq;
2485
2486         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2487         if (!cic)
2488                 return NULL;
2489
2490         cfqq = cic_to_cfqq(cic, op_is_sync(bio->bi_opf));
2491         if (cfqq)
2492                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2493
2494         return NULL;
2495 }
2496
2497 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2498 {
2499         struct cfq_data *cfqd = q->elevator->elevator_data;
2500
2501         cfqd->rq_in_driver++;
2502         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2503                                                 cfqd->rq_in_driver);
2504
2505         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2506 }
2507
2508 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2509 {
2510         struct cfq_data *cfqd = q->elevator->elevator_data;
2511
2512         WARN_ON(!cfqd->rq_in_driver);
2513         cfqd->rq_in_driver--;
2514         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2515                                                 cfqd->rq_in_driver);
2516 }
2517
2518 static void cfq_remove_request(struct request *rq)
2519 {
2520         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2521
2522         if (cfqq->next_rq == rq)
2523                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2524
2525         list_del_init(&rq->queuelist);
2526         cfq_del_rq_rb(rq);
2527
2528         cfqq->cfqd->rq_queued--;
2529         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2530         if (rq->cmd_flags & REQ_PRIO) {
2531                 WARN_ON(!cfqq->prio_pending);
2532                 cfqq->prio_pending--;
2533         }
2534 }
2535
2536 static enum elv_merge cfq_merge(struct request_queue *q, struct request **req,
2537                      struct bio *bio)
2538 {
2539         struct cfq_data *cfqd = q->elevator->elevator_data;
2540         struct request *__rq;
2541
2542         __rq = cfq_find_rq_fmerge(cfqd, bio);
2543         if (__rq && elv_bio_merge_ok(__rq, bio)) {
2544                 *req = __rq;
2545                 return ELEVATOR_FRONT_MERGE;
2546         }
2547
2548         return ELEVATOR_NO_MERGE;
2549 }
2550
2551 static void cfq_merged_request(struct request_queue *q, struct request *req,
2552                                enum elv_merge type)
2553 {
2554         if (type == ELEVATOR_FRONT_MERGE) {
2555                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2556
2557                 cfq_reposition_rq_rb(cfqq, req);
2558         }
2559 }
2560
2561 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2562                                 struct bio *bio)
2563 {
2564         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_opf);
2565 }
2566
2567 static void
2568 cfq_merged_requests(struct request_queue *q, struct request *rq,
2569                     struct request *next)
2570 {
2571         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2572         struct cfq_data *cfqd = q->elevator->elevator_data;
2573
2574         /*
2575          * reposition in fifo if next is older than rq
2576          */
2577         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2578             next->fifo_time < rq->fifo_time &&
2579             cfqq == RQ_CFQQ(next)) {
2580                 list_move(&rq->queuelist, &next->queuelist);
2581                 rq->fifo_time = next->fifo_time;
2582         }
2583
2584         if (cfqq->next_rq == next)
2585                 cfqq->next_rq = rq;
2586         cfq_remove_request(next);
2587         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2588
2589         cfqq = RQ_CFQQ(next);
2590         /*
2591          * all requests of this queue are merged to other queues, delete it
2592          * from the service tree. If it's the active_queue,
2593          * cfq_dispatch_requests() will choose to expire it or do idle
2594          */
2595         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2596             cfqq != cfqd->active_queue)
2597                 cfq_del_cfqq_rr(cfqd, cfqq);
2598 }
2599
2600 static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2601                                struct bio *bio)
2602 {
2603         struct cfq_data *cfqd = q->elevator->elevator_data;
2604         bool is_sync = op_is_sync(bio->bi_opf);
2605         struct cfq_io_cq *cic;
2606         struct cfq_queue *cfqq;
2607
2608         /*
2609          * Disallow merge of a sync bio into an async request.
2610          */
2611         if (is_sync && !rq_is_sync(rq))
2612                 return false;
2613
2614         /*
2615          * Lookup the cfqq that this bio will be queued with and allow
2616          * merge only if rq is queued there.
2617          */
2618         cic = cfq_cic_lookup(cfqd, current->io_context);
2619         if (!cic)
2620                 return false;
2621
2622         cfqq = cic_to_cfqq(cic, is_sync);
2623         return cfqq == RQ_CFQQ(rq);
2624 }
2625
2626 static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2627                               struct request *next)
2628 {
2629         return RQ_CFQQ(rq) == RQ_CFQQ(next);
2630 }
2631
2632 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2633 {
2634         hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2635         cfqg_stats_update_idle_time(cfqq->cfqg);
2636 }
2637
2638 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2639                                    struct cfq_queue *cfqq)
2640 {
2641         if (cfqq) {
2642                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2643                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2644                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2645                 cfqq->slice_start = 0;
2646                 cfqq->dispatch_start = ktime_get_ns();
2647                 cfqq->allocated_slice = 0;
2648                 cfqq->slice_end = 0;
2649                 cfqq->slice_dispatch = 0;
2650                 cfqq->nr_sectors = 0;
2651
2652                 cfq_clear_cfqq_wait_request(cfqq);
2653                 cfq_clear_cfqq_must_dispatch(cfqq);
2654                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2655                 cfq_clear_cfqq_fifo_expire(cfqq);
2656                 cfq_mark_cfqq_slice_new(cfqq);
2657
2658                 cfq_del_timer(cfqd, cfqq);
2659         }
2660
2661         cfqd->active_queue = cfqq;
2662 }
2663
2664 /*
2665  * current cfqq expired its slice (or was too idle), select new one
2666  */
2667 static void
2668 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2669                     bool timed_out)
2670 {
2671         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2672
2673         if (cfq_cfqq_wait_request(cfqq))
2674                 cfq_del_timer(cfqd, cfqq);
2675
2676         cfq_clear_cfqq_wait_request(cfqq);
2677         cfq_clear_cfqq_wait_busy(cfqq);
2678
2679         /*
2680          * If this cfqq is shared between multiple processes, check to
2681          * make sure that those processes are still issuing I/Os within
2682          * the mean seek distance.  If not, it may be time to break the
2683          * queues apart again.
2684          */
2685         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2686                 cfq_mark_cfqq_split_coop(cfqq);
2687
2688         /*
2689          * store what was left of this slice, if the queue idled/timed out
2690          */
2691         if (timed_out) {
2692                 if (cfq_cfqq_slice_new(cfqq))
2693                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2694                 else
2695                         cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2696                 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2697         }
2698
2699         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2700
2701         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2702                 cfq_del_cfqq_rr(cfqd, cfqq);
2703
2704         cfq_resort_rr_list(cfqd, cfqq);
2705
2706         if (cfqq == cfqd->active_queue)
2707                 cfqd->active_queue = NULL;
2708
2709         if (cfqd->active_cic) {
2710                 put_io_context(cfqd->active_cic->icq.ioc);
2711                 cfqd->active_cic = NULL;
2712         }
2713 }
2714
2715 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2716 {
2717         struct cfq_queue *cfqq = cfqd->active_queue;
2718
2719         if (cfqq)
2720                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2721 }
2722
2723 /*
2724  * Get next queue for service. Unless we have a queue preemption,
2725  * we'll simply select the first cfqq in the service tree.
2726  */
2727 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2728 {
2729         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2730                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2731
2732         if (!cfqd->rq_queued)
2733                 return NULL;
2734
2735         /* There is nothing to dispatch */
2736         if (!st)
2737                 return NULL;
2738         if (RB_EMPTY_ROOT(&st->rb))
2739                 return NULL;
2740         return cfq_rb_first(st);
2741 }
2742
2743 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2744 {
2745         struct cfq_group *cfqg;
2746         struct cfq_queue *cfqq;
2747         int i, j;
2748         struct cfq_rb_root *st;
2749
2750         if (!cfqd->rq_queued)
2751                 return NULL;
2752
2753         cfqg = cfq_get_next_cfqg(cfqd);
2754         if (!cfqg)
2755                 return NULL;
2756
2757         for_each_cfqg_st(cfqg, i, j, st) {
2758                 cfqq = cfq_rb_first(st);
2759                 if (cfqq)
2760                         return cfqq;
2761         }
2762         return NULL;
2763 }
2764
2765 /*
2766  * Get and set a new active queue for service.
2767  */
2768 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2769                                               struct cfq_queue *cfqq)
2770 {
2771         if (!cfqq)
2772                 cfqq = cfq_get_next_queue(cfqd);
2773
2774         __cfq_set_active_queue(cfqd, cfqq);
2775         return cfqq;
2776 }
2777
2778 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2779                                           struct request *rq)
2780 {
2781         if (blk_rq_pos(rq) >= cfqd->last_position)
2782                 return blk_rq_pos(rq) - cfqd->last_position;
2783         else
2784                 return cfqd->last_position - blk_rq_pos(rq);
2785 }
2786
2787 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2788                                struct request *rq)
2789 {
2790         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2791 }
2792
2793 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2794                                     struct cfq_queue *cur_cfqq)
2795 {
2796         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2797         struct rb_node *parent, *node;
2798         struct cfq_queue *__cfqq;
2799         sector_t sector = cfqd->last_position;
2800
2801         if (RB_EMPTY_ROOT(root))
2802                 return NULL;
2803
2804         /*
2805          * First, if we find a request starting at the end of the last
2806          * request, choose it.
