block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/ktime.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16 #include <linux/blktrace_api.h>
  17 #include <linux/blk-cgroup.h>
  18 #include "blk.h"
  19
  20 /*
  21  * tunables
  22  */
  23 /* max queue in one round of service */
  24 static const int cfq_quantum = 8;
  25 static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
  26 /* maximum backwards seek, in KiB */
  27 static const int cfq_back_max = 16 * 1024;
  28 /* penalty of a backwards seek */
  29 static const int cfq_back_penalty = 2;
  30 static const u64 cfq_slice_sync = NSEC_PER_SEC / 10;
  31 static u64 cfq_slice_async = NSEC_PER_SEC / 25;
  32 static const int cfq_slice_async_rq = 2;
  33 static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
  34 static u64 cfq_group_idle = NSEC_PER_SEC / 125;
  35 static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
  36 static const int cfq_hist_divisor = 4;
  37
  38 /*
  39  * offset from end of service tree
  40  */
  41 #define CFQ_IDLE_DELAY          (NSEC_PER_SEC / 5)
  42
  43 /*
  44  * below this threshold, we consider thinktime immediate
  45  */
  46 #define CFQ_MIN_TT              (2 * NSEC_PER_SEC / HZ)
  47
  48 #define CFQ_SLICE_SCALE         (5)
  49 #define CFQ_HW_QUEUE_MIN        (5)
  50 #define CFQ_SERVICE_SHIFT       12
  51
  52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  56
  57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  60
  61 static struct kmem_cache *cfq_pool;
  62
  63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  69
  70 /* blkio-related constants */
  71 #define CFQ_WEIGHT_LEGACY_MIN   10
  72 #define CFQ_WEIGHT_LEGACY_DFL   500
  73 #define CFQ_WEIGHT_LEGACY_MAX   1000
  74
  75 struct cfq_ttime {
  76         u64 last_end_request;
  77
  78         u64 ttime_total;
  79         u64 ttime_mean;
  80         unsigned long ttime_samples;
  81 };
  82
  83 /*
  84  * Most of our rbtree usage is for sorting with min extraction, so
  85  * if we cache the leftmost node we don't have to walk down the tree
  86  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  87  * move this into the elevator for the rq sorting as well.
  88  */
  89 struct cfq_rb_root {
  90         struct rb_root rb;
  91         struct rb_node *left;
  92         unsigned count;
  93         u64 min_vdisktime;
  94         struct cfq_ttime ttime;
  95 };
  96 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  97                         .ttime = {.last_end_request = ktime_get_ns(),},}
  98
  99 /*
 100  * Per process-grouping structure
 101  */
 102 struct cfq_queue {
 103         /* reference count */
 104         int ref;
 105         /* various state flags, see below */
 106         unsigned int flags;
 107         /* parent cfq_data */
 108         struct cfq_data *cfqd;
 109         /* service_tree member */
 110         struct rb_node rb_node;
 111         /* service_tree key */
 112         u64 rb_key;
 113         /* prio tree member */
 114         struct rb_node p_node;
 115         /* prio tree root we belong to, if any */
 116         struct rb_root *p_root;
 117         /* sorted list of pending requests */
 118         struct rb_root sort_list;
 119         /* if fifo isn't expired, next request to serve */
 120         struct request *next_rq;
 121         /* requests queued in sort_list */
 122         int queued[2];
 123         /* currently allocated requests */
 124         int allocated[2];
 125         /* fifo list of requests in sort_list */
 126         struct list_head fifo;
 127
 128         /* time when queue got scheduled in to dispatch first request. */
 129         u64 dispatch_start;
 130         u64 allocated_slice;
 131         u64 slice_dispatch;
 132         /* time when first request from queue completed and slice started. */
 133         u64 slice_start;
 134         u64 slice_end;
 135         s64 slice_resid;
 136
 137         /* pending priority requests */
 138         int prio_pending;
 139         /* number of requests that are on the dispatch list or inside driver */
 140         int dispatched;
 141
 142         /* io prio of this group */
 143         unsigned short ioprio, org_ioprio;
 144         unsigned short ioprio_class, org_ioprio_class;
 145
 146         pid_t pid;
 147
 148         u32 seek_history;
 149         sector_t last_request_pos;
 150
 151         struct cfq_rb_root *service_tree;
 152         struct cfq_queue *new_cfqq;
 153         struct cfq_group *cfqg;
 154         /* Number of sectors dispatched from queue in single dispatch round */
 155         unsigned long nr_sectors;
 156 };
 157
 158 /*
 159  * First index in the service_trees.
 160  * IDLE is handled separately, so it has negative index
 161  */
 162 enum wl_class_t {
 163         BE_WORKLOAD = 0,
 164         RT_WORKLOAD = 1,
 165         IDLE_WORKLOAD = 2,
 166         CFQ_PRIO_NR,
 167 };
 168
 169 /*
 170  * Second index in the service_trees.
 171  */
 172 enum wl_type_t {
 173         ASYNC_WORKLOAD = 0,
 174         SYNC_NOIDLE_WORKLOAD = 1,
 175         SYNC_WORKLOAD = 2
 176 };
 177
 178 struct cfqg_stats {
 179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 180         /* number of ios merged */
 181         struct blkg_rwstat              merged;
 182         /* total time spent on device in ns, may not be accurate w/ queueing */
 183         struct blkg_rwstat              service_time;
 184         /* total time spent waiting in scheduler queue in ns */
 185         struct blkg_rwstat              wait_time;
 186         /* number of IOs queued up */
 187         struct blkg_rwstat              queued;
 188         /* total disk time and nr sectors dispatched by this group */
 189         struct blkg_stat                time;
 190 #ifdef CONFIG_DEBUG_BLK_CGROUP
 191         /* time not charged to this cgroup */
 192         struct blkg_stat                unaccounted_time;
 193         /* sum of number of ios queued across all samples */
 194         struct blkg_stat                avg_queue_size_sum;
 195         /* count of samples taken for average */
 196         struct blkg_stat                avg_queue_size_samples;
 197         /* how many times this group has been removed from service tree */
 198         struct blkg_stat                dequeue;
 199         /* total time spent waiting for it to be assigned a timeslice. */
 200         struct blkg_stat                group_wait_time;
 201         /* time spent idling for this blkcg_gq */
 202         struct blkg_stat                idle_time;
 203         /* total time with empty current active q with other requests queued */
 204         struct blkg_stat                empty_time;
 205         /* fields after this shouldn't be cleared on stat reset */
 206         uint64_t                        start_group_wait_time;
 207         uint64_t                        start_idle_time;
 208         uint64_t                        start_empty_time;
 209         uint16_t                        flags;
 210 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 211 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 212 };
 213
 214 /* Per-cgroup data */
 215 struct cfq_group_data {
 216         /* must be the first member */
 217         struct blkcg_policy_data cpd;
 218
 219         unsigned int weight;
 220         unsigned int leaf_weight;
 221 };
 222
 223 /* This is per cgroup per device grouping structure */
 224 struct cfq_group {
 225         /* must be the first member */
 226         struct blkg_policy_data pd;
 227
 228         /* group service_tree member */
 229         struct rb_node rb_node;
 230
 231         /* group service_tree key */
 232         u64 vdisktime;
 233
 234         /*
 235          * The number of active cfqgs and sum of their weights under this
 236          * cfqg.  This covers this cfqg's leaf_weight and all children's
 237          * weights, but does not cover weights of further descendants.
 238          *
 239          * If a cfqg is on the service tree, it's active.  An active cfqg
 240          * also activates its parent and contributes to the children_weight
 241          * of the parent.
 242          */
 243         int nr_active;
 244         unsigned int children_weight;
 245
 246         /*
 247          * vfraction is the fraction of vdisktime that the tasks in this
 248          * cfqg are entitled to.  This is determined by compounding the
 249          * ratios walking up from this cfqg to the root.
 250          *
 251          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 252          * vfractions on a service tree is approximately 1.  The sum may
 253          * deviate a bit due to rounding errors and fluctuations caused by
 254          * cfqgs entering and leaving the service tree.
 255          */
 256         unsigned int vfraction;
 257
 258         /*
 259          * There are two weights - (internal) weight is the weight of this
 260          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 261          * this cfqg against the child cfqgs.  For the root cfqg, both
 262          * weights are kept in sync for backward compatibility.
 263          */
 264         unsigned int weight;
 265         unsigned int new_weight;
 266         unsigned int dev_weight;
 267
 268         unsigned int leaf_weight;
 269         unsigned int new_leaf_weight;
 270         unsigned int dev_leaf_weight;
 271
 272         /* number of cfqq currently on this group */
 273         int nr_cfqq;
 274
 275         /*
 276          * Per group busy queues average. Useful for workload slice calc. We
 277          * create the array for each prio class but at run time it is used
 278          * only for RT and BE class and slot for IDLE class remains unused.
 279          * This is primarily done to avoid confusion and a gcc warning.
 280          */
 281         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 282         /*
 283          * rr lists of queues with requests. We maintain service trees for
 284          * RT and BE classes. These trees are subdivided in subclasses
 285          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 286          * class there is no subclassification and all the cfq queues go on
 287          * a single tree service_tree_idle.
 288          * Counts are embedded in the cfq_rb_root
 289          */
 290         struct cfq_rb_root service_trees[2][3];
 291         struct cfq_rb_root service_tree_idle;
 292
 293         u64 saved_wl_slice;
 294         enum wl_type_t saved_wl_type;
 295         enum wl_class_t saved_wl_class;
 296
 297         /* number of requests that are on the dispatch list or inside driver */
 298         int dispatched;
 299         struct cfq_ttime ttime;
 300         struct cfqg_stats stats;        /* stats for this cfqg */
 301
 302         /* async queue for each priority case */
 303         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 304         struct cfq_queue *async_idle_cfqq;
 305
 306 };
 307
 308 struct cfq_io_cq {
 309         struct io_cq            icq;            /* must be the first member */
 310         struct cfq_queue        *cfqq[2];
 311         struct cfq_ttime        ttime;
 312         int                     ioprio;         /* the current ioprio */
 313 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 314         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 315 #endif
 316 };
 317
 318 /*
 319  * Per block device queue structure
 320  */
 321 struct cfq_data {
 322         struct request_queue *queue;
 323         /* Root service tree for cfq_groups */
 324         struct cfq_rb_root grp_service_tree;
 325         struct cfq_group *root_group;
 326
 327         /*
 328          * The priority currently being served
 329          */
 330         enum wl_class_t serving_wl_class;
 331         enum wl_type_t serving_wl_type;
 332         u64 workload_expires;
 333         struct cfq_group *serving_group;
 334
 335         /*
 336          * Each priority tree is sorted by next_request position.  These
 337          * trees are used when determining if two or more queues are
 338          * interleaving requests (see cfq_close_cooperator).