2807          */
2808         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2809         if (__cfqq)
2810                 return __cfqq;
2811
2812         /*
2813          * If the exact sector wasn't found, the parent of the NULL leaf
2814          * will contain the closest sector.
2815          */
2816         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2817         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2818                 return __cfqq;
2819
2820         if (blk_rq_pos(__cfqq->next_rq) < sector)
2821                 node = rb_next(&__cfqq->p_node);
2822         else
2823                 node = rb_prev(&__cfqq->p_node);
2824         if (!node)
2825                 return NULL;
2826
2827         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2828         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2829                 return __cfqq;
2830
2831         return NULL;
2832 }
2833
2834 /*
2835  * cfqd - obvious
2836  * cur_cfqq - passed in so that we don't decide that the current queue is
2837  *            closely cooperating with itself.
2838  *
2839  * So, basically we're assuming that that cur_cfqq has dispatched at least
2840  * one request, and that cfqd->last_position reflects a position on the disk
2841  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2842  * assumption.
2843  */
2844 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2845                                               struct cfq_queue *cur_cfqq)
2846 {
2847         struct cfq_queue *cfqq;
2848
2849         if (cfq_class_idle(cur_cfqq))
2850                 return NULL;
2851         if (!cfq_cfqq_sync(cur_cfqq))
2852                 return NULL;
2853         if (CFQQ_SEEKY(cur_cfqq))
2854                 return NULL;
2855
2856         /*
2857          * Don't search priority tree if it's the only queue in the group.
2858          */
2859         if (cur_cfqq->cfqg->nr_cfqq == 1)
2860                 return NULL;
2861
2862         /*
2863          * We should notice if some of the queues are cooperating, eg
2864          * working closely on the same area of the disk. In that case,
2865          * we can group them together and don't waste time idling.
2866          */
2867         cfqq = cfqq_close(cfqd, cur_cfqq);
2868         if (!cfqq)
2869                 return NULL;
2870
2871         /* If new queue belongs to different cfq_group, don't choose it */
2872         if (cur_cfqq->cfqg != cfqq->cfqg)
2873                 return NULL;
2874
2875         /*
2876          * It only makes sense to merge sync queues.
2877          */
2878         if (!cfq_cfqq_sync(cfqq))
2879                 return NULL;
2880         if (CFQQ_SEEKY(cfqq))
2881                 return NULL;
2882
2883         /*
2884          * Do not merge queues of different priority classes
2885          */
2886         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2887                 return NULL;
2888
2889         return cfqq;
2890 }
2891
2892 /*
2893  * Determine whether we should enforce idle window for this queue.
2894  */
2895
2896 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2897 {
2898         enum wl_class_t wl_class = cfqq_class(cfqq);
2899         struct cfq_rb_root *st = cfqq->service_tree;
2900
2901         BUG_ON(!st);
2902         BUG_ON(!st->count);
2903
2904         if (!cfqd->cfq_slice_idle)
2905                 return false;
2906
2907         /* We never do for idle class queues. */
2908         if (wl_class == IDLE_WORKLOAD)
2909                 return false;
2910
2911         /* We do for queues that were marked with idle window flag. */
2912         if (cfq_cfqq_idle_window(cfqq) &&
2913            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2914                 return true;
2915
2916         /*
2917          * Otherwise, we do only if they are the last ones
2918          * in their service tree.
2919          */
2920         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2921            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2922                 return true;
2923         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2924         return false;
2925 }
2926
2927 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2928 {
2929         struct cfq_queue *cfqq = cfqd->active_queue;
2930         struct cfq_rb_root *st = cfqq->service_tree;
2931         struct cfq_io_cq *cic;
2932         u64 sl, group_idle = 0;
2933         u64 now = ktime_get_ns();
2934
2935         /*
2936          * SSD device without seek penalty, disable idling. But only do so
2937          * for devices that support queuing, otherwise we still have a problem
2938          * with sync vs async workloads.
2939          */
2940         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2941                 return;
2942
2943         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2944         WARN_ON(cfq_cfqq_slice_new(cfqq));
2945
2946         /*
2947          * idle is disabled, either manually or by past process history
2948          */
2949         if (!cfq_should_idle(cfqd, cfqq)) {
2950                 /* no queue idling. Check for group idling */
2951                 if (cfqd->cfq_group_idle)
2952                         group_idle = cfqd->cfq_group_idle;
2953                 else
2954                         return;
2955         }
2956
2957         /*
2958          * still active requests from this queue, don't idle
2959          */
2960         if (cfqq->dispatched)
2961                 return;
2962
2963         /*
2964          * task has exited, don't wait
2965          */
2966         cic = cfqd->active_cic;
2967         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2968                 return;
2969
2970         /*
2971          * If our average think time is larger than the remaining time
2972          * slice, then don't idle. This avoids overrunning the allotted
2973          * time slice.
2974          */
2975         if (sample_valid(cic->ttime.ttime_samples) &&
2976             (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2977                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2978                              cic->ttime.ttime_mean);
2979                 return;
2980         }
2981
2982         /*
2983          * There are other queues in the group or this is the only group and
2984          * it has too big thinktime, don't do group idle.
2985          */
2986         if (group_idle &&
2987             (cfqq->cfqg->nr_cfqq > 1 ||
2988              cfq_io_thinktime_big(cfqd, &st->ttime, true)))
2989                 return;
2990
2991         cfq_mark_cfqq_wait_request(cfqq);
2992
2993         if (group_idle)
2994                 sl = cfqd->cfq_group_idle;
2995         else
2996                 sl = cfqd->cfq_slice_idle;
2997
2998         hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
2999                       HRTIMER_MODE_REL);
3000         cfqg_stats_set_start_idle_time(cfqq->cfqg);
3001         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
3002                         group_idle ? 1 : 0);
3003 }
3004
3005 /*
3006  * Move request from internal lists to the request queue dispatch list.
3007  */
3008 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
3009 {
3010         struct cfq_data *cfqd = q->elevator->elevator_data;
3011         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3012
3013         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3014
3015         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
3016         cfq_remove_request(rq);
3017         cfqq->dispatched++;
3018         (RQ_CFQG(rq))->dispatched++;
3019         elv_dispatch_sort(q, rq);
3020
3021         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
3022         cfqq->nr_sectors += blk_rq_sectors(rq);
3023 }
3024
3025 /*
3026  * return expired entry, or NULL to just start from scratch in rbtree
3027  */
3028 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3029 {
3030         struct request *rq = NULL;
3031
3032         if (cfq_cfqq_fifo_expire(cfqq))
3033                 return NULL;
3034
3035         cfq_mark_cfqq_fifo_expire(cfqq);
3036
3037         if (list_empty(&cfqq->fifo))
3038                 return NULL;
3039
3040         rq = rq_entry_fifo(cfqq->fifo.next);
3041         if (ktime_get_ns() < rq->fifo_time)
3042                 rq = NULL;
3043
3044         return rq;
3045 }
3046
3047 static inline int
3048 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3049 {
3050         const int base_rq = cfqd->cfq_slice_async_rq;
3051
3052         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3053
3054         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3055 }
3056
3057 /*
3058  * Must be called with the queue_lock held.
3059  */
3060 static int cfqq_process_refs(struct cfq_queue *cfqq)
3061 {
3062         int process_refs, io_refs;
3063
3064         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3065         process_refs = cfqq->ref - io_refs;
3066         BUG_ON(process_refs < 0);
3067         return process_refs;
3068 }
3069
3070 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3071 {
3072         int process_refs, new_process_refs;
3073         struct cfq_queue *__cfqq;
3074
3075         /*
3076          * If there are no process references on the new_cfqq, then it is
3077          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3078          * chain may have dropped their last reference (not just their
3079          * last process reference).
3080          */
3081         if (!cfqq_process_refs(new_cfqq))
3082                 return;
3083
3084         /* Avoid a circular list and skip interim queue merges */
3085         while ((__cfqq = new_cfqq->new_cfqq)) {
3086                 if (__cfqq == cfqq)
3087                         return;
3088                 new_cfqq = __cfqq;
3089         }
3090
3091         process_refs = cfqq_process_refs(cfqq);
3092         new_process_refs = cfqq_process_refs(new_cfqq);
3093         /*
3094          * If the process for the cfqq has gone away, there is no
3095          * sense in merging the queues.
3096          */
3097         if (process_refs == 0 || new_process_refs == 0)
3098                 return;
3099
3100         /*
3101          * Merge in the direction of the lesser amount of work.