 339          */
 340         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 341
 342         unsigned int busy_queues;
 343         unsigned int busy_sync_queues;
 344
 345         int rq_in_driver;
 346         int rq_in_flight[2];
 347
 348         /*
 349          * queue-depth detection
 350          */
 351         int rq_queued;
 352         int hw_tag;
 353         /*
 354          * hw_tag can be
 355          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 356          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 357          *  0 => no NCQ
 358          */
 359         int hw_tag_est_depth;
 360         unsigned int hw_tag_samples;
 361
 362         /*
 363          * idle window management
 364          */
 365         struct hrtimer idle_slice_timer;
 366         struct work_struct unplug_work;
 367
 368         struct cfq_queue *active_queue;
 369         struct cfq_io_cq *active_cic;
 370
 371         sector_t last_position;
 372
 373         /*
 374          * tunables, see top of file
 375          */
 376         unsigned int cfq_quantum;
 377         unsigned int cfq_back_penalty;
 378         unsigned int cfq_back_max;
 379         unsigned int cfq_slice_async_rq;
 380         unsigned int cfq_latency;
 381         u64 cfq_fifo_expire[2];
 382         u64 cfq_slice[2];
 383         u64 cfq_slice_idle;
 384         u64 cfq_group_idle;
 385         u64 cfq_target_latency;
 386
 387         /*
 388          * Fallback dummy cfqq for extreme OOM conditions
 389          */
 390         struct cfq_queue oom_cfqq;
 391
 392         u64 last_delayed_sync;
 393 };
 394
 395 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 396 static void cfq_put_queue(struct cfq_queue *cfqq);
 397
 398 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 399                                             enum wl_class_t class,
 400                                             enum wl_type_t type)
 401 {
 402         if (!cfqg)
 403                 return NULL;
 404
 405         if (class == IDLE_WORKLOAD)
 406                 return &cfqg->service_tree_idle;
 407
 408         return &cfqg->service_trees[class][type];
 409 }
 410
 411 enum cfqq_state_flags {
 412         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 413         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 414         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 415         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 416         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 417         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 418         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 419         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 420         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 421         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 422         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 423         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 424         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 425 };
 426
 427 #define CFQ_CFQQ_FNS(name)                                              \
 428 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 429 {                                                                       \
 430         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 431 }                                                                       \
 432 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 433 {                                                                       \
 434         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 435 }                                                                       \
 436 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 437 {                                                                       \
 438         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 439 }
 440
 441 CFQ_CFQQ_FNS(on_rr);
 442 CFQ_CFQQ_FNS(wait_request);
 443 CFQ_CFQQ_FNS(must_dispatch);
 444 CFQ_CFQQ_FNS(must_alloc_slice);
 445 CFQ_CFQQ_FNS(fifo_expire);
 446 CFQ_CFQQ_FNS(idle_window);
 447 CFQ_CFQQ_FNS(prio_changed);
 448 CFQ_CFQQ_FNS(slice_new);
 449 CFQ_CFQQ_FNS(sync);
 450 CFQ_CFQQ_FNS(coop);
 451 CFQ_CFQQ_FNS(split_coop);
 452 CFQ_CFQQ_FNS(deep);
 453 CFQ_CFQQ_FNS(wait_busy);
 454 #undef CFQ_CFQQ_FNS
 455
 456 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 457
 458 /* cfqg stats flags */
 459 enum cfqg_stats_flags {
 460         CFQG_stats_waiting = 0,
 461         CFQG_stats_idling,
 462         CFQG_stats_empty,
 463 };
 464
 465 #define CFQG_FLAG_FNS(name)                                             \
 466 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 467 {                                                                       \
 468         stats->flags |= (1 << CFQG_stats_##name);                       \
 469 }                                                                       \
 470 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 471 {                                                                       \
 472         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 473 }                                                                       \
 474 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 475 {                                                                       \
 476         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 477 }                                                                       \
 478
 479 CFQG_FLAG_FNS(waiting)
 480 CFQG_FLAG_FNS(idling)
 481 CFQG_FLAG_FNS(empty)
 482 #undef CFQG_FLAG_FNS
 483
 484 /* This should be called with the queue_lock held. */
 485 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 486 {
 487         unsigned long long now;
 488
 489         if (!cfqg_stats_waiting(stats))
 490                 return;
 491
 492         now = sched_clock();
 493         if (time_after64(now, stats->start_group_wait_time))
 494                 blkg_stat_add(&stats->group_wait_time,
 495                               now - stats->start_group_wait_time);
 496         cfqg_stats_clear_waiting(stats);
 497 }
 498
 499 /* This should be called with the queue_lock held. */
 500 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 501                                                  struct cfq_group *curr_cfqg)
 502 {
 503         struct cfqg_stats *stats = &cfqg->stats;
 504
 505         if (cfqg_stats_waiting(stats))
 506                 return;
 507         if (cfqg == curr_cfqg)
 508                 return;
 509         stats->start_group_wait_time = sched_clock();
 510         cfqg_stats_mark_waiting(stats);
 511 }
 512
 513 /* This should be called with the queue_lock held. */
 514 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 515 {
 516         unsigned long long now;
 517
 518         if (!cfqg_stats_empty(stats))
 519                 return;
 520
 521         now = sched_clock();
 522         if (time_after64(now, stats->start_empty_time))
 523                 blkg_stat_add(&stats->empty_time,
 524                               now - stats->start_empty_time);
 525         cfqg_stats_clear_empty(stats);
 526 }
 527
 528 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 529 {
 530         blkg_stat_add(&cfqg->stats.dequeue, 1);
 531 }
 532
 533 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 534 {
 535         struct cfqg_stats *stats = &cfqg->stats;
 536
 537         if (blkg_rwstat_total(&stats->queued))
 538                 return;
 539
 540         /*
 541          * group is already marked empty. This can happen if cfqq got new
 542          * request in parent group and moved to this group while being added
 543          * to service tree. Just ignore the event and move on.
 544          */
 545         if (cfqg_stats_empty(stats))
 546                 return;
 547
 548         stats->start_empty_time = sched_clock();
 549         cfqg_stats_mark_empty(stats);
 550 }
 551
 552 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 553 {
 554         struct cfqg_stats *stats = &cfqg->stats;
 555
 556         if (cfqg_stats_idling(stats)) {
 557                 unsigned long long now = sched_clock();
 558
 559                 if (time_after64(now, stats->start_idle_time))
 560                         blkg_stat_add(&stats->idle_time,
 561                                       now - stats->start_idle_time);
 562                 cfqg_stats_clear_idling(stats);
 563         }
 564 }
 565
 566 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 567 {
 568         struct cfqg_stats *stats = &cfqg->stats;
 569
 570         BUG_ON(cfqg_stats_idling(stats));
 571
 572         stats->start_idle_time = sched_clock();
 573         cfqg_stats_mark_idling(stats);
 574 }
 575
 576 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 577 {
 578         struct cfqg_stats *stats = &cfqg->stats;
 579
 580         blkg_stat_add(&stats->avg_queue_size_sum,
 581                       blkg_rwstat_total(&stats->queued));
 582         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 583         cfqg_stats_update_group_wait_time(stats);
 584 }
 585
 586 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 587
 588 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 589 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 590 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 591 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 592 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 593 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 594 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 595
 596 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 597
 598 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 599
 600 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 601 {
 602         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 603 }
 604
 605 static struct cfq_group_data
 606 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 607 {
 608         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 609 }
 610
 611 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 612 {
 613         return pd_to_blkg(&cfqg->pd);
 614 }
 615
 616 static struct blkcg_policy blkcg_policy_cfq;
 617
 618 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 619 {
 620         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 621 }
 622
 623 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 624 {
 625         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 626 }
 627
 628 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 629 {
 630         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 631
 632         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 633 }
 634
 635 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 636                                       struct cfq_group *ancestor)
 637 {
 638         return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
 639                                     cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
 640 }
 641
 642 static inline void cfqg_get(struct cfq_group *cfqg)
 643 {
 644         return blkg_get(cfqg_to_blkg(cfqg));
 645 }
 646
 647 static inline void cfqg_put(struct cfq_group *cfqg)
 648 {
 649         return blkg_put(cfqg_to_blkg(cfqg));
 650 }
 651
 652 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 653         char __pbuf[128];                                               \
 654                                                                         \
 655         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 656         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 657                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 658                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 659                           __pbuf, ##args);                              \
 660 } while (0)
 661
 662 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 663         char __pbuf[128];                                               \
 664                                                                         \
 665         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 666         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 667 } while (0)
 668
 669 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 670                                             struct cfq_group *curr_cfqg, int op,
 671                                             int op_flags)
 672 {
 673         blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, 1);
 674         cfqg_stats_end_empty_time(&cfqg->stats);
 675         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 676 }
 677
 678 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 679                         uint64_t time, unsigned long unaccounted_time)
 680 {
 681         blkg_stat_add(&cfqg->stats.time, time);
 682 #ifdef CONFIG_DEBUG_BLK_CGROUP
 683         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 684 #endif
 685 }
 686
 687 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
 688                                                int op_flags)
 689 {
 690         blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, -1);
 691 }
 692
 693 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
 694                                                int op_flags)
 695 {
 696         blkg_rwstat_add(&cfqg->stats.merged, op, op_flags, 1);
 697 }
 698
 699 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 700                         uint64_t start_time, uint64_t io_start_time, int op,
 701                         int op_flags)
 702 {
 703         struct cfqg_stats *stats = &cfqg->stats;
 704         unsigned long long now = sched_clock();
 705
 706         if (time_after64(now, io_start_time))
 707                 blkg_rwstat_add(&stats->service_time, op, op_flags,
 708                                 now - io_start_time);
 709         if (time_after64(io_start_time, start_time))
 710                 blkg_rwstat_add(&stats->wait_time, op, op_flags,
 711                                 io_start_time - start_time);
 712 }
 713
 714 /* @stats = 0 */
 715 static void cfqg_stats_reset(struct cfqg_stats *stats)
 716 {
 717         /* queued stats shouldn't be cleared */
 718         blkg_rwstat_reset(&stats->merged);
 719         blkg_rwstat_reset(&stats->service_time);
 720         blkg_rwstat_reset(&stats->wait_time);
 721         blkg_stat_reset(&stats->time);
 722 #ifdef CONFIG_DEBUG_BLK_CGROUP
 723         blkg_stat_reset(&stats->unaccounted_time);
 724         blkg_stat_reset(&stats->avg_queue_size_sum);
 725         blkg_stat_reset(&stats->avg_queue_size_samples);
 726         blkg_stat_reset(&stats->dequeue);
 727         blkg_stat_reset(&stats->group_wait_time);
 728         blkg_stat_reset(&stats->idle_time);
 729         blkg_stat_reset(&stats->empty_time);
 730 #endif
 731 }
 732
 733 /* @to += @from */
 734 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 735 {
 736         /* queued stats shouldn't be cleared */
 737         blkg_rwstat_add_aux(&to->merged, &from->merged);
 738         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 739         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 740         blkg_stat_add_aux(&from->time, &from->time);
 741 #ifdef CONFIG_DEBUG_BLK_CGROUP
 742         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 743         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 744         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 745         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 746         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 747         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 748         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 749 #endif
 750 }
 751
 752 /*
 753  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 754  * recursive stats can still account for the amount used by this cfqg after
 755  * it's gone.
 756  */
 757 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 758 {
 759         struct cfq_group *parent = cfqg_parent(cfqg);
 760
 761         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 762
 763         if (unlikely(!parent))
 764                 return;
 765
 766         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 767         cfqg_stats_reset(&cfqg->stats);
 768 }
 769
 770 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 771
 772 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 773 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 774                                       struct cfq_group *ancestor)
 775 {
 776         return true;
 777 }
 778 static inline void cfqg_get(struct cfq_group *cfqg) { }
 779 static inline void cfqg_put(struct cfq_group *cfqg) { }
 780
 781 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 782         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 783                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 784                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 785                                 ##args)
 786 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 787
 788 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 789                         struct cfq_group *curr_cfqg, int op, int op_flags) { }
 790 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 791                         uint64_t time, unsigned long unaccounted_time) { }
 792 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
 793                         int op_flags) { }
 794 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
 795                         int op_flags) { }
 796 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 797                         uint64_t start_time, uint64_t io_start_time, int op,
 798                         int op_flags) { }
 799
 800 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 801
 802 #define cfq_log(cfqd, fmt, args...)     \
 803         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 804
 805 /* Traverses through cfq group service trees */
 806 #define for_each_cfqg_st(cfqg, i, j, st) \
 807         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 808                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 809                         : &cfqg->service_tree_idle; \
 810                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 811                         (i == IDLE_WORKLOAD && j == 0); \
 812                         j++, st = i < IDLE_WORKLOAD ? \
 813                         &cfqg->service_trees[i][j]: NULL) \
 814
 815 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 816         struct cfq_ttime *ttime, bool group_idle)
 817 {
 818         u64 slice;
 819         if (!sample_valid(ttime->ttime_samples))
 820                 return false;
 821         if (group_idle)
 822                 slice = cfqd->cfq_group_idle;
 823         else
 824                 slice = cfqd->cfq_slice_idle;
 825         return ttime->ttime_mean > slice;
 826 }
 827
 828 static inline bool iops_mode(struct cfq_data *cfqd)
 829 {
 830         /*
 831          * If we are not idling on queues and it is a NCQ drive, parallel
 832          * execution of requests is on and measuring time is not possible
 833          * in most of the cases until and unless we drive shallower queue
 834          * depths and that becomes a performance bottleneck. In such cases
 835          * switch to start providing fairness in terms of number of IOs.
 836          */
 837         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 838                 return true;
 839         else
 840                 return false;
 841 }
 842
 843 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 844 {
 845         if (cfq_class_idle(cfqq))
 846                 return IDLE_WORKLOAD;
 847         if (cfq_class_rt(cfqq))
 848                 return RT_WORKLOAD;
 849         return BE_WORKLOAD;
 850 }
 851
 852
 853 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 854 {
 855         if (!cfq_cfqq_sync(cfqq))
 856                 return ASYNC_WORKLOAD;
 857         if (!cfq_cfqq_idle_window(cfqq))
 858                 return SYNC_NOIDLE_WORKLOAD;
 859         return SYNC_WORKLOAD;
 860 }
 861
 862 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 863                                         struct cfq_data *cfqd,
 864                                         struct cfq_group *cfqg)
 865 {
 866         if (wl_class == IDLE_WORKLOAD)
 867                 return cfqg->service_tree_idle.count;
 868
 869         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 870                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 871                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 872 }
 873
 874 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 875                                         struct cfq_group *cfqg)
 876 {
 877         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 878                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 879 }
 880
 881 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 882 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 883                                        struct cfq_io_cq *cic, struct bio *bio);
 884
 885 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 886 {
 887         /* cic->icq is the first member, %NULL will convert to %NULL */
 888         return container_of(icq, struct cfq_io_cq, icq);
 889 }
 890
 891 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 892                                                struct io_context *ioc)
 893 {
 894         if (ioc)
 895                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 896         return NULL;
 897 }
 898
 899 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 900 {
 901         return cic->cfqq[is_sync];
 902 }
 903
 904 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 905                                 bool is_sync)
 906 {
 907         cic->cfqq[is_sync] = cfqq;
 908 }
 909
 910 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 911 {
 912         return cic->icq.q->elevator->elevator_data;
 913 }
 914
 915 /*
 916  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 917  * set (in which case it could also be direct WRITE).
 918  */
 919 static inline bool cfq_bio_sync(struct bio *bio)
 920 {
 921         return bio_data_dir(bio) == READ || (bio->bi_opf & REQ_SYNC);
 922 }
 923
 924 /*
 925  * scheduler run of queue, if there are requests pending and no one in the
 926  * driver that will restart queueing
 927  */
 928 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 929 {
 930         if (cfqd->busy_queues) {
 931                 cfq_log(cfqd, "schedule dispatch");
 932                 kblockd_schedule_work(&cfqd->unplug_work);
 933         }
 934 }
 935
 936 /*
 937  * Scale schedule slice based on io priority. Use the sync time slice only
 938  * if a queue is marked sync and has sync io queued. A sync queue with async
 939  * io only, should not get full sync slice length.
 940  */
 941 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 942                                  unsigned short prio)
 943 {
 944         u64 base_slice = cfqd->cfq_slice[sync];
 945         u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
 946
 947         WARN_ON(prio >= IOPRIO_BE_NR);
 948
 949         return base_slice + (slice * (4 - prio));
 950 }
 951
 952 static inline u64
 953 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 954 {
 955         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 956 }
 957
 958 /**
 959  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 960  * @charge: disk time being charged
 961  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 962  *
 963  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 964  * scaling is inversely proportional.
 965  *
 966  * scaled = charge / vfraction
 967  *
 968  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 969  */
 970 static inline u64 cfqg_scale_charge(u64 charge,
 971                                     unsigned int vfraction)
 972 {
 973         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 974
 975         /* charge / vfraction */
 976         c <<= CFQ_SERVICE_SHIFT;
 977         return div_u64(c, vfraction);
 978 }
 979
 980 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 981 {
 982         s64 delta = (s64)(vdisktime - min_vdisktime);
 983         if (delta > 0)
 984                 min_vdisktime = vdisktime;
 985
 986         return min_vdisktime;
 987 }
 988
 989 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 990 {
 991         s64 delta = (s64)(vdisktime - min_vdisktime);
 992         if (delta < 0)
 993                 min_vdisktime = vdisktime;
 994
 995         return min_vdisktime;
 996 }
 997
 998 static void update_min_vdisktime(struct cfq_rb_root *st)
 999 {
1000         struct cfq_group *cfqg;
1001
1002         if (st->left) {
1003                 cfqg = rb_entry_cfqg(st->left);
1004                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1005                                                   cfqg->vdisktime);
1006         }
1007 }
1008
1009 /*
1010  * get averaged number of queues of RT/BE priority.
1011  * average is updated, with a formula that gives more weight to higher numbers,
1012  * to quickly follows sudden increases and decrease slowly
1013  */
1014
1015 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1016                                         struct cfq_group *cfqg, bool rt)
1017 {
1018         unsigned min_q, max_q;
1019         unsigned mult  = cfq_hist_divisor - 1;
1020         unsigned round = cfq_hist_divisor / 2;
1021         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1022
1023         min_q = min(cfqg->busy_queues_avg[rt], busy);
1024         max_q = max(cfqg->busy_queues_avg[rt], busy);
1025         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1026                 cfq_hist_divisor;
1027         return cfqg->busy_queues_avg[rt];
1028 }
1029
1030 static inline u64
1031 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1032 {
1033         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1034 }
1035
1036 static inline u64
1037 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1038 {
1039         u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1040         if (cfqd->cfq_latency) {
1041                 /*
1042                  * interested queues (we consider only the ones with the same
1043                  * priority class in the cfq group)
1044                  */
1045                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1046                                                 cfq_class_rt(cfqq));
1047                 u64 sync_slice = cfqd->cfq_slice[1];
1048                 u64 expect_latency = sync_slice * iq;
1049                 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1050
1051                 if (expect_latency > group_slice) {
1052                         u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1053                         u64 low_slice;
1054
1055                         /* scale low_slice according to IO priority
1056                          * and sync vs async */
1057                         low_slice = div64_u64(base_low_slice*slice, sync_slice);
1058                         low_slice = min(slice, low_slice);
1059                         /* the adapted slice value is scaled to fit all iqs
1060                          * into the target latency */
1061                         slice = div64_u64(slice*group_slice, expect_latency);
1062                         slice = max(slice, low_slice);
1063                 }
1064         }
1065         return slice;
1066 }
1067
1068 static inline void
1069 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1070 {
1071         u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1072         u64 now = ktime_get_ns();
1073
1074         cfqq->slice_start = now;
1075         cfqq->slice_end = now + slice;
1076         cfqq->allocated_slice = slice;
1077         cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1078 }
1079
1080 /*
1081  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1082  * isn't valid until the first request from the dispatch is activated
1083  * and the slice time set.
1084  */
1085 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1086 {
1087         if (cfq_cfqq_slice_new(cfqq))
1088                 return false;
1089         if (ktime_get_ns() < cfqq->slice_end)
1090                 return false;
1091
1092         return true;
1093 }
1094
1095 /*
1096  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1097  * We choose the request that is closest to the head right now. Distance
1098  * behind the head is penalized and only allowed to a certain extent.
1099  */
1100 static struct request *
1101 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1102 {
1103         sector_t s1, s2, d1 = 0, d2 = 0;
1104         unsigned long back_max;
1105 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1106 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1107         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1108
1109         if (rq1 == NULL || rq1 == rq2)
1110                 return rq2;
1111         if (rq2 == NULL)
1112                 return rq1;
1113
1114         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1115                 return rq_is_sync(rq1) ? rq1 : rq2;
1116
1117         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1118                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1119
1120         s1 = blk_rq_pos(rq1);
1121         s2 = blk_rq_pos(rq2);
1122
1123         /*
1124          * by definition, 1KiB is 2 sectors
1125          */
1126         back_max = cfqd->cfq_back_max * 2;
1127
1128         /*
1129          * Strict one way elevator _except_ in the case where we allow
1130          * short backward seeks which are biased as twice the cost of a
1131          * similar forward seek.
1132          */
1133         if (s1 >= last)
1134                 d1 = s1 - last;
1135         else if (s1 + back_max >= last)
1136                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1137         else
1138                 wrap |= CFQ_RQ1_WRAP;
1139
1140         if (s2 >= last)
1141                 d2 = s2 - last;
1142         else if (s2 + back_max >= last)
1143                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1144         else
1145                 wrap |= CFQ_RQ2_WRAP;
1146
1147         /* Found required data */
1148
1149         /*
1150          * By doing switch() on the bit mask "wrap" we avoid having to
1151          * check two variables for all permutations: --> faster!