3102          */
3103         if (new_process_refs >= process_refs) {
3104                 cfqq->new_cfqq = new_cfqq;
3105                 new_cfqq->ref += process_refs;
3106         } else {
3107                 new_cfqq->new_cfqq = cfqq;
3108                 cfqq->ref += new_process_refs;
3109         }
3110 }
3111
3112 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3113                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3114 {
3115         struct cfq_queue *queue;
3116         int i;
3117         bool key_valid = false;
3118         u64 lowest_key = 0;
3119         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3120
3121         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3122                 /* select the one with lowest rb_key */
3123                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3124                 if (queue &&
3125                     (!key_valid || queue->rb_key < lowest_key)) {
3126                         lowest_key = queue->rb_key;
3127                         cur_best = i;
3128                         key_valid = true;
3129                 }
3130         }
3131
3132         return cur_best;
3133 }
3134
3135 static void
3136 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3137 {
3138         u64 slice;
3139         unsigned count;
3140         struct cfq_rb_root *st;
3141         u64 group_slice;
3142         enum wl_class_t original_class = cfqd->serving_wl_class;
3143         u64 now = ktime_get_ns();
3144
3145         /* Choose next priority. RT > BE > IDLE */
3146         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3147                 cfqd->serving_wl_class = RT_WORKLOAD;
3148         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3149                 cfqd->serving_wl_class = BE_WORKLOAD;
3150         else {
3151                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3152                 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3153                 return;
3154         }
3155
3156         if (original_class != cfqd->serving_wl_class)
3157                 goto new_workload;
3158
3159         /*
3160          * For RT and BE, we have to choose also the type
3161          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3162          * expiration time
3163          */
3164         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3165         count = st->count;
3166
3167         /*
3168          * check workload expiration, and that we still have other queues ready
3169          */
3170         if (count && !(now > cfqd->workload_expires))
3171                 return;
3172
3173 new_workload:
3174         /* otherwise select new workload type */
3175         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3176                                         cfqd->serving_wl_class);
3177         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3178         count = st->count;
3179
3180         /*
3181          * the workload slice is computed as a fraction of target latency
3182          * proportional to the number of queues in that workload, over
3183          * all the queues in the same priority class
3184          */
3185         group_slice = cfq_group_slice(cfqd, cfqg);
3186
3187         slice = div_u64(group_slice * count,
3188                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3189                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3190                                         cfqg)));
3191
3192         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3193                 u64 tmp;
3194
3195                 /*
3196                  * Async queues are currently system wide. Just taking
3197                  * proportion of queues with-in same group will lead to higher
3198                  * async ratio system wide as generally root group is going
3199                  * to have higher weight. A more accurate thing would be to
3200                  * calculate system wide asnc/sync ratio.
3201                  */
3202                 tmp = cfqd->cfq_target_latency *
3203                         cfqg_busy_async_queues(cfqd, cfqg);
3204                 tmp = div_u64(tmp, cfqd->busy_queues);
3205                 slice = min_t(u64, slice, tmp);
3206
3207                 /* async workload slice is scaled down according to
3208                  * the sync/async slice ratio. */
3209                 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3210         } else
3211                 /* sync workload slice is at least 2 * cfq_slice_idle */
3212                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3213
3214         slice = max_t(u64, slice, CFQ_MIN_TT);
3215         cfq_log(cfqd, "workload slice:%llu", slice);
3216         cfqd->workload_expires = now + slice;
3217 }
3218
3219 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3220 {
3221         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3222         struct cfq_group *cfqg;
3223
3224         if (RB_EMPTY_ROOT(&st->rb))
3225                 return NULL;
3226         cfqg = cfq_rb_first_group(st);
3227         update_min_vdisktime(st);
3228         return cfqg;
3229 }
3230
3231 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3232 {
3233         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3234         u64 now = ktime_get_ns();
3235
3236         cfqd->serving_group = cfqg;
3237
3238         /* Restore the workload type data */
3239         if (cfqg->saved_wl_slice) {
3240                 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3241                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3242                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3243         } else
3244                 cfqd->workload_expires = now - 1;
3245
3246         choose_wl_class_and_type(cfqd, cfqg);
3247 }
3248
3249 /*
3250  * Select a queue for service. If we have a current active queue,
3251  * check whether to continue servicing it, or retrieve and set a new one.
3252  */
3253 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3254 {
3255         struct cfq_queue *cfqq, *new_cfqq = NULL;
3256         u64 now = ktime_get_ns();
3257
3258         cfqq = cfqd->active_queue;
3259         if (!cfqq)
3260                 goto new_queue;
3261
3262         if (!cfqd->rq_queued)
3263                 return NULL;
3264
3265         /*
3266          * We were waiting for group to get backlogged. Expire the queue
3267          */
3268         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3269                 goto expire;
3270
3271         /*
3272          * The active queue has run out of time, expire it and select new.
3273          */
3274         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3275                 /*
3276                  * If slice had not expired at the completion of last request
3277                  * we might not have turned on wait_busy flag. Don't expire
3278                  * the queue yet. Allow the group to get backlogged.
3279                  *
3280                  * The very fact that we have used the slice, that means we
3281                  * have been idling all along on this queue and it should be
3282                  * ok to wait for this request to complete.
3283                  */
3284                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3285                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3286                         cfqq = NULL;
3287                         goto keep_queue;
3288                 } else
3289                         goto check_group_idle;
3290         }
3291
3292         /*
3293          * The active queue has requests and isn't expired, allow it to
3294          * dispatch.
3295          */
3296         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3297                 goto keep_queue;
3298
3299         /*
3300          * If another queue has a request waiting within our mean seek
3301          * distance, let it run.  The expire code will check for close
3302          * cooperators and put the close queue at the front of the service
3303          * tree.  If possible, merge the expiring queue with the new cfqq.
3304          */
3305         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3306         if (new_cfqq) {
3307                 if (!cfqq->new_cfqq)
3308                         cfq_setup_merge(cfqq, new_cfqq);
3309                 goto expire;
3310         }
3311
3312         /*
3313          * No requests pending. If the active queue still has requests in
3314          * flight or is idling for a new request, allow either of these
3315          * conditions to happen (or time out) before selecting a new queue.
3316          */
3317         if (hrtimer_active(&cfqd->idle_slice_timer)) {
3318                 cfqq = NULL;
3319                 goto keep_queue;
3320         }
3321
3322         /*
3323          * This is a deep seek queue, but the device is much faster than
3324          * the queue can deliver, don't idle
3325          **/
3326         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3327             (cfq_cfqq_slice_new(cfqq) ||
3328             (cfqq->slice_end - now > now - cfqq->slice_start))) {
3329                 cfq_clear_cfqq_deep(cfqq);
3330                 cfq_clear_cfqq_idle_window(cfqq);
3331         }
3332
3333         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3334                 cfqq = NULL;
3335                 goto keep_queue;
3336         }
3337
3338         /*
3339          * If group idle is enabled and there are requests dispatched from
3340          * this group, wait for requests to complete.
3341          */
3342 check_group_idle:
3343         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3344             cfqq->cfqg->dispatched &&
3345             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3346                 cfqq = NULL;
3347                 goto keep_queue;
3348         }
3349
3350 expire:
3351         cfq_slice_expired(cfqd, 0);
3352 new_queue:
3353         /*
3354          * Current queue expired. Check if we have to switch to a new
3355          * service tree
3356          */
3357         if (!new_cfqq)
3358                 cfq_choose_cfqg(cfqd);
3359
3360         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3361 keep_queue:
3362         return cfqq;
3363 }
3364
3365 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3366 {
3367         int dispatched = 0;
3368
3369         while (cfqq->next_rq) {
3370                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3371                 dispatched++;
3372         }
3373
3374         BUG_ON(!list_empty(&cfqq->fifo));
3375
3376         /* By default cfqq is not expired if it is empty. Do it explicitly */
3377         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3378         return dispatched;
3379 }
3380
3381 /*
3382  * Drain our current requests. Used for barriers and when switching
3383  * io schedulers on-the-fly.
3384  */
3385 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3386 {
3387         struct cfq_queue *cfqq;
3388         int dispatched = 0;
3389
3390         /* Expire the timeslice of the current active queue first */
3391         cfq_slice_expired(cfqd, 0);
3392         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3393                 __cfq_set_active_queue(cfqd, cfqq);
3394                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3395         }
3396
3397         BUG_ON(cfqd->busy_queues);
3398
3399         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3400         return dispatched;
3401 }
3402
3403 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3404         struct cfq_queue *cfqq)
3405 {
3406         u64 now = ktime_get_ns();
3407
3408         /* the queue hasn't finished any request, can't estimate */
3409         if (cfq_cfqq_slice_new(cfqq))
3410                 return true;
3411         if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end)
3412                 return true;
3413
3414         return false;
3415 }
3416
3417 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3418 {
3419         unsigned int max_dispatch;
3420
3421         if (cfq_cfqq_must_dispatch(cfqq))
3422                 return true;
3423
3424         /*
3425          * Drain async requests before we start sync IO
3426          */
3427         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3428                 return false;
3429
3430         /*
3431          * If this is an async queue and we have sync IO in flight, let it wait
3432          */
3433         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3434                 return false;
3435
3436         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3437         if (cfq_class_idle(cfqq))
3438                 max_dispatch = 1;
3439
3440         /*
3441          * Does this cfqq already have too much IO in flight?
3442          */
3443         if (cfqq->dispatched >= max_dispatch) {
3444                 bool promote_sync = false;
3445                 /*
3446                  * idle queue must always only have a single IO in flight
3447                  */
3448                 if (cfq_class_idle(cfqq))
3449                         return false;
3450
3451                 /*
3452                  * If there is only one sync queue
3453                  * we can ignore async queue here and give the sync
3454                  * queue no dispatch limit. The reason is a sync queue can
3455                  * preempt async queue, limiting the sync queue doesn't make
3456                  * sense. This is useful for aiostress test.
3457                  */
3458                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3459                         promote_sync = true;
3460
3461                 /*
3462                  * We have other queues, don't allow more IO from this one
3463                  */
3464                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3465                                 !promote_sync)
3466                         return false;
3467
3468                 /*
3469                  * Sole queue user, no limit
3470                  */
3471                 if (cfqd->busy_queues == 1 || promote_sync)
3472                         max_dispatch = -1;
3473                 else
3474                         /*
3475                          * Normally we start throttling cfqq when cfq_quantum/2
3476                          * requests have been dispatched. But we can drive
3477                          * deeper queue depths at the beginning of slice
3478                          * subjected to upper limit of cfq_quantum.