1152          */
1153         switch (wrap) {
1154         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1155                 if (d1 < d2)
1156                         return rq1;
1157                 else if (d2 < d1)
1158                         return rq2;
1159                 else {
1160                         if (s1 >= s2)
1161                                 return rq1;
1162                         else
1163                                 return rq2;
1164                 }
1165
1166         case CFQ_RQ2_WRAP:
1167                 return rq1;
1168         case CFQ_RQ1_WRAP:
1169                 return rq2;
1170         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1171         default:
1172                 /*
1173                  * Since both rqs are wrapped,
1174                  * start with the one that's further behind head
1175                  * (--> only *one* back seek required),
1176                  * since back seek takes more time than forward.
1177                  */
1178                 if (s1 <= s2)
1179                         return rq1;
1180                 else
1181                         return rq2;
1182         }
1183 }
1184
1185 /*
1186  * The below is leftmost cache rbtree addon
1187  */
1188 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1189 {
1190         /* Service tree is empty */
1191         if (!root->count)
1192                 return NULL;
1193
1194         if (!root->left)
1195                 root->left = rb_first(&root->rb);
1196
1197         if (root->left)
1198                 return rb_entry(root->left, struct cfq_queue, rb_node);
1199
1200         return NULL;
1201 }
1202
1203 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1204 {
1205         if (!root->left)
1206                 root->left = rb_first(&root->rb);
1207
1208         if (root->left)
1209                 return rb_entry_cfqg(root->left);
1210
1211         return NULL;
1212 }
1213
1214 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1215 {
1216         rb_erase(n, root);
1217         RB_CLEAR_NODE(n);
1218 }
1219
1220 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1221 {
1222         if (root->left == n)
1223                 root->left = NULL;
1224         rb_erase_init(n, &root->rb);
1225         --root->count;
1226 }
1227
1228 /*
1229  * would be nice to take fifo expire time into account as well
1230  */
1231 static struct request *
1232 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1233                   struct request *last)
1234 {
1235         struct rb_node *rbnext = rb_next(&last->rb_node);
1236         struct rb_node *rbprev = rb_prev(&last->rb_node);
1237         struct request *next = NULL, *prev = NULL;
1238
1239         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1240
1241         if (rbprev)
1242                 prev = rb_entry_rq(rbprev);
1243
1244         if (rbnext)
1245                 next = rb_entry_rq(rbnext);
1246         else {
1247                 rbnext = rb_first(&cfqq->sort_list);
1248                 if (rbnext && rbnext != &last->rb_node)
1249                         next = rb_entry_rq(rbnext);
1250         }
1251
1252         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1253 }
1254
1255 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1256                             struct cfq_queue *cfqq)
1257 {
1258         /*
1259          * just an approximation, should be ok.
1260          */
1261         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1262                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1263 }
1264
1265 static inline s64
1266 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1267 {
1268         return cfqg->vdisktime - st->min_vdisktime;
1269 }
1270
1271 static void
1272 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1273 {
1274         struct rb_node **node = &st->rb.rb_node;
1275         struct rb_node *parent = NULL;
1276         struct cfq_group *__cfqg;
1277         s64 key = cfqg_key(st, cfqg);
1278         int left = 1;
1279
1280         while (*node != NULL) {
1281                 parent = *node;
1282                 __cfqg = rb_entry_cfqg(parent);
1283
1284                 if (key < cfqg_key(st, __cfqg))
1285                         node = &parent->rb_left;
1286                 else {
1287                         node = &parent->rb_right;
1288                         left = 0;
1289                 }
1290         }
1291
1292         if (left)
1293                 st->left = &cfqg->rb_node;
1294
1295         rb_link_node(&cfqg->rb_node, parent, node);
1296         rb_insert_color(&cfqg->rb_node, &st->rb);
1297 }
1298
1299 /*
1300  * This has to be called only on activation of cfqg
1301  */
1302 static void
1303 cfq_update_group_weight(struct cfq_group *cfqg)
1304 {
1305         if (cfqg->new_weight) {
1306                 cfqg->weight = cfqg->new_weight;
1307                 cfqg->new_weight = 0;
1308         }
1309 }
1310
1311 static void
1312 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1313 {
1314         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1315
1316         if (cfqg->new_leaf_weight) {
1317                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1318                 cfqg->new_leaf_weight = 0;
1319         }
1320 }
1321
1322 static void
1323 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1324 {
1325         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1326         struct cfq_group *pos = cfqg;
1327         struct cfq_group *parent;
1328         bool propagate;
1329
1330         /* add to the service tree */
1331         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1332
1333         /*
1334          * Update leaf_weight.  We cannot update weight at this point
1335          * because cfqg might already have been activated and is
1336          * contributing its current weight to the parent's child_weight.
1337          */
1338         cfq_update_group_leaf_weight(cfqg);
1339         __cfq_group_service_tree_add(st, cfqg);
1340
1341         /*
1342          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1343          * entitled to.  vfraction is calculated by walking the tree
1344          * towards the root calculating the fraction it has at each level.
1345          * The compounded ratio is how much vfraction @cfqg owns.
1346          *
1347          * Start with the proportion tasks in this cfqg has against active
1348          * children cfqgs - its leaf_weight against children_weight.
1349          */
1350         propagate = !pos->nr_active++;
1351         pos->children_weight += pos->leaf_weight;
1352         vfr = vfr * pos->leaf_weight / pos->children_weight;
1353
1354         /*
1355          * Compound ->weight walking up the tree.  Both activation and
1356          * vfraction calculation are done in the same loop.  Propagation
1357          * stops once an already activated node is met.  vfraction
1358          * calculation should always continue to the root.
1359          */
1360         while ((parent = cfqg_parent(pos))) {
1361                 if (propagate) {
1362                         cfq_update_group_weight(pos);
1363                         propagate = !parent->nr_active++;
1364                         parent->children_weight += pos->weight;
1365                 }
1366                 vfr = vfr * pos->weight / parent->children_weight;
1367                 pos = parent;
1368         }
1369
1370         cfqg->vfraction = max_t(unsigned, vfr, 1);
1371 }
1372
1373 static void
1374 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1375 {
1376         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1377         struct cfq_group *__cfqg;
1378         struct rb_node *n;
1379
1380         cfqg->nr_cfqq++;
1381         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1382                 return;
1383
1384         /*
1385          * Currently put the group at the end. Later implement something
1386          * so that groups get lesser vtime based on their weights, so that
1387          * if group does not loose all if it was not continuously backlogged.
1388          */
1389         n = rb_last(&st->rb);
1390         if (n) {
1391                 __cfqg = rb_entry_cfqg(n);
1392                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1393         } else
1394                 cfqg->vdisktime = st->min_vdisktime;
1395         cfq_group_service_tree_add(st, cfqg);
1396 }
1397
1398 static void
1399 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1400 {
1401         struct cfq_group *pos = cfqg;
1402         bool propagate;
1403
1404         /*
1405          * Undo activation from cfq_group_service_tree_add().  Deactivate
1406          * @cfqg and propagate deactivation upwards.
1407          */
1408         propagate = !--pos->nr_active;
1409         pos->children_weight -= pos->leaf_weight;
1410
1411         while (propagate) {
1412                 struct cfq_group *parent = cfqg_parent(pos);
1413
1414                 /* @pos has 0 nr_active at this point */
1415                 WARN_ON_ONCE(pos->children_weight);
1416                 pos->vfraction = 0;
1417
1418                 if (!parent)
1419                         break;
1420
1421                 propagate = !--parent->nr_active;
1422                 parent->children_weight -= pos->weight;
1423                 pos = parent;
1424         }
1425
1426         /* remove from the service tree */
1427         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1428                 cfq_rb_erase(&cfqg->rb_node, st);
1429 }
1430
1431 static void
1432 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1433 {
1434         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1435
1436         BUG_ON(cfqg->nr_cfqq < 1);
1437         cfqg->nr_cfqq--;
1438
1439         /* If there are other cfq queues under this group, don't delete it */
1440         if (cfqg->nr_cfqq)
1441                 return;
1442
1443         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1444         cfq_group_service_tree_del(st, cfqg);
1445         cfqg->saved_wl_slice = 0;
1446         cfqg_stats_update_dequeue(cfqg);
1447 }
1448
1449 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1450                                        u64 *unaccounted_time)
1451 {
1452         u64 slice_used;
1453         u64 now = ktime_get_ns();
1454
1455         /*
1456          * Queue got expired before even a single request completed or
1457          * got expired immediately after first request completion.
1458          */
1459         if (!cfqq->slice_start || cfqq->slice_start == now) {
1460                 /*
1461                  * Also charge the seek time incurred to the group, otherwise
1462                  * if there are mutiple queues in the group, each can dispatch
1463                  * a single request on seeky media and cause lots of seek time
1464                  * and group will never know it.
1465                  */
1466                 slice_used = max_t(u64, (now - cfqq->dispatch_start),
1467                                         jiffies_to_nsecs(1));
1468         } else {
1469                 slice_used = now - cfqq->slice_start;
1470                 if (slice_used > cfqq->allocated_slice) {
1471                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1472                         slice_used = cfqq->allocated_slice;
1473                 }
1474                 if (cfqq->slice_start > cfqq->dispatch_start)
1475                         *unaccounted_time += cfqq->slice_start -
1476                                         cfqq->dispatch_start;
1477         }
1478
1479         return slice_used;
1480 }
1481
1482 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1483                                 struct cfq_queue *cfqq)
1484 {
1485         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1486         u64 used_sl, charge, unaccounted_sl = 0;
1487         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1488                         - cfqg->service_tree_idle.count;
1489         unsigned int vfr;
1490         u64 now = ktime_get_ns();
1491
1492         BUG_ON(nr_sync < 0);
1493         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1494
1495         if (iops_mode(cfqd))
1496                 charge = cfqq->slice_dispatch;
1497         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1498                 charge = cfqq->allocated_slice;
1499
1500         /*
1501          * Can't update vdisktime while on service tree and cfqg->vfraction
1502          * is valid only while on it.  Cache vfr, leave the service tree,
1503          * update vdisktime and go back on.  The re-addition to the tree
1504          * will also update the weights as necessary.