3479                          * */
3480                         max_dispatch = cfqd->cfq_quantum;
3481         }
3482
3483         /*
3484          * Async queues must wait a bit before being allowed dispatch.
3485          * We also ramp up the dispatch depth gradually for async IO,
3486          * based on the last sync IO we serviced
3487          */
3488         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3489                 u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync;
3490                 unsigned int depth;
3491
3492                 depth = div64_u64(last_sync, cfqd->cfq_slice[1]);
3493                 if (!depth && !cfqq->dispatched)
3494                         depth = 1;
3495                 if (depth < max_dispatch)
3496                         max_dispatch = depth;
3497         }
3498
3499         /*
3500          * If we're below the current max, allow a dispatch
3501          */
3502         return cfqq->dispatched < max_dispatch;
3503 }
3504
3505 /*
3506  * Dispatch a request from cfqq, moving them to the request queue
3507  * dispatch list.
3508  */
3509 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3510 {
3511         struct request *rq;
3512
3513         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3514
3515         rq = cfq_check_fifo(cfqq);
3516         if (rq)
3517                 cfq_mark_cfqq_must_dispatch(cfqq);
3518
3519         if (!cfq_may_dispatch(cfqd, cfqq))
3520                 return false;
3521
3522         /*
3523          * follow expired path, else get first next available
3524          */
3525         if (!rq)
3526                 rq = cfqq->next_rq;
3527         else
3528                 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
3529
3530         /*
3531          * insert request into driver dispatch list
3532          */
3533         cfq_dispatch_insert(cfqd->queue, rq);
3534
3535         if (!cfqd->active_cic) {
3536                 struct cfq_io_cq *cic = RQ_CIC(rq);
3537
3538                 atomic_long_inc(&cic->icq.ioc->refcount);
3539                 cfqd->active_cic = cic;
3540         }
3541
3542         return true;
3543 }
3544
3545 /*
3546  * Find the cfqq that we need to service and move a request from that to the
3547  * dispatch list
3548  */
3549 static int cfq_dispatch_requests(struct request_queue *q, int force)
3550 {
3551         struct cfq_data *cfqd = q->elevator->elevator_data;
3552         struct cfq_queue *cfqq;
3553
3554         if (!cfqd->busy_queues)
3555                 return 0;
3556
3557         if (unlikely(force))
3558                 return cfq_forced_dispatch(cfqd);
3559
3560         cfqq = cfq_select_queue(cfqd);
3561         if (!cfqq)
3562                 return 0;
3563
3564         /*
3565          * Dispatch a request from this cfqq, if it is allowed
3566          */
3567         if (!cfq_dispatch_request(cfqd, cfqq))
3568                 return 0;
3569
3570         cfqq->slice_dispatch++;
3571         cfq_clear_cfqq_must_dispatch(cfqq);
3572
3573         /*
3574          * expire an async queue immediately if it has used up its slice. idle
3575          * queue always expire after 1 dispatch round.
3576          */
3577         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3578             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3579             cfq_class_idle(cfqq))) {
3580                 cfqq->slice_end = ktime_get_ns() + 1;
3581                 cfq_slice_expired(cfqd, 0);
3582         }
3583
3584         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3585         return 1;
3586 }
3587
3588 /*
3589  * task holds one reference to the queue, dropped when task exits. each rq
3590  * in-flight on this queue also holds a reference, dropped when rq is freed.
3591  *
3592  * Each cfq queue took a reference on the parent group. Drop it now.
3593  * queue lock must be held here.
3594  */
3595 static void cfq_put_queue(struct cfq_queue *cfqq)
3596 {
3597         struct cfq_data *cfqd = cfqq->cfqd;
3598         struct cfq_group *cfqg;
3599
3600         BUG_ON(cfqq->ref <= 0);
3601
3602         cfqq->ref--;
3603         if (cfqq->ref)
3604                 return;
3605
3606         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3607         BUG_ON(rb_first(&cfqq->sort_list));
3608         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3609         cfqg = cfqq->cfqg;
3610
3611         if (unlikely(cfqd->active_queue == cfqq)) {
3612                 __cfq_slice_expired(cfqd, cfqq, 0);
3613                 cfq_schedule_dispatch(cfqd);
3614         }
3615
3616         BUG_ON(cfq_cfqq_on_rr(cfqq));
3617         kmem_cache_free(cfq_pool, cfqq);
3618         cfqg_put(cfqg);
3619 }
3620
3621 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3622 {
3623         struct cfq_queue *__cfqq, *next;
3624
3625         /*
3626          * If this queue was scheduled to merge with another queue, be
3627          * sure to drop the reference taken on that queue (and others in
3628          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3629          */
3630         __cfqq = cfqq->new_cfqq;
3631         while (__cfqq) {
3632                 if (__cfqq == cfqq) {
3633                         WARN(1, "cfqq->new_cfqq loop detected\n");
3634                         break;
3635                 }
3636                 next = __cfqq->new_cfqq;
3637                 cfq_put_queue(__cfqq);
3638                 __cfqq = next;
3639         }
3640 }
3641
3642 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3643 {
3644         if (unlikely(cfqq == cfqd->active_queue)) {
3645                 __cfq_slice_expired(cfqd, cfqq, 0);
3646                 cfq_schedule_dispatch(cfqd);
3647         }
3648
3649         cfq_put_cooperator(cfqq);
3650
3651         cfq_put_queue(cfqq);
3652 }
3653
3654 static void cfq_init_icq(struct io_cq *icq)
3655 {
3656         struct cfq_io_cq *cic = icq_to_cic(icq);
3657
3658         cic->ttime.last_end_request = ktime_get_ns();
3659 }
3660
3661 static void cfq_exit_icq(struct io_cq *icq)
3662 {
3663         struct cfq_io_cq *cic = icq_to_cic(icq);
3664         struct cfq_data *cfqd = cic_to_cfqd(cic);
3665
3666         if (cic_to_cfqq(cic, false)) {
3667                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3668                 cic_set_cfqq(cic, NULL, false);
3669         }
3670
3671         if (cic_to_cfqq(cic, true)) {
3672                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3673                 cic_set_cfqq(cic, NULL, true);
3674         }
3675 }
3676
3677 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3678 {
3679         struct task_struct *tsk = current;
3680         int ioprio_class;
3681
3682         if (!cfq_cfqq_prio_changed(cfqq))
3683                 return;
3684
3685         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3686         switch (ioprio_class) {
3687         default:
3688                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3689         case IOPRIO_CLASS_NONE:
3690                 /*
3691                  * no prio set, inherit CPU scheduling settings
3692                  */
3693                 cfqq->ioprio = task_nice_ioprio(tsk);
3694                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3695                 break;
3696         case IOPRIO_CLASS_RT:
3697                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3698                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3699                 break;
3700         case IOPRIO_CLASS_BE:
3701                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3702                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3703                 break;
3704         case IOPRIO_CLASS_IDLE:
3705                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3706                 cfqq->ioprio = 7;
3707                 cfq_clear_cfqq_idle_window(cfqq);
3708                 break;
3709         }
3710
3711         /*
3712          * keep track of original prio settings in case we have to temporarily
3713          * elevate the priority of this queue
3714          */
3715         cfqq->org_ioprio = cfqq->ioprio;
3716         cfqq->org_ioprio_class = cfqq->ioprio_class;
3717         cfq_clear_cfqq_prio_changed(cfqq);
3718 }
3719
3720 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3721 {
3722         int ioprio = cic->icq.ioc->ioprio;
3723         struct cfq_data *cfqd = cic_to_cfqd(cic);
3724         struct cfq_queue *cfqq;
3725
3726         /*
3727          * Check whether ioprio has changed.  The condition may trigger
3728          * spuriously on a newly created cic but there's no harm.
3729          */
3730         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3731                 return;
3732
3733         cfqq = cic_to_cfqq(cic, false);
3734         if (cfqq) {
3735                 cfq_put_queue(cfqq);
3736                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3737                 cic_set_cfqq(cic, cfqq, false);
3738         }
3739
3740         cfqq = cic_to_cfqq(cic, true);
3741         if (cfqq)
3742                 cfq_mark_cfqq_prio_changed(cfqq);
3743
3744         cic->ioprio = ioprio;
3745 }
3746
3747 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3748                           pid_t pid, bool is_sync)
3749 {
3750         RB_CLEAR_NODE(&cfqq->rb_node);
3751         RB_CLEAR_NODE(&cfqq->p_node);
3752         INIT_LIST_HEAD(&cfqq->fifo);
3753
3754         cfqq->ref = 0;
3755         cfqq->cfqd = cfqd;
3756
3757         cfq_mark_cfqq_prio_changed(cfqq);
3758
3759         if (is_sync) {
3760                 if (!cfq_class_idle(cfqq))
3761                         cfq_mark_cfqq_idle_window(cfqq);
3762                 cfq_mark_cfqq_sync(cfqq);
3763         }
3764         cfqq->pid = pid;
3765 }
3766
3767 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3768 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3769 {
3770         struct cfq_data *cfqd = cic_to_cfqd(cic);
3771         struct cfq_queue *cfqq;
3772         uint64_t serial_nr;
3773
3774         rcu_read_lock();
3775         serial_nr = bio_blkcg(bio)->css.serial_nr;
3776         rcu_read_unlock();
3777
3778         /*
3779          * Check whether blkcg has changed.  The condition may trigger
3780          * spuriously on a newly created cic but there's no harm.