1505          */
1506         vfr = cfqg->vfraction;
1507         cfq_group_service_tree_del(st, cfqg);
1508         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1509         cfq_group_service_tree_add(st, cfqg);
1510
1511         /* This group is being expired. Save the context */
1512         if (cfqd->workload_expires > now) {
1513                 cfqg->saved_wl_slice = cfqd->workload_expires - now;
1514                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1515                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1516         } else
1517                 cfqg->saved_wl_slice = 0;
1518
1519         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1520                                         st->min_vdisktime);
1521         cfq_log_cfqq(cfqq->cfqd, cfqq,
1522                      "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
1523                      used_sl, cfqq->slice_dispatch, charge,
1524                      iops_mode(cfqd), cfqq->nr_sectors);
1525         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1526         cfqg_stats_set_start_empty_time(cfqg);
1527 }
1528
1529 /**
1530  * cfq_init_cfqg_base - initialize base part of a cfq_group
1531  * @cfqg: cfq_group to initialize
1532  *
1533  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1534  * is enabled or not.
1535  */
1536 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1537 {
1538         struct cfq_rb_root *st;
1539         int i, j;
1540
1541         for_each_cfqg_st(cfqg, i, j, st)
1542                 *st = CFQ_RB_ROOT;
1543         RB_CLEAR_NODE(&cfqg->rb_node);
1544
1545         cfqg->ttime.last_end_request = ktime_get_ns();
1546 }
1547
1548 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1549 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1550                             bool on_dfl, bool reset_dev, bool is_leaf_weight);
1551
1552 static void cfqg_stats_exit(struct cfqg_stats *stats)
1553 {
1554         blkg_rwstat_exit(&stats->merged);
1555         blkg_rwstat_exit(&stats->service_time);
1556         blkg_rwstat_exit(&stats->wait_time);
1557         blkg_rwstat_exit(&stats->queued);
1558         blkg_stat_exit(&stats->time);
1559 #ifdef CONFIG_DEBUG_BLK_CGROUP
1560         blkg_stat_exit(&stats->unaccounted_time);
1561         blkg_stat_exit(&stats->avg_queue_size_sum);
1562         blkg_stat_exit(&stats->avg_queue_size_samples);
1563         blkg_stat_exit(&stats->dequeue);
1564         blkg_stat_exit(&stats->group_wait_time);
1565         blkg_stat_exit(&stats->idle_time);
1566         blkg_stat_exit(&stats->empty_time);
1567 #endif
1568 }
1569
1570 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1571 {
1572         if (blkg_rwstat_init(&stats->merged, gfp) ||
1573             blkg_rwstat_init(&stats->service_time, gfp) ||
1574             blkg_rwstat_init(&stats->wait_time, gfp) ||
1575             blkg_rwstat_init(&stats->queued, gfp) ||
1576             blkg_stat_init(&stats->time, gfp))
1577                 goto err;
1578
1579 #ifdef CONFIG_DEBUG_BLK_CGROUP
1580         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1581             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1582             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1583             blkg_stat_init(&stats->dequeue, gfp) ||
1584             blkg_stat_init(&stats->group_wait_time, gfp) ||
1585             blkg_stat_init(&stats->idle_time, gfp) ||
1586             blkg_stat_init(&stats->empty_time, gfp))
1587                 goto err;
1588 #endif
1589         return 0;
1590 err:
1591         cfqg_stats_exit(stats);
1592         return -ENOMEM;
1593 }
1594
1595 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1596 {
1597         struct cfq_group_data *cgd;
1598
1599         cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1600         if (!cgd)
1601                 return NULL;
1602         return &cgd->cpd;
1603 }
1604
1605 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1606 {
1607         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1608         unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1609                               CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1610
1611         if (cpd_to_blkcg(cpd) == &blkcg_root)
1612                 weight *= 2;
1613
1614         cgd->weight = weight;
1615         cgd->leaf_weight = weight;
1616 }
1617
1618 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1619 {
1620         kfree(cpd_to_cfqgd(cpd));
1621 }
1622
1623 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1624 {
1625         struct blkcg *blkcg = cpd_to_blkcg(cpd);
1626         bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1627         unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1628
1629         if (blkcg == &blkcg_root)
1630                 weight *= 2;
1631
1632         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1633         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1634 }
1635
1636 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1637 {
1638         struct cfq_group *cfqg;
1639
1640         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1641         if (!cfqg)
1642                 return NULL;
1643
1644         cfq_init_cfqg_base(cfqg);
1645         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1646                 kfree(cfqg);
1647                 return NULL;
1648         }
1649
1650         return &cfqg->pd;
1651 }
1652
1653 static void cfq_pd_init(struct blkg_policy_data *pd)
1654 {
1655         struct cfq_group *cfqg = pd_to_cfqg(pd);
1656         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1657
1658         cfqg->weight = cgd->weight;
1659         cfqg->leaf_weight = cgd->leaf_weight;
1660 }
1661
1662 static void cfq_pd_offline(struct blkg_policy_data *pd)
1663 {
1664         struct cfq_group *cfqg = pd_to_cfqg(pd);
1665         int i;
1666
1667         for (i = 0; i < IOPRIO_BE_NR; i++) {
1668                 if (cfqg->async_cfqq[0][i])
1669                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1670                 if (cfqg->async_cfqq[1][i])
1671                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1672         }
1673
1674         if (cfqg->async_idle_cfqq)
1675                 cfq_put_queue(cfqg->async_idle_cfqq);
1676
1677         /*
1678          * @blkg is going offline and will be ignored by
1679          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1680          * that they don't get lost.  If IOs complete after this point, the
1681          * stats for them will be lost.  Oh well...
1682          */
1683         cfqg_stats_xfer_dead(cfqg);
1684 }
1685
1686 static void cfq_pd_free(struct blkg_policy_data *pd)
1687 {
1688         struct cfq_group *cfqg = pd_to_cfqg(pd);
1689
1690         cfqg_stats_exit(&cfqg->stats);
1691         return kfree(cfqg);
1692 }
1693
1694 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1695 {
1696         struct cfq_group *cfqg = pd_to_cfqg(pd);
1697
1698         cfqg_stats_reset(&cfqg->stats);
1699 }
1700
1701 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1702                                          struct blkcg *blkcg)
1703 {
1704         struct blkcg_gq *blkg;
1705
1706         blkg = blkg_lookup(blkcg, cfqd->queue);
1707         if (likely(blkg))
1708                 return blkg_to_cfqg(blkg);
1709         return NULL;
1710 }
1711
1712 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1713 {
1714         cfqq->cfqg = cfqg;
1715         /* cfqq reference on cfqg */
1716         cfqg_get(cfqg);
1717 }
1718
1719 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1720                                      struct blkg_policy_data *pd, int off)
1721 {
1722         struct cfq_group *cfqg = pd_to_cfqg(pd);
1723
1724         if (!cfqg->dev_weight)
1725                 return 0;
1726         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1727 }
1728
1729 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1730 {
1731         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1732                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1733                           0, false);
1734         return 0;
1735 }
1736
1737 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1738                                           struct blkg_policy_data *pd, int off)
1739 {
1740         struct cfq_group *cfqg = pd_to_cfqg(pd);
1741
1742         if (!cfqg->dev_leaf_weight)
1743                 return 0;
1744         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1745 }
1746
1747 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1748 {
1749         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1750                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1751                           0, false);
1752         return 0;
1753 }
1754
1755 static int cfq_print_weight(struct seq_file *sf, void *v)
1756 {
1757         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1758         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1759         unsigned int val = 0;
1760
1761         if (cgd)
1762                 val = cgd->weight;
1763
1764         seq_printf(sf, "%u\n", val);
1765         return 0;
1766 }
1767
1768 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1769 {
1770         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1771         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1772         unsigned int val = 0;
1773
1774         if (cgd)
1775                 val = cgd->leaf_weight;
1776
1777         seq_printf(sf, "%u\n", val);
1778         return 0;
1779 }
1780
1781 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1782                                         char *buf, size_t nbytes, loff_t off,
1783                                         bool on_dfl, bool is_leaf_weight)
1784 {
1785         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1786         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1787         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1788         struct blkg_conf_ctx ctx;
1789         struct cfq_group *cfqg;
1790         struct cfq_group_data *cfqgd;
1791         int ret;
1792         u64 v;
1793
1794         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1795         if (ret)
1796                 return ret;
1797
1798         if (sscanf(ctx.body, "%llu", &v) == 1) {
1799                 /* require "default" on dfl */
1800                 ret = -ERANGE;
1801                 if (!v && on_dfl)
1802                         goto out_finish;
1803         } else if (!strcmp(strim(ctx.body), "default")) {
1804                 v = 0;
1805         } else {
1806                 ret = -EINVAL;
1807                 goto out_finish;
1808         }
1809
1810         cfqg = blkg_to_cfqg(ctx.blkg);
1811         cfqgd = blkcg_to_cfqgd(blkcg);
1812
1813         ret = -ERANGE;
1814         if (!v || (v >= min && v <= max)) {
1815                 if (!is_leaf_weight) {
1816                         cfqg->dev_weight = v;
1817                         cfqg->new_weight = v ?: cfqgd->weight;
1818                 } else {
1819                         cfqg->dev_leaf_weight = v;
1820                         cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1821                 }
1822                 ret = 0;
1823         }
1824 out_finish:
1825         blkg_conf_finish(&ctx);
1826         return ret ?: nbytes;
1827 }
1828
1829 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1830                                       char *buf, size_t nbytes, loff_t off)
1831 {
1832         return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1833 }
1834
1835 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1836                                            char *buf, size_t nbytes, loff_t off)
1837 {
1838         return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1839 }
1840
1841 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1842                             bool on_dfl, bool reset_dev, bool is_leaf_weight)
1843 {
1844         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1845         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1846         struct blkcg *blkcg = css_to_blkcg(css);
1847         struct blkcg_gq *blkg;
1848         struct cfq_group_data *cfqgd;
1849         int ret = 0;
1850
1851         if (val < min || val > max)
1852                 return -ERANGE;
1853
1854         spin_lock_irq(&blkcg->lock);
1855         cfqgd = blkcg_to_cfqgd(blkcg);
1856         if (!cfqgd) {
1857                 ret = -EINVAL;
1858                 goto out;
1859         }
1860
1861         if (!is_leaf_weight)
1862                 cfqgd->weight = val;
1863         else
1864                 cfqgd->leaf_weight = val;
1865
1866         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1867                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1868
1869                 if (!cfqg)
1870                         continue;
1871
1872                 if (!is_leaf_weight) {
1873                         if (reset_dev)
1874                                 cfqg->dev_weight = 0;
1875                         if (!cfqg->dev_weight)
1876                                 cfqg->new_weight = cfqgd->weight;
1877                 } else {
1878                         if (reset_dev)
1879                                 cfqg->dev_leaf_weight = 0;
1880                         if (!cfqg->dev_leaf_weight)
1881                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1882                 }
1883         }
1884
1885 out:
1886         spin_unlock_irq(&blkcg->lock);
1887         return ret;
1888 }
1889
1890 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1891                           u64 val)
1892 {
1893         return __cfq_set_weight(css, val, false, false, false);
1894 }
1895
1896 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1897                                struct cftype *cft, u64 val)
1898 {
1899         return __cfq_set_weight(css, val, false, false, true);
1900 }
1901
1902 static int cfqg_print_stat(struct seq_file *sf, void *v)
1903 {
1904         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1905                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1906         return 0;
1907 }
1908