3781          */
3782         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3783                 return;
3784
3785         /*
3786          * Drop reference to queues.  New queues will be assigned in new
3787          * group upon arrival of fresh requests.
3788          */
3789         cfqq = cic_to_cfqq(cic, false);
3790         if (cfqq) {
3791                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3792                 cic_set_cfqq(cic, NULL, false);
3793                 cfq_put_queue(cfqq);
3794         }
3795
3796         cfqq = cic_to_cfqq(cic, true);
3797         if (cfqq) {
3798                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3799                 cic_set_cfqq(cic, NULL, true);
3800                 cfq_put_queue(cfqq);
3801         }
3802
3803         cic->blkcg_serial_nr = serial_nr;
3804 }
3805 #else
3806 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3807 {
3808 }
3809 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3810
3811 static struct cfq_queue **
3812 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3813 {
3814         switch (ioprio_class) {
3815         case IOPRIO_CLASS_RT:
3816                 return &cfqg->async_cfqq[0][ioprio];
3817         case IOPRIO_CLASS_NONE:
3818                 ioprio = IOPRIO_NORM;
3819                 /* fall through */
3820         case IOPRIO_CLASS_BE:
3821                 return &cfqg->async_cfqq[1][ioprio];
3822         case IOPRIO_CLASS_IDLE:
3823                 return &cfqg->async_idle_cfqq;
3824         default:
3825                 BUG();
3826         }
3827 }
3828
3829 static struct cfq_queue *
3830 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3831               struct bio *bio)
3832 {
3833         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3834         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3835         struct cfq_queue **async_cfqq = NULL;
3836         struct cfq_queue *cfqq;
3837         struct cfq_group *cfqg;
3838
3839         rcu_read_lock();
3840         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3841         if (!cfqg) {
3842                 cfqq = &cfqd->oom_cfqq;
3843                 goto out;
3844         }
3845
3846         if (!is_sync) {
3847                 if (!ioprio_valid(cic->ioprio)) {
3848                         struct task_struct *tsk = current;
3849                         ioprio = task_nice_ioprio(tsk);
3850                         ioprio_class = task_nice_ioclass(tsk);
3851                 }
3852                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3853                 cfqq = *async_cfqq;
3854                 if (cfqq)
3855                         goto out;
3856         }
3857
3858         cfqq = kmem_cache_alloc_node(cfq_pool,
3859                                      GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
3860                                      cfqd->queue->node);
3861         if (!cfqq) {
3862                 cfqq = &cfqd->oom_cfqq;
3863                 goto out;
3864         }
3865
3866         /* cfq_init_cfqq() assumes cfqq->ioprio_class is initialized. */
3867         cfqq->ioprio_class = IOPRIO_CLASS_NONE;
3868         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3869         cfq_init_prio_data(cfqq, cic);
3870         cfq_link_cfqq_cfqg(cfqq, cfqg);
3871         cfq_log_cfqq(cfqd, cfqq, "alloced");
3872
3873         if (async_cfqq) {
3874                 /* a new async queue is created, pin and remember */
3875                 cfqq->ref++;
3876                 *async_cfqq = cfqq;
3877         }
3878 out:
3879         cfqq->ref++;
3880         rcu_read_unlock();
3881         return cfqq;
3882 }
3883
3884 static void
3885 __cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle)
3886 {
3887         u64 elapsed = ktime_get_ns() - ttime->last_end_request;
3888         elapsed = min(elapsed, 2UL * slice_idle);
3889
3890         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3891         ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
3892         ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3893                                      ttime->ttime_samples);
3894 }
3895
3896 static void
3897 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3898                         struct cfq_io_cq *cic)
3899 {
3900         if (cfq_cfqq_sync(cfqq)) {
3901                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3902                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3903                         cfqd->cfq_slice_idle);
3904         }
3905 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3906         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3907 #endif
3908 }
3909
3910 static void
3911 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3912                        struct request *rq)
3913 {
3914         sector_t sdist = 0;
3915         sector_t n_sec = blk_rq_sectors(rq);
3916         if (cfqq->last_request_pos) {
3917                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3918                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3919                 else
3920                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3921         }
3922
3923         cfqq->seek_history <<= 1;
3924         if (blk_queue_nonrot(cfqd->queue))
3925                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3926         else
3927                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3928 }
3929
3930 static inline bool req_noidle(struct request *req)
3931 {
3932         return req_op(req) == REQ_OP_WRITE &&
3933                 (req->cmd_flags & (REQ_SYNC | REQ_IDLE)) == REQ_SYNC;
3934 }
3935
3936 /*
3937  * Disable idle window if the process thinks too long or seeks so much that
3938  * it doesn't matter
3939  */
3940 static void
3941 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3942                        struct cfq_io_cq *cic)
3943 {
3944         int old_idle, enable_idle;
3945
3946         /*
3947          * Don't idle for async or idle io prio class
3948          */
3949         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3950                 return;
3951
3952         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3953
3954         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3955                 cfq_mark_cfqq_deep(cfqq);
3956
3957         if (cfqq->next_rq && req_noidle(cfqq->next_rq))
3958                 enable_idle = 0;
3959         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3960                  !cfqd->cfq_slice_idle ||
3961                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3962                 enable_idle = 0;
3963         else if (sample_valid(cic->ttime.ttime_samples)) {
3964                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3965                         enable_idle = 0;
3966                 else
3967                         enable_idle = 1;
3968         }
3969
3970         if (old_idle != enable_idle) {
3971                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3972                 if (enable_idle)
3973                         cfq_mark_cfqq_idle_window(cfqq);
3974                 else
3975                         cfq_clear_cfqq_idle_window(cfqq);
3976         }
3977 }
3978
3979 /*
3980  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3981  * no or if we aren't sure, a 1 will cause a preempt.
3982  */
3983 static bool
3984 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3985                    struct request *rq)
3986 {
3987         struct cfq_queue *cfqq;
3988
3989         cfqq = cfqd->active_queue;
3990         if (!cfqq)
3991                 return false;
3992
3993         if (cfq_class_idle(new_cfqq))
3994                 return false;
3995
3996         if (cfq_class_idle(cfqq))
3997                 return true;
3998
3999         /*
4000          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
4001          */
4002         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
4003                 return false;
4004
4005         /*
4006          * if the new request is sync, but the currently running queue is
4007          * not, let the sync request have priority.
4008          */
4009         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
4010                 return true;
4011
4012         /*
4013          * Treat ancestors of current cgroup the same way as current cgroup.
4014          * For anybody else we disallow preemption to guarantee service
4015          * fairness among cgroups.
4016          */
4017         if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
4018                 return false;
4019
4020         if (cfq_slice_used(cfqq))
4021                 return true;
4022
4023         /*
4024          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
4025          */
4026         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
4027                 return true;
4028
4029         WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
4030         /* Allow preemption only if we are idling on sync-noidle tree */
4031         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
4032             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
4033             RB_EMPTY_ROOT(&cfqq->sort_list))
4034                 return true;
4035
4036         /*
4037          * So both queues are sync. Let the new request get disk time if
4038          * it's a metadata request and the current queue is doing regular IO.
4039          */
4040         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
4041                 return true;
4042
4043         /* An idle queue should not be idle now for some reason */
4044         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4045                 return true;
4046
4047         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
4048                 return false;
4049
4050         /*
4051          * if this request is as-good as one we would expect from the
4052          * current cfqq, let it preempt
4053          */
4054         if (cfq_rq_close(cfqd, cfqq, rq))
4055                 return true;
4056
4057         return false;
4058 }
4059
4060 /*
4061  * cfqq preempts the active queue. if we allowed preempt with no slice left,
4062  * let it have half of its nominal slice.
4063  */
4064 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4065 {
4066         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4067
4068         cfq_log_cfqq(cfqd, cfqq, "preempt");
4069         cfq_slice_expired(cfqd, 1);
4070
4071         /*
4072          * workload type is changed, don't save slice, otherwise preempt
4073          * doesn't happen
4074          */
4075         if (old_type != cfqq_type(cfqq))
4076                 cfqq->cfqg->saved_wl_slice = 0;
4077
4078         /*
4079          * Put the new queue at the front of the of the current list,
4080          * so we know that it will be selected next.
4081          */
4082         BUG_ON(!cfq_cfqq_on_rr(cfqq));
4083
4084         cfq_service_tree_add(cfqd, cfqq, 1);
4085
4086         cfqq->slice_end = 0;
4087         cfq_mark_cfqq_slice_new(cfqq);
4088 }
4089
4090 /*
4091  * Called when a new fs request (rq) is added (to cfqq). Check if there's
4092  * something we should do about it
4093  */
4094 static void
4095 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4096                 struct request *rq)
4097 {
4098         struct cfq_io_cq *cic = RQ_CIC(rq);
4099
4100         cfqd->rq_queued++;
4101         if (rq->cmd_flags & REQ_PRIO)
4102                 cfqq->prio_pending++;
4103
4104         cfq_update_io_thinktime(cfqd, cfqq, cic);
4105         cfq_update_io_seektime(cfqd, cfqq, rq);
4106         cfq_update_idle_window(cfqd, cfqq, cic);
4107
4108         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4109
4110         if (cfqq == cfqd->active_queue) {
4111                 /*
4112                  * Remember that we saw a request from this process, but
4113                  * don't start queuing just yet. Otherwise we risk seeing lots
4114                  * of tiny requests, because we disrupt the normal plugging
4115                  * and merging. If the request is already larger than a single
4116                  * page, let it rip immediately. For that case we assume that
4117                  * merging is already done. Ditto for a busy system that
4118                  * has other work pending, don't risk delaying until the
4119                  * idle timer unplug to continue working.
4120                  */
4121                 if (cfq_cfqq_wait_request(cfqq)) {
4122                         if (blk_rq_bytes(rq) > PAGE_SIZE ||
4123                             cfqd->busy_queues > 1) {
4124                                 cfq_del_timer(cfqd, cfqq);
4125                                 cfq_clear_cfqq_wait_request(cfqq);
4126                                 __blk_run_queue(cfqd->queue);
4127                         } else {
4128                                 cfqg_stats_update_idle_time(cfqq->cfqg);
4129                                 cfq_mark_cfqq_must_dispatch(cfqq);
4130                         }
4131                 }
4132         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
4133                 /*
4134                  * not the active queue - expire current slice if it is
4135                  * idle and has expired it's mean thinktime or this new queue
4136                  * has some old slice time left and is of higher priority or
4137                  * this new queue is RT and the current one is BE
4138                  */
4139                 cfq_preempt_queue(cfqd, cfqq);
4140                 __blk_run_queue(cfqd->queue);
4141         }
4142 }
4143
4144 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4145 {
4146         struct cfq_data *cfqd = q->elevator->elevator_data;
4147         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4148
4149         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4150         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4151
4152         rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4153         list_add_tail(&rq->queuelist, &cfqq->fifo);
4154         cfq_add_rq_rb(rq);
4155         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4156                                  rq->cmd_flags);
4157         cfq_rq_enqueued(cfqd, cfqq, rq);
4158 }
4159
4160 /*
4161  * Update hw_tag based on peak queue depth over 50 samples under
4162  * sufficient load.
4163  */
4164 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4165 {
4166         struct cfq_queue *cfqq = cfqd->active_queue;
4167
4168         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4169                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4170
4171         if (cfqd->hw_tag == 1)
4172                 return;
4173
4174         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4175             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4176                 return;
4177
4178         /*
4179          * If active queue hasn't enough requests and can idle, cfq might not
4180          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4181          * case
4182          */
4183         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4184             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4185             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4186                 return;
4187
4188         if (cfqd->hw_tag_samples++ < 50)
4189                 return;
4190
4191         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4192                 cfqd->hw_tag = 1;
4193         else
4194                 cfqd->hw_tag = 0;
4195 }
4196
4197 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4198 {
4199         struct cfq_io_cq *cic = cfqd->active_cic;
4200         u64 now = ktime_get_ns();
4201
4202         /* If the queue already has requests, don't wait */
4203         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4204                 return false;
4205
4206         /* If there are other queues in the group, don't wait */
4207         if (cfqq->cfqg->nr_cfqq > 1)
4208                 return false;
4209
4210         /* the only queue in the group, but think time is big */
4211         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4212                 return false;
4213
4214         if (cfq_slice_used(cfqq))
4215                 return true;
4216
4217         /* if slice left is less than think time, wait busy */
4218         if (cic && sample_valid(cic->ttime.ttime_samples)
4219             && (cfqq->slice_end - now < cic->ttime.ttime_mean))
4220                 return true;
4221
4222         /*
4223          * If think times is less than a jiffy than ttime_mean=0 and above
4224          * will not be true. It might happen that slice has not expired yet
4225          * but will expire soon (4-5 ns) during select_queue(). To cover the
4226          * case where think time is less than a jiffy, mark the queue wait
4227          * busy if only 1 jiffy is left in the slice.
4228          */
4229         if (cfqq->slice_end - now <= jiffies_to_nsecs(1))
4230                 return true;
4231
4232         return false;
4233 }
4234
4235 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4236 {
4237         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4238         struct cfq_data *cfqd = cfqq->cfqd;
4239         const int sync = rq_is_sync(rq);
4240         u64 now = ktime_get_ns();
4241
4242         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", req_noidle(rq));
4243
4244         cfq_update_hw_tag(cfqd);
4245
4246         WARN_ON(!cfqd->rq_in_driver);
4247         WARN_ON(!cfqq->dispatched);
4248         cfqd->rq_in_driver--;
4249         cfqq->dispatched--;
4250         (RQ_CFQG(rq))->dispatched--;
4251         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4252                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4253
4254         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4255
4256         if (sync) {
4257                 struct cfq_rb_root *st;
4258
4259                 RQ_CIC(rq)->ttime.last_end_request = now;
4260
4261                 if (cfq_cfqq_on_rr(cfqq))
4262                         st = cfqq->service_tree;
4263                 else
4264                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4265                                         cfqq_type(cfqq));
4266
4267                 st->ttime.last_end_request = now;
4268                 /*
4269                  * We have to do this check in jiffies since start_time is in
4270                  * jiffies and it is not trivial to convert to ns. If
4271                  * cfq_fifo_expire[1] ever comes close to 1 jiffie, this test
4272                  * will become problematic but so far we are fine (the default
4273                  * is 128 ms).
4274                  */
4275                 if (!time_after(rq->start_time +
4276                                   nsecs_to_jiffies(cfqd->cfq_fifo_expire[1]),
4277                                 jiffies))
4278                         cfqd->last_delayed_sync = now;
4279         }
4280
4281 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4282         cfqq->cfqg->ttime.last_end_request = now;
4283 #endif
4284
4285         /*
4286          * If this is the active queue, check if it needs to be expired,
4287          * or if we want to idle in case it has no pending requests.
4288          */
4289         if (cfqd->active_queue == cfqq) {
4290                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4291
4292                 if (cfq_cfqq_slice_new(cfqq)) {
4293                         cfq_set_prio_slice(cfqd, cfqq);
4294                         cfq_clear_cfqq_slice_new(cfqq);
4295                 }
4296
4297                 /*
4298                  * Should we wait for next request to come in before we expire
4299                  * the queue.
4300                  */
4301                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4302                         u64 extend_sl = cfqd->cfq_slice_idle;
4303                         if (!cfqd->cfq_slice_idle)
4304                                 extend_sl = cfqd->cfq_group_idle;
4305                         cfqq->slice_end = now + extend_sl;
4306                         cfq_mark_cfqq_wait_busy(cfqq);
4307                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4308                 }
4309
4310                 /*
4311                  * Idling is not enabled on:
4312                  * - expired queues
4313                  * - idle-priority queues
4314                  * - async queues
4315                  * - queues with still some requests queued
4316                  * - when there is a close cooperator
4317                  */
4318                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4319                         cfq_slice_expired(cfqd, 1);
4320                 else if (sync && cfqq_empty &&
4321                          !cfq_close_cooperator(cfqd, cfqq)) {
4322                         cfq_arm_slice_timer(cfqd);
4323                 }
4324         }
4325
4326         if (!cfqd->rq_in_driver)
4327                 cfq_schedule_dispatch(cfqd);
4328 }
4329
4330 static void cfqq_boost_on_prio(struct cfq_queue *cfqq, unsigned int op)
4331 {
4332         /*
4333          * If REQ_PRIO is set, boost class and prio level, if it's below
4334          * BE/NORM. If prio is not set, restore the potentially boosted
4335          * class/prio level.
4336          */
4337         if (!(op & REQ_PRIO)) {
4338                 cfqq->ioprio_class = cfqq->org_ioprio_class;
4339                 cfqq->ioprio = cfqq->org_ioprio;
4340         } else {
4341                 if (cfq_class_idle(cfqq))
4342                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
4343                 if (cfqq->ioprio > IOPRIO_NORM)
4344                         cfqq->ioprio = IOPRIO_NORM;
4345         }
4346 }
4347
4348 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4349 {
4350         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4351                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4352                 return ELV_MQUEUE_MUST;
4353         }
4354
4355         return ELV_MQUEUE_MAY;
4356 }
4357
4358 static int cfq_may_queue(struct request_queue *q, unsigned int op)
4359 {
4360         struct cfq_data *cfqd = q->elevator->elevator_data;
4361         struct task_struct *tsk = current;
4362         struct cfq_io_cq *cic;
4363         struct cfq_queue *cfqq;
4364
4365         /*
4366          * don't force setup of a queue from here, as a call to may_queue
4367          * does not necessarily imply that a request actually will be queued.
4368          * so just lookup a possibly existing queue, or return 'may queue'
4369          * if that fails
4370          */
4371         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4372         if (!cic)
4373                 return ELV_MQUEUE_MAY;
4374
4375         cfqq = cic_to_cfqq(cic, op_is_sync(op));
4376         if (cfqq) {
4377                 cfq_init_prio_data(cfqq, cic);
4378                 cfqq_boost_on_prio(cfqq, op);
4379
4380                 return __cfq_may_queue(cfqq);
4381         }
4382
4383         return ELV_MQUEUE_MAY;
4384 }
4385
4386 /*
4387  * queue lock held here
4388  */
4389 static void cfq_put_request(struct request *rq)
4390 {
4391         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4392
4393         if (cfqq) {
4394                 const int rw = rq_data_dir(rq);
4395
4396                 BUG_ON(!cfqq->allocated[rw]);
4397                 cfqq->allocated[rw]--;
4398
4399                 /* Put down rq reference on cfqg */
4400                 cfqg_put(RQ_CFQG(rq));
4401                 rq->elv.priv[0] = NULL;
4402                 rq->elv.priv[1] = NULL;
4403
4404                 cfq_put_queue(cfqq);
4405         }
4406 }
4407
4408 static struct cfq_queue *
4409 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4410                 struct cfq_queue *cfqq)
4411 {
4412         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4413         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4414         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4415         cfq_put_queue(cfqq);
4416         return cic_to_cfqq(cic, 1);
4417 }
4418
4419 /*
4420  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4421  * was the last process referring to said cfqq.