1909 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1910 {
1911         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1912                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1913         return 0;
1914 }
1915
1916 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1917                                       struct blkg_policy_data *pd, int off)
1918 {
1919         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1920                                           &blkcg_policy_cfq, off);
1921         return __blkg_prfill_u64(sf, pd, sum);
1922 }
1923
1924 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1925                                         struct blkg_policy_data *pd, int off)
1926 {
1927         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1928                                                         &blkcg_policy_cfq, off);
1929         return __blkg_prfill_rwstat(sf, pd, &sum);
1930 }
1931
1932 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1933 {
1934         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1935                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1936                           seq_cft(sf)->private, false);
1937         return 0;
1938 }
1939
1940 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1941 {
1942         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1943                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1944                           seq_cft(sf)->private, true);
1945         return 0;
1946 }
1947
1948 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1949                                int off)
1950 {
1951         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1952
1953         return __blkg_prfill_u64(sf, pd, sum >> 9);
1954 }
1955
1956 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1957 {
1958         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1959                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1960         return 0;
1961 }
1962
1963 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1964                                          struct blkg_policy_data *pd, int off)
1965 {
1966         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1967                                         offsetof(struct blkcg_gq, stat_bytes));
1968         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1969                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1970
1971         return __blkg_prfill_u64(sf, pd, sum >> 9);
1972 }
1973
1974 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1975 {
1976         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1977                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1978                           false);
1979         return 0;
1980 }
1981
1982 #ifdef CONFIG_DEBUG_BLK_CGROUP
1983 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1984                                       struct blkg_policy_data *pd, int off)
1985 {
1986         struct cfq_group *cfqg = pd_to_cfqg(pd);
1987         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1988         u64 v = 0;
1989
1990         if (samples) {
1991                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1992                 v = div64_u64(v, samples);
1993         }
1994         __blkg_prfill_u64(sf, pd, v);
1995         return 0;
1996 }
1997
1998 /* print avg_queue_size */
1999 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
2000 {
2001         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
2002                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
2003                           0, false);
2004         return 0;
2005 }
2006 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2007
2008 static struct cftype cfq_blkcg_legacy_files[] = {
2009         /* on root, weight is mapped to leaf_weight */
2010         {
2011                 .name = "weight_device",
2012                 .flags = CFTYPE_ONLY_ON_ROOT,
2013                 .seq_show = cfqg_print_leaf_weight_device,
2014                 .write = cfqg_set_leaf_weight_device,
2015         },
2016         {
2017                 .name = "weight",
2018                 .flags = CFTYPE_ONLY_ON_ROOT,
2019                 .seq_show = cfq_print_leaf_weight,
2020                 .write_u64 = cfq_set_leaf_weight,
2021         },
2022
2023         /* no such mapping necessary for !roots */
2024         {
2025                 .name = "weight_device",
2026                 .flags = CFTYPE_NOT_ON_ROOT,
2027                 .seq_show = cfqg_print_weight_device,
2028                 .write = cfqg_set_weight_device,
2029         },
2030         {
2031                 .name = "weight",
2032                 .flags = CFTYPE_NOT_ON_ROOT,
2033                 .seq_show = cfq_print_weight,
2034                 .write_u64 = cfq_set_weight,
2035         },
2036
2037         {
2038                 .name = "leaf_weight_device",
2039                 .seq_show = cfqg_print_leaf_weight_device,
2040                 .write = cfqg_set_leaf_weight_device,
2041         },
2042         {
2043                 .name = "leaf_weight",
2044                 .seq_show = cfq_print_leaf_weight,
2045                 .write_u64 = cfq_set_leaf_weight,
2046         },
2047
2048         /* statistics, covers only the tasks in the cfqg */
2049         {
2050                 .name = "time",
2051                 .private = offsetof(struct cfq_group, stats.time),
2052                 .seq_show = cfqg_print_stat,
2053         },
2054         {
2055                 .name = "sectors",
2056                 .seq_show = cfqg_print_stat_sectors,
2057         },
2058         {
2059                 .name = "io_service_bytes",
2060                 .private = (unsigned long)&blkcg_policy_cfq,
2061                 .seq_show = blkg_print_stat_bytes,
2062         },
2063         {
2064                 .name = "io_serviced",
2065                 .private = (unsigned long)&blkcg_policy_cfq,
2066                 .seq_show = blkg_print_stat_ios,
2067         },
2068         {
2069                 .name = "io_service_time",
2070                 .private = offsetof(struct cfq_group, stats.service_time),
2071                 .seq_show = cfqg_print_rwstat,
2072         },
2073         {
2074                 .name = "io_wait_time",
2075                 .private = offsetof(struct cfq_group, stats.wait_time),
2076                 .seq_show = cfqg_print_rwstat,
2077         },
2078         {
2079                 .name = "io_merged",
2080                 .private = offsetof(struct cfq_group, stats.merged),
2081                 .seq_show = cfqg_print_rwstat,
2082         },
2083         {
2084                 .name = "io_queued",
2085                 .private = offsetof(struct cfq_group, stats.queued),
2086                 .seq_show = cfqg_print_rwstat,
2087         },
2088
2089         /* the same statictics which cover the cfqg and its descendants */
2090         {
2091                 .name = "time_recursive",
2092                 .private = offsetof(struct cfq_group, stats.time),
2093                 .seq_show = cfqg_print_stat_recursive,
2094         },
2095         {
2096                 .name = "sectors_recursive",
2097                 .seq_show = cfqg_print_stat_sectors_recursive,
2098         },
2099         {
2100                 .name = "io_service_bytes_recursive",
2101                 .private = (unsigned long)&blkcg_policy_cfq,
2102                 .seq_show = blkg_print_stat_bytes_recursive,
2103         },
2104         {
2105                 .name = "io_serviced_recursive",
2106                 .private = (unsigned long)&blkcg_policy_cfq,
2107                 .seq_show = blkg_print_stat_ios_recursive,
2108         },
2109         {
2110                 .name = "io_service_time_recursive",
2111                 .private = offsetof(struct cfq_group, stats.service_time),
2112                 .seq_show = cfqg_print_rwstat_recursive,
2113         },
2114         {
2115                 .name = "io_wait_time_recursive",
2116                 .private = offsetof(struct cfq_group, stats.wait_time),
2117                 .seq_show = cfqg_print_rwstat_recursive,
2118         },
2119         {
2120                 .name = "io_merged_recursive",
2121                 .private = offsetof(struct cfq_group, stats.merged),
2122                 .seq_show = cfqg_print_rwstat_recursive,
2123         },
2124         {
2125                 .name = "io_queued_recursive",
2126                 .private = offsetof(struct cfq_group, stats.queued),
2127                 .seq_show = cfqg_print_rwstat_recursive,
2128         },
2129 #ifdef CONFIG_DEBUG_BLK_CGROUP
2130         {
2131                 .name = "avg_queue_size",
2132                 .seq_show = cfqg_print_avg_queue_size,
2133         },
2134         {
2135                 .name = "group_wait_time",
2136                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2137                 .seq_show = cfqg_print_stat,
2138         },
2139         {
2140                 .name = "idle_time",
2141                 .private = offsetof(struct cfq_group, stats.idle_time),
2142                 .seq_show = cfqg_print_stat,
2143         },
2144         {
2145                 .name = "empty_time",
2146                 .private = offsetof(struct cfq_group, stats.empty_time),
2147                 .seq_show = cfqg_print_stat,
2148         },
2149         {
2150                 .name = "dequeue",
2151                 .private = offsetof(struct cfq_group, stats.dequeue),
2152                 .seq_show = cfqg_print_stat,
2153         },
2154         {
2155                 .name = "unaccounted_time",
2156                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2157                 .seq_show = cfqg_print_stat,
2158         },
2159 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2160         { }     /* terminate */
2161 };
2162
2163 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2164 {
2165         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2166         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2167
2168         seq_printf(sf, "default %u\n", cgd->weight);
2169         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2170                           &blkcg_policy_cfq, 0, false);
2171         return 0;
2172 }
2173
2174 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2175                                      char *buf, size_t nbytes, loff_t off)
2176 {
2177         char *endp;
2178         int ret;
2179         u64 v;
2180
2181         buf = strim(buf);
2182
2183         /* "WEIGHT" or "default WEIGHT" sets the default weight */
2184         v = simple_strtoull(buf, &endp, 0);
2185         if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2186                 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2187                 return ret ?: nbytes;
2188         }
2189
2190         /* "MAJ:MIN WEIGHT" */
2191         return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2192 }
2193
2194 static struct cftype cfq_blkcg_files[] = {
2195         {
2196                 .name = "weight",
2197                 .flags = CFTYPE_NOT_ON_ROOT,
2198                 .seq_show = cfq_print_weight_on_dfl,
2199                 .write = cfq_set_weight_on_dfl,
2200         },
2201         { }     /* terminate */
2202 };
2203
2204 #else /* GROUP_IOSCHED */
2205 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2206                                          struct blkcg *blkcg)
2207 {
2208         return cfqd->root_group;
2209 }
2210
2211 static inline void
2212 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2213         cfqq->cfqg = cfqg;
2214 }
2215
2216 #endif /* GROUP_IOSCHED */
2217
2218 /*
2219  * The cfqd->service_trees holds all pending cfq_queue's that have
2220  * requests waiting to be processed. It is sorted in the order that
2221  * we will service the queues.
2222  */
2223 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2224                                  bool add_front)
2225 {
2226         struct rb_node **p, *parent;
2227         struct cfq_queue *__cfqq;
2228         u64 rb_key;
2229         struct cfq_rb_root *st;
2230         int left;
2231         int new_cfqq = 1;
2232         u64 now = ktime_get_ns();
2233
2234         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2235         if (cfq_class_idle(cfqq)) {
2236                 rb_key = CFQ_IDLE_DELAY;
2237                 parent = rb_last(&st->rb);
2238                 if (parent && parent != &cfqq->rb_node) {
2239                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2240                         rb_key += __cfqq->rb_key;
2241                 } else
2242                         rb_key += now;
2243         } else if (!add_front) {
2244                 /*
2245                  * Get our rb key offset. Subtract any residual slice
2246                  * value carried from last service. A negative resid
2247                  * count indicates slice overrun, and this should position
2248                  * the next service time further away in the tree.
2249                  */
2250                 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2251                 rb_key -= cfqq->slice_resid;
2252                 cfqq->slice_resid = 0;
2253         } else {
2254                 rb_key = -NSEC_PER_SEC;
2255                 __cfqq = cfq_rb_first(st);
2256                 rb_key += __cfqq ? __cfqq->rb_key : now;
2257         }
2258
2259         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2260                 new_cfqq = 0;
2261                 /*
2262                  * same position, nothing more to do
2263                  */
2264                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2265                         return;
2266
2267                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2268                 cfqq->service_tree = NULL;
2269         }
2270
2271         left = 1;
2272         parent = NULL;
2273         cfqq->service_tree = st;
2274         p = &st->rb.rb_node;
2275         while (*p) {
2276                 parent = *p;
2277                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2278
2279                 /*
2280                  * sort by key, that represents service time.