4422  */
4423 static struct cfq_queue *
4424 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4425 {
4426         if (cfqq_process_refs(cfqq) == 1) {
4427                 cfqq->pid = current->pid;
4428                 cfq_clear_cfqq_coop(cfqq);
4429                 cfq_clear_cfqq_split_coop(cfqq);
4430                 return cfqq;
4431         }
4432
4433         cic_set_cfqq(cic, NULL, 1);
4434
4435         cfq_put_cooperator(cfqq);
4436
4437         cfq_put_queue(cfqq);
4438         return NULL;
4439 }
4440 /*
4441  * Allocate cfq data structures associated with this request.
4442  */
4443 static int
4444 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4445                 gfp_t gfp_mask)
4446 {
4447         struct cfq_data *cfqd = q->elevator->elevator_data;
4448         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4449         const int rw = rq_data_dir(rq);
4450         const bool is_sync = rq_is_sync(rq);
4451         struct cfq_queue *cfqq;
4452
4453         spin_lock_irq(q->queue_lock);
4454
4455         check_ioprio_changed(cic, bio);
4456         check_blkcg_changed(cic, bio);
4457 new_queue:
4458         cfqq = cic_to_cfqq(cic, is_sync);
4459         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4460                 if (cfqq)
4461                         cfq_put_queue(cfqq);
4462                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4463                 cic_set_cfqq(cic, cfqq, is_sync);
4464         } else {
4465                 /*
4466                  * If the queue was seeky for too long, break it apart.
4467                  */
4468                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4469                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4470                         cfqq = split_cfqq(cic, cfqq);
4471                         if (!cfqq)
4472                                 goto new_queue;
4473                 }
4474
4475                 /*
4476                  * Check to see if this queue is scheduled to merge with
4477                  * another, closely cooperating queue.  The merging of
4478                  * queues happens here as it must be done in process context.
4479                  * The reference on new_cfqq was taken in merge_cfqqs.
4480                  */
4481                 if (cfqq->new_cfqq)
4482                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4483         }
4484
4485         cfqq->allocated[rw]++;
4486
4487         cfqq->ref++;
4488         cfqg_get(cfqq->cfqg);
4489         rq->elv.priv[0] = cfqq;
4490         rq->elv.priv[1] = cfqq->cfqg;
4491         spin_unlock_irq(q->queue_lock);
4492
4493         return 0;
4494 }
4495
4496 static void cfq_kick_queue(struct work_struct *work)
4497 {
4498         struct cfq_data *cfqd =
4499                 container_of(work, struct cfq_data, unplug_work);
4500         struct request_queue *q = cfqd->queue;
4501
4502         spin_lock_irq(q->queue_lock);
4503         __blk_run_queue(cfqd->queue);
4504         spin_unlock_irq(q->queue_lock);
4505 }
4506
4507 /*
4508  * Timer running if the active_queue is currently idling inside its time slice
4509  */
4510 static enum hrtimer_restart cfq_idle_slice_timer(struct hrtimer *timer)
4511 {
4512         struct cfq_data *cfqd = container_of(timer, struct cfq_data,
4513                                              idle_slice_timer);
4514         struct cfq_queue *cfqq;
4515         unsigned long flags;
4516         int timed_out = 1;
4517
4518         cfq_log(cfqd, "idle timer fired");
4519
4520         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4521
4522         cfqq = cfqd->active_queue;
4523         if (cfqq) {
4524                 timed_out = 0;
4525
4526                 /*
4527                  * We saw a request before the queue expired, let it through
4528                  */
4529                 if (cfq_cfqq_must_dispatch(cfqq))
4530                         goto out_kick;
4531
4532                 /*
4533                  * expired
4534                  */
4535                 if (cfq_slice_used(cfqq))
4536                         goto expire;
4537
4538                 /*
4539                  * only expire and reinvoke request handler, if there are
4540                  * other queues with pending requests
4541                  */
4542                 if (!cfqd->busy_queues)
4543                         goto out_cont;
4544
4545                 /*
4546                  * not expired and it has a request pending, let it dispatch
4547                  */
4548                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4549                         goto out_kick;
4550
4551                 /*
4552                  * Queue depth flag is reset only when the idle didn't succeed
4553                  */
4554                 cfq_clear_cfqq_deep(cfqq);
4555         }
4556 expire:
4557         cfq_slice_expired(cfqd, timed_out);
4558 out_kick:
4559         cfq_schedule_dispatch(cfqd);
4560 out_cont:
4561         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4562         return HRTIMER_NORESTART;
4563 }
4564
4565 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4566 {
4567         hrtimer_cancel(&cfqd->idle_slice_timer);
4568         cancel_work_sync(&cfqd->unplug_work);
4569 }
4570
4571 static void cfq_exit_queue(struct elevator_queue *e)
4572 {
4573         struct cfq_data *cfqd = e->elevator_data;
4574         struct request_queue *q = cfqd->queue;
4575
4576         cfq_shutdown_timer_wq(cfqd);
4577
4578         spin_lock_irq(q->queue_lock);
4579
4580         if (cfqd->active_queue)
4581                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4582
4583         spin_unlock_irq(q->queue_lock);
4584
4585         cfq_shutdown_timer_wq(cfqd);
4586
4587 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4588         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4589 #else
4590         kfree(cfqd->root_group);
4591 #endif
4592         kfree(cfqd);
4593 }
4594
4595 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4596 {
4597         struct cfq_data *cfqd;
4598         struct blkcg_gq *blkg __maybe_unused;
4599         int i, ret;
4600         struct elevator_queue *eq;
4601
4602         eq = elevator_alloc(q, e);
4603         if (!eq)
4604                 return -ENOMEM;
4605
4606         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4607         if (!cfqd) {
4608                 kobject_put(&eq->kobj);
4609                 return -ENOMEM;
4610         }
4611         eq->elevator_data = cfqd;
4612
4613         cfqd->queue = q;
4614         spin_lock_irq(q->queue_lock);
4615         q->elevator = eq;
4616         spin_unlock_irq(q->queue_lock);
4617
4618         /* Init root service tree */
4619         cfqd->grp_service_tree = CFQ_RB_ROOT;
4620
4621         /* Init root group and prefer root group over other groups by default */
4622 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4623         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4624         if (ret)
4625                 goto out_free;
4626
4627         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4628 #else
4629         ret = -ENOMEM;
4630         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4631                                         GFP_KERNEL, cfqd->queue->node);
4632         if (!cfqd->root_group)
4633                 goto out_free;
4634
4635         cfq_init_cfqg_base(cfqd->root_group);
4636         cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4637         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4638 #endif
4639
4640         /*
4641          * Not strictly needed (since RB_ROOT just clears the node and we
4642          * zeroed cfqd on alloc), but better be safe in case someone decides
4643          * to add magic to the rb code
4644          */
4645         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4646                 cfqd->prio_trees[i] = RB_ROOT;
4647
4648         /*
4649          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4650          * Grab a permanent reference to it, so that the normal code flow
4651          * will not attempt to free it.  oom_cfqq is linked to root_group
4652          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4653          * the reference from linking right away.
4654          */
4655         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4656         cfqd->oom_cfqq.ref++;
4657
4658         spin_lock_irq(q->queue_lock);
4659         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4660         cfqg_put(cfqd->root_group);
4661         spin_unlock_irq(q->queue_lock);
4662
4663         hrtimer_init(&cfqd->idle_slice_timer, CLOCK_MONOTONIC,
4664                      HRTIMER_MODE_REL);
4665         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4666
4667         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4668
4669         cfqd->cfq_quantum = cfq_quantum;
4670         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4671         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4672         cfqd->cfq_back_max = cfq_back_max;
4673         cfqd->cfq_back_penalty = cfq_back_penalty;
4674         cfqd->cfq_slice[0] = cfq_slice_async;
4675         cfqd->cfq_slice[1] = cfq_slice_sync;
4676         cfqd->cfq_target_latency = cfq_target_latency;
4677         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4678         cfqd->cfq_slice_idle = cfq_slice_idle;
4679         cfqd->cfq_group_idle = cfq_group_idle;
4680         cfqd->cfq_latency = 1;
4681         cfqd->hw_tag = -1;
4682         /*
4683          * we optimistically start assuming sync ops weren't delayed in last
4684          * second, in order to have larger depth for async operations.