2281                  */
2282                 if (rb_key < __cfqq->rb_key)
2283                         p = &parent->rb_left;
2284                 else {
2285                         p = &parent->rb_right;
2286                         left = 0;
2287                 }
2288         }
2289
2290         if (left)
2291                 st->left = &cfqq->rb_node;
2292
2293         cfqq->rb_key = rb_key;
2294         rb_link_node(&cfqq->rb_node, parent, p);
2295         rb_insert_color(&cfqq->rb_node, &st->rb);
2296         st->count++;
2297         if (add_front || !new_cfqq)
2298                 return;
2299         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2300 }
2301
2302 static struct cfq_queue *
2303 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2304                      sector_t sector, struct rb_node **ret_parent,
2305                      struct rb_node ***rb_link)
2306 {
2307         struct rb_node **p, *parent;
2308         struct cfq_queue *cfqq = NULL;
2309
2310         parent = NULL;
2311         p = &root->rb_node;
2312         while (*p) {
2313                 struct rb_node **n;
2314
2315                 parent = *p;
2316                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2317
2318                 /*
2319                  * Sort strictly based on sector.  Smallest to the left,
2320                  * largest to the right.
2321                  */
2322                 if (sector > blk_rq_pos(cfqq->next_rq))
2323                         n = &(*p)->rb_right;
2324                 else if (sector < blk_rq_pos(cfqq->next_rq))
2325                         n = &(*p)->rb_left;
2326                 else
2327                         break;
2328                 p = n;
2329                 cfqq = NULL;
2330         }
2331
2332         *ret_parent = parent;
2333         if (rb_link)
2334                 *rb_link = p;
2335         return cfqq;
2336 }
2337
2338 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2339 {
2340         struct rb_node **p, *parent;
2341         struct cfq_queue *__cfqq;
2342
2343         if (cfqq->p_root) {
2344                 rb_erase(&cfqq->p_node, cfqq->p_root);
2345                 cfqq->p_root = NULL;
2346         }
2347
2348         if (cfq_class_idle(cfqq))
2349                 return;
2350         if (!cfqq->next_rq)
2351                 return;
2352
2353         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2354         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2355                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2356         if (!__cfqq) {
2357                 rb_link_node(&cfqq->p_node, parent, p);
2358                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2359         } else
2360                 cfqq->p_root = NULL;
2361 }
2362
2363 /*
2364  * Update cfqq's position in the service tree.
2365  */
2366 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2367 {
2368         /*
2369          * Resorting requires the cfqq to be on the RR list already.
2370          */
2371         if (cfq_cfqq_on_rr(cfqq)) {
2372                 cfq_service_tree_add(cfqd, cfqq, 0);
2373                 cfq_prio_tree_add(cfqd, cfqq);
2374         }
2375 }
2376
2377 /*
2378  * add to busy list of queues for service, trying to be fair in ordering
2379  * the pending list according to last request service
2380  */
2381 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2382 {
2383         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2384         BUG_ON(cfq_cfqq_on_rr(cfqq));
2385         cfq_mark_cfqq_on_rr(cfqq);
2386         cfqd->busy_queues++;
2387         if (cfq_cfqq_sync(cfqq))
2388                 cfqd->busy_sync_queues++;
2389
2390         cfq_resort_rr_list(cfqd, cfqq);
2391 }
2392
2393 /*
2394  * Called when the cfqq no longer has requests pending, remove it from
2395  * the service tree.
2396  */
2397 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2398 {
2399         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2400         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2401         cfq_clear_cfqq_on_rr(cfqq);
2402
2403         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2404                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2405                 cfqq->service_tree = NULL;
2406         }
2407         if (cfqq->p_root) {
2408                 rb_erase(&cfqq->p_node, cfqq->p_root);
2409                 cfqq->p_root = NULL;
2410         }
2411
2412         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2413         BUG_ON(!cfqd->busy_queues);
2414         cfqd->busy_queues--;
2415         if (cfq_cfqq_sync(cfqq))
2416                 cfqd->busy_sync_queues--;
2417 }
2418
2419 /*
2420  * rb tree support functions
2421  */
2422 static void cfq_del_rq_rb(struct request *rq)
2423 {
2424         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2425         const int sync = rq_is_sync(rq);
2426
2427         BUG_ON(!cfqq->queued[sync]);
2428         cfqq->queued[sync]--;
2429
2430         elv_rb_del(&cfqq->sort_list, rq);
2431
2432         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2433                 /*
2434                  * Queue will be deleted from service tree when we actually
2435                  * expire it later. Right now just remove it from prio tree
2436                  * as it is empty.
2437                  */
2438                 if (cfqq->p_root) {
2439                         rb_erase(&cfqq->p_node, cfqq->p_root);
2440                         cfqq->p_root = NULL;
2441                 }
2442         }
2443 }
2444
2445 static void cfq_add_rq_rb(struct request *rq)
2446 {
2447         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2448         struct cfq_data *cfqd = cfqq->cfqd;
2449         struct request *prev;
2450
2451         cfqq->queued[rq_is_sync(rq)]++;
2452
2453         elv_rb_add(&cfqq->sort_list, rq);
2454
2455         if (!cfq_cfqq_on_rr(cfqq))
2456                 cfq_add_cfqq_rr(cfqd, cfqq);
2457
2458         /*
2459          * check if this request is a better next-serve candidate
2460          */
2461         prev = cfqq->next_rq;
2462         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2463
2464         /*
2465          * adjust priority tree position, if ->next_rq changes
2466          */
2467         if (prev != cfqq->next_rq)
2468                 cfq_prio_tree_add(cfqd, cfqq);
2469
2470         BUG_ON(!cfqq->next_rq);
2471 }
2472
2473 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2474 {
2475         elv_rb_del(&cfqq->sort_list, rq);
2476         cfqq->queued[rq_is_sync(rq)]--;
2477         cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
2478         cfq_add_rq_rb(rq);
2479         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2480                                  req_op(rq), rq->cmd_flags);
2481 }
2482
2483 static struct request *
2484 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2485 {
2486         struct task_struct *tsk = current;
2487         struct cfq_io_cq *cic;
2488         struct cfq_queue *cfqq;
2489
2490         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2491         if (!cic)
2492                 return NULL;
2493
2494         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2495         if (cfqq)
2496                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2497
2498         return NULL;
2499 }
2500
2501 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2502 {
2503         struct cfq_data *cfqd = q->elevator->elevator_data;
2504
2505         cfqd->rq_in_driver++;
2506         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2507                                                 cfqd->rq_in_driver);
2508
2509         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2510 }
2511
2512 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2513 {
2514         struct cfq_data *cfqd = q->elevator->elevator_data;
2515
2516         WARN_ON(!cfqd->rq_in_driver);
2517         cfqd->rq_in_driver--;
2518         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2519                                                 cfqd->rq_in_driver);
2520 }
2521
2522 static void cfq_remove_request(struct request *rq)
2523 {
2524         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2525
2526         if (cfqq->next_rq == rq)
2527                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2528
2529         list_del_init(&rq->queuelist);
2530         cfq_del_rq_rb(rq);
2531
2532         cfqq->cfqd->rq_queued--;
2533         cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
2534         if (rq->cmd_flags & REQ_PRIO) {
2535                 WARN_ON(!cfqq->prio_pending);
2536                 cfqq->prio_pending--;
2537         }
2538 }
2539
2540 static int cfq_merge(struct request_queue *q, struct request **req,
2541                      struct bio *bio)
2542 {
2543         struct cfq_data *cfqd = q->elevator->elevator_data;
2544         struct request *__rq;
2545
2546         __rq = cfq_find_rq_fmerge(cfqd, bio);
2547         if (__rq && elv_bio_merge_ok(__rq, bio)) {
2548                 *req = __rq;
2549                 return ELEVATOR_FRONT_MERGE;
2550         }
2551
2552         return ELEVATOR_NO_MERGE;
2553 }
2554
2555 static void cfq_merged_request(struct request_queue *q, struct request *req,
2556                                int type)
2557 {
2558         if (type == ELEVATOR_FRONT_MERGE) {
2559                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2560
2561                 cfq_reposition_rq_rb(cfqq, req);
2562         }
2563 }
2564
2565 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2566                                 struct bio *bio)
2567 {
2568         cfqg_stats_update_io_merged(RQ_CFQG(req), bio_op(bio), bio->bi_opf);
2569 }
2570
2571 static void
2572 cfq_merged_requests(struct request_queue *q, struct request *rq,
2573                     struct request *next)
2574 {
2575         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2576         struct cfq_data *cfqd = q->elevator->elevator_data;
2577
2578         /*
2579          * reposition in fifo if next is older than rq
2580          */
2581         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2582             next->fifo_time < rq->fifo_time &&
2583             cfqq == RQ_CFQQ(next)) {
2584                 list_move(&rq->queuelist, &next->queuelist);
2585                 rq->fifo_time = next->fifo_time;
2586         }
2587
2588         if (cfqq->next_rq == next)
2589                 cfqq->next_rq = rq;
2590         cfq_remove_request(next);
2591         cfqg_stats_update_io_merged(RQ_CFQG(rq), req_op(next), next->cmd_flags);
2592
2593         cfqq = RQ_CFQQ(next);
2594         /*
2595          * all requests of this queue are merged to other queues, delete it
2596          * from the service tree. If it's the active_queue,
2597          * cfq_dispatch_requests() will choose to expire it or do idle
2598          */
2599         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2600             cfqq != cfqd->active_queue)
2601                 cfq_del_cfqq_rr(cfqd, cfqq);
2602 }
2603
2604 static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2605                                struct bio *bio)
2606 {
2607         struct cfq_data *cfqd = q->elevator->elevator_data;
2608         struct cfq_io_cq *cic;
2609         struct cfq_queue *cfqq;
2610
2611         /*
2612          * Disallow merge of a sync bio into an async request.
2613          */
2614         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2615                 return false;
2616
2617         /*
2618          * Lookup the cfqq that this bio will be queued with and allow
2619          * merge only if rq is queued there.