4685          */
4686         cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC;
4687         return 0;
4688
4689 out_free:
4690         kfree(cfqd);
4691         kobject_put(&eq->kobj);
4692         return ret;
4693 }
4694
4695 static void cfq_registered_queue(struct request_queue *q)
4696 {
4697         struct elevator_queue *e = q->elevator;
4698         struct cfq_data *cfqd = e->elevator_data;
4699
4700         /*
4701          * Default to IOPS mode with no idling for SSDs
4702          */
4703         if (blk_queue_nonrot(q))
4704                 cfqd->cfq_slice_idle = 0;
4705         wbt_disable_default(q);
4706 }
4707
4708 /*
4709  * sysfs parts below -->
4710  */
4711 static ssize_t
4712 cfq_var_show(unsigned int var, char *page)
4713 {
4714         return sprintf(page, "%u\n", var);
4715 }
4716
4717 static ssize_t
4718 cfq_var_store(unsigned int *var, const char *page, size_t count)
4719 {
4720         char *p = (char *) page;
4721
4722         *var = simple_strtoul(p, &p, 10);
4723         return count;
4724 }
4725
4726 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4727 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4728 {                                                                       \
4729         struct cfq_data *cfqd = e->elevator_data;                       \
4730         u64 __data = __VAR;                                             \
4731         if (__CONV)                                                     \
4732                 __data = div_u64(__data, NSEC_PER_MSEC);                        \
4733         return cfq_var_show(__data, (page));                            \
4734 }
4735 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4736 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4737 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4738 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4739 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4740 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4741 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4742 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4743 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4744 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4745 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4746 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4747 #undef SHOW_FUNCTION
4748
4749 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)                               \
4750 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4751 {                                                                       \
4752         struct cfq_data *cfqd = e->elevator_data;                       \
4753         u64 __data = __VAR;                                             \
4754         __data = div_u64(__data, NSEC_PER_USEC);                        \
4755         return cfq_var_show(__data, (page));                            \
4756 }
4757 USEC_SHOW_FUNCTION(cfq_slice_idle_us_show, cfqd->cfq_slice_idle);
4758 USEC_SHOW_FUNCTION(cfq_group_idle_us_show, cfqd->cfq_group_idle);
4759 USEC_SHOW_FUNCTION(cfq_slice_sync_us_show, cfqd->cfq_slice[1]);
4760 USEC_SHOW_FUNCTION(cfq_slice_async_us_show, cfqd->cfq_slice[0]);
4761 USEC_SHOW_FUNCTION(cfq_target_latency_us_show, cfqd->cfq_target_latency);
4762 #undef USEC_SHOW_FUNCTION
4763
4764 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4765 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4766 {                                                                       \
4767         struct cfq_data *cfqd = e->elevator_data;                       \
4768         unsigned int __data;                                            \
4769         int ret = cfq_var_store(&__data, (page), count);                \
4770         if (__data < (MIN))                                             \
4771                 __data = (MIN);                                         \
4772         else if (__data > (MAX))                                        \
4773                 __data = (MAX);                                         \
4774         if (__CONV)                                                     \
4775                 *(__PTR) = (u64)__data * NSEC_PER_MSEC;                 \
4776         else                                                            \
4777                 *(__PTR) = __data;                                      \
4778         return ret;                                                     \
4779 }
4780 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4781 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4782                 UINT_MAX, 1);
4783 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4784                 UINT_MAX, 1);
4785 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4786 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4787                 UINT_MAX, 0);
4788 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4789 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4790 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4791 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4792 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4793                 UINT_MAX, 0);
4794 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4795 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4796 #undef STORE_FUNCTION
4797
4798 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)                    \
4799 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4800 {                                                                       \
4801         struct cfq_data *cfqd = e->elevator_data;                       \
4802         unsigned int __data;                                            \
4803         int ret = cfq_var_store(&__data, (page), count);                \
4804         if (__data < (MIN))                                             \
4805                 __data = (MIN);                                         \
4806         else if (__data > (MAX))                                        \
4807                 __data = (MAX);                                         \
4808         *(__PTR) = (u64)__data * NSEC_PER_USEC;                         \
4809         return ret;                                                     \
4810 }
4811 USEC_STORE_FUNCTION(cfq_slice_idle_us_store, &cfqd->cfq_slice_idle, 0, UINT_MAX);
4812 USEC_STORE_FUNCTION(cfq_group_idle_us_store, &cfqd->cfq_group_idle, 0, UINT_MAX);
4813 USEC_STORE_FUNCTION(cfq_slice_sync_us_store, &cfqd->cfq_slice[1], 1, UINT_MAX);
4814 USEC_STORE_FUNCTION(cfq_slice_async_us_store, &cfqd->cfq_slice[0], 1, UINT_MAX);
4815 USEC_STORE_FUNCTION(cfq_target_latency_us_store, &cfqd->cfq_target_latency, 1, UINT_MAX);
4816 #undef USEC_STORE_FUNCTION
4817
4818 #define CFQ_ATTR(name) \
4819         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4820
4821 static struct elv_fs_entry cfq_attrs[] = {
4822         CFQ_ATTR(quantum),
4823         CFQ_ATTR(fifo_expire_sync),
4824         CFQ_ATTR(fifo_expire_async),
4825         CFQ_ATTR(back_seek_max),
4826         CFQ_ATTR(back_seek_penalty),
4827         CFQ_ATTR(slice_sync),
4828         CFQ_ATTR(slice_sync_us),
4829         CFQ_ATTR(slice_async),
4830         CFQ_ATTR(slice_async_us),
4831         CFQ_ATTR(slice_async_rq),
4832         CFQ_ATTR(slice_idle),
4833         CFQ_ATTR(slice_idle_us),
4834         CFQ_ATTR(group_idle),
4835         CFQ_ATTR(group_idle_us),
4836         CFQ_ATTR(low_latency),
4837         CFQ_ATTR(target_latency),
4838         CFQ_ATTR(target_latency_us),
4839         __ATTR_NULL
4840 };
4841
4842 static struct elevator_type iosched_cfq = {
4843         .ops.sq = {
4844                 .elevator_merge_fn =            cfq_merge,
4845                 .elevator_merged_fn =           cfq_merged_request,
4846                 .elevator_merge_req_fn =        cfq_merged_requests,
4847                 .elevator_allow_bio_merge_fn =  cfq_allow_bio_merge,
4848                 .elevator_allow_rq_merge_fn =   cfq_allow_rq_merge,
4849                 .elevator_bio_merged_fn =       cfq_bio_merged,
4850                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4851                 .elevator_add_req_fn =          cfq_insert_request,
4852                 .elevator_activate_req_fn =     cfq_activate_request,
4853                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4854                 .elevator_completed_req_fn =    cfq_completed_request,
4855                 .elevator_former_req_fn =       elv_rb_former_request,
4856                 .elevator_latter_req_fn =       elv_rb_latter_request,
4857                 .elevator_init_icq_fn =         cfq_init_icq,
4858                 .elevator_exit_icq_fn =         cfq_exit_icq,
4859                 .elevator_set_req_fn =          cfq_set_request,
4860                 .elevator_put_req_fn =          cfq_put_request,
4861                 .elevator_may_queue_fn =        cfq_may_queue,
4862                 .elevator_init_fn =             cfq_init_queue,
4863                 .elevator_exit_fn =             cfq_exit_queue,
4864                 .elevator_registered_fn =       cfq_registered_queue,
4865         },
4866         .icq_size       =       sizeof(struct cfq_io_cq),
4867         .icq_align      =       __alignof__(struct cfq_io_cq),
4868         .elevator_attrs =       cfq_attrs,
4869         .elevator_name  =       "cfq",
4870         .elevator_owner =       THIS_MODULE,
4871 };
4872
4873 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4874 static struct blkcg_policy blkcg_policy_cfq = {
4875         .dfl_cftypes            = cfq_blkcg_files,
4876         .legacy_cftypes         = cfq_blkcg_legacy_files,
4877
4878         .cpd_alloc_fn           = cfq_cpd_alloc,
4879         .cpd_init_fn            = cfq_cpd_init,
4880         .cpd_free_fn            = cfq_cpd_free,
4881         .cpd_bind_fn            = cfq_cpd_bind,
4882
4883         .pd_alloc_fn            = cfq_pd_alloc,
4884         .pd_init_fn             = cfq_pd_init,
4885         .pd_offline_fn          = cfq_pd_offline,
4886         .pd_free_fn             = cfq_pd_free,
4887         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4888 };
4889 #endif
4890
4891 static int __init cfq_init(void)
4892 {
4893         int ret;
4894
4895 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4896         ret = blkcg_policy_register(&blkcg_policy_cfq);
4897         if (ret)
4898                 return ret;
4899 #else
4900         cfq_group_idle = 0;
4901 #endif
4902
4903         ret = -ENOMEM;
4904         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4905         if (!cfq_pool)
4906                 goto err_pol_unreg;
4907
4908         ret = elv_register(&iosched_cfq);
4909         if (ret)
4910                 goto err_free_pool;
4911
4912         return 0;
4913
4914 err_free_pool:
4915         kmem_cache_destroy(cfq_pool);
4916 err_pol_unreg:
4917 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4918         blkcg_policy_unregister(&blkcg_policy_cfq);
4919 #endif
4920         return ret;
4921 }
4922
4923 static void __exit cfq_exit(void)
4924 {
4925 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4926         blkcg_policy_unregister(&blkcg_policy_cfq);
4927 #endif
4928         elv_unregister(&iosched_cfq);
4929         kmem_cache_destroy(cfq_pool);
4930 }
4931
4932 module_init(cfq_init);
4933 module_exit(cfq_exit);
4934
4935 MODULE_AUTHOR("Jens Axboe");
4936 MODULE_LICENSE("GPL");
4937 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");