2620          */
2621         cic = cfq_cic_lookup(cfqd, current->io_context);
2622         if (!cic)
2623                 return false;
2624
2625         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2626         return cfqq == RQ_CFQQ(rq);
2627 }
2628
2629 static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2630                               struct request *next)
2631 {
2632         return RQ_CFQQ(rq) == RQ_CFQQ(next);
2633 }
2634
2635 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2636 {
2637         hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2638         cfqg_stats_update_idle_time(cfqq->cfqg);
2639 }
2640
2641 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2642                                    struct cfq_queue *cfqq)
2643 {
2644         if (cfqq) {
2645                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2646                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2647                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2648                 cfqq->slice_start = 0;
2649                 cfqq->dispatch_start = ktime_get_ns();
2650                 cfqq->allocated_slice = 0;
2651                 cfqq->slice_end = 0;
2652                 cfqq->slice_dispatch = 0;
2653                 cfqq->nr_sectors = 0;
2654
2655                 cfq_clear_cfqq_wait_request(cfqq);
2656                 cfq_clear_cfqq_must_dispatch(cfqq);
2657                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2658                 cfq_clear_cfqq_fifo_expire(cfqq);
2659                 cfq_mark_cfqq_slice_new(cfqq);
2660
2661                 cfq_del_timer(cfqd, cfqq);
2662         }
2663
2664         cfqd->active_queue = cfqq;
2665 }
2666
2667 /*
2668  * current cfqq expired its slice (or was too idle), select new one
2669  */
2670 static void
2671 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2672                     bool timed_out)
2673 {
2674         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2675
2676         if (cfq_cfqq_wait_request(cfqq))
2677                 cfq_del_timer(cfqd, cfqq);
2678
2679         cfq_clear_cfqq_wait_request(cfqq);
2680         cfq_clear_cfqq_wait_busy(cfqq);
2681
2682         /*
2683          * If this cfqq is shared between multiple processes, check to
2684          * make sure that those processes are still issuing I/Os within
2685          * the mean seek distance.  If not, it may be time to break the
2686          * queues apart again.
2687          */
2688         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2689                 cfq_mark_cfqq_split_coop(cfqq);
2690
2691         /*
2692          * store what was left of this slice, if the queue idled/timed out
2693          */
2694         if (timed_out) {
2695                 if (cfq_cfqq_slice_new(cfqq))
2696                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2697                 else
2698                         cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2699                 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2700         }
2701
2702         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2703
2704         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2705                 cfq_del_cfqq_rr(cfqd, cfqq);
2706
2707         cfq_resort_rr_list(cfqd, cfqq);
2708
2709         if (cfqq == cfqd->active_queue)
2710                 cfqd->active_queue = NULL;
2711
2712         if (cfqd->active_cic) {
2713                 put_io_context(cfqd->active_cic->icq.ioc);
2714                 cfqd->active_cic = NULL;
2715         }
2716 }
2717
2718 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2719 {
2720         struct cfq_queue *cfqq = cfqd->active_queue;
2721
2722         if (cfqq)
2723                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2724 }
2725
2726 /*
2727  * Get next queue for service. Unless we have a queue preemption,
2728  * we'll simply select the first cfqq in the service tree.
2729  */
2730 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2731 {
2732         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2733                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2734
2735         if (!cfqd->rq_queued)
2736                 return NULL;
2737
2738         /* There is nothing to dispatch */
2739         if (!st)
2740                 return NULL;
2741         if (RB_EMPTY_ROOT(&st->rb))
2742                 return NULL;
2743         return cfq_rb_first(st);
2744 }
2745
2746 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2747 {
2748         struct cfq_group *cfqg;
2749         struct cfq_queue *cfqq;
2750         int i, j;
2751         struct cfq_rb_root *st;
2752
2753         if (!cfqd->rq_queued)
2754                 return NULL;
2755
2756         cfqg = cfq_get_next_cfqg(cfqd);
2757         if (!cfqg)
2758                 return NULL;
2759
2760         for_each_cfqg_st(cfqg, i, j, st)
2761                 if ((cfqq = cfq_rb_first(st)) != NULL)
2762                         return cfqq;
2763         return NULL;
2764 }
2765
2766 /*
2767  * Get and set a new active queue for service.
2768  */
2769 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2770                                               struct cfq_queue *cfqq)
2771 {
2772         if (!cfqq)
2773                 cfqq = cfq_get_next_queue(cfqd);
2774
2775         __cfq_set_active_queue(cfqd, cfqq);
2776         return cfqq;
2777 }
2778
2779 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2780                                           struct request *rq)
2781 {
2782         if (blk_rq_pos(rq) >= cfqd->last_position)
2783                 return blk_rq_pos(rq) - cfqd->last_position;
2784         else
2785                 return cfqd->last_position - blk_rq_pos(rq);
2786 }
2787
2788 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2789                                struct request *rq)
2790 {
2791         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2792 }
2793
2794 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2795                                     struct cfq_queue *cur_cfqq)
2796 {
2797         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2798         struct rb_node *parent, *node;
2799         struct cfq_queue *__cfqq;
2800         sector_t sector = cfqd->last_position;
2801
2802         if (RB_EMPTY_ROOT(root))
2803                 return NULL;
2804
2805         /*
2806          * First, if we find a request starting at the end of the last
2807          * request, choose it.
2808          */
2809         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2810         if (__cfqq)
2811                 return __cfqq;
2812
2813         /*
2814          * If the exact sector wasn't found, the parent of the NULL leaf
2815          * will contain the closest sector.
2816          */
2817         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2818         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2819                 return __cfqq;
2820
2821         if (blk_rq_pos(__cfqq->next_rq) < sector)
2822                 node = rb_next(&__cfqq->p_node);
2823         else
2824                 node = rb_prev(&__cfqq->p_node);
2825         if (!node)
2826                 return NULL;
2827
2828         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2829         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2830                 return __cfqq;
2831
2832         return NULL;
2833 }
2834
2835 /*
2836  * cfqd - obvious
2837  * cur_cfqq - passed in so that we don't decide that the current queue is
2838  *            closely cooperating with itself.
2839  *
2840  * So, basically we're assuming that that cur_cfqq has dispatched at least
2841  * one request, and that cfqd->last_position reflects a position on the disk
2842  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2843  * assumption.
2844  */
2845 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2846                                               struct cfq_queue *cur_cfqq)
2847 {
2848         struct cfq_queue *cfqq;
2849
2850         if (cfq_class_idle(cur_cfqq))
2851                 return NULL;
2852         if (!cfq_cfqq_sync(cur_cfqq))
2853                 return NULL;
2854         if (CFQQ_SEEKY(cur_cfqq))
2855                 return NULL;
2856
2857         /*
2858          * Don't search priority tree if it's the only queue in the group.
2859          */
2860         if (cur_cfqq->cfqg->nr_cfqq == 1)
2861                 return NULL;
2862
2863         /*
2864          * We should notice if some of the queues are cooperating, eg
2865          * working closely on the same area of the disk. In that case,
2866          * we can group them together and don't waste time idling.
2867          */
2868         cfqq = cfqq_close(cfqd, cur_cfqq);
2869         if (!cfqq)
2870                 return NULL;
2871
2872         /* If new queue belongs to different cfq_group, don't choose it */
2873         if (cur_cfqq->cfqg != cfqq->cfqg)
2874                 return NULL;
2875
2876         /*
2877          * It only makes sense to merge sync queues.
2878          */
2879         if (!cfq_cfqq_sync(cfqq))
2880                 return NULL;
2881         if (CFQQ_SEEKY(cfqq))
2882                 return NULL;
2883
2884         /*
2885          * Do not merge queues of different priority classes
2886          */
2887         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2888                 return NULL;
2889
2890         return cfqq;
2891 }
2892
2893 /*
2894  * Determine whether we should enforce idle window for this queue.
2895  */
2896
2897 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2898 {
2899         enum wl_class_t wl_class = cfqq_class(cfqq);
2900         struct cfq_rb_root *st = cfqq->service_tree;
2901
2902         BUG_ON(!st);
2903         BUG_ON(!st->count);
2904
2905         if (!cfqd->cfq_slice_idle)
2906                 return false;
2907
2908         /* We never do for idle class queues. */
2909         if (wl_class == IDLE_WORKLOAD)
2910                 return false;
2911
2912         /* We do for queues that were marked with idle window flag. */
2913         if (cfq_cfqq_idle_window(cfqq) &&
2914            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2915                 return true;
2916
2917         /*
2918          * Otherwise, we do only if they are the last ones
2919          * in their service tree.
2920          */
2921         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2922            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2923                 return true;
2924         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2925         return false;
2926 }
2927
2928 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2929 {
2930         struct cfq_queue *cfqq = cfqd->active_queue;
2931         struct cfq_rb_root *st = cfqq->service_tree;
2932         struct cfq_io_cq *cic;
2933         u64 sl, group_idle = 0;
2934         u64 now = ktime_get_ns();
2935
2936         /*
2937          * SSD device without seek penalty, disable idling. But only do so
2938          * for devices that support queuing, otherwise we still have a problem
2939          * with sync vs async workloads.
2940          */
2941         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2942                 return;
2943
2944         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2945         WARN_ON(cfq_cfqq_slice_new(cfqq));
2946
2947         /*
2948          * idle is disabled, either manually or by past process history
2949          */
2950         if (!cfq_should_idle(cfqd, cfqq)) {
2951                 /* no queue idling. Check for group idling */
2952                 if (cfqd->cfq_group_idle)
2953                         group_idle = cfqd->cfq_group_idle;
2954                 else
2955                         return;
2956         }
2957
2958         /*
2959          * still active requests from this queue, don't idle
2960          */
2961         if (cfqq->dispatched)
2962                 return;
2963
2964         /*
2965          * task has exited, don't wait
2966          */
2967         cic = cfqd->active_cic;
2968         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2969                 return;
2970
2971         /*
2972          * If our average think time is larger than the remaining time
2973          * slice, then don't idle. This avoids overrunning the allotted
2974          * time slice.
2975          */
2976         if (sample_valid(cic->ttime.ttime_samples) &&
2977             (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2978                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2979                              cic->ttime.ttime_mean);
2980                 return;
2981         }
2982
2983         /*
2984          * There are other queues in the group or this is the only group and
2985          * it has too big thinktime, don't do group idle.
2986          */
2987         if (group_idle &&
2988             (cfqq->cfqg->nr_cfqq > 1 ||
2989              cfq_io_thinktime_big(cfqd, &st->ttime, true)))
2990                 return;
2991
2992         cfq_mark_cfqq_wait_request(cfqq);
2993
2994         if (group_idle)
2995                 sl = cfqd->cfq_group_idle;
2996         else
2997                 sl = cfqd->cfq_slice_idle;
2998
2999         hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
3000                       HRTIMER_MODE_REL);
3001         cfqg_stats_set_start_idle_time(cfqq->cfqg);
3002         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
3003                         group_idle ? 1 : 0);
3004 }
3005
3006 /*
3007  * Move request from internal lists to the request queue dispatch list.
3008  */
3009 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
3010 {
3011         struct cfq_data *cfqd = q->elevator->elevator_data;
3012         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3013
3014         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3015
3016         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
3017         cfq_remove_request(rq);
3018         cfqq->dispatched++;
3019         (RQ_CFQG(rq))->dispatched++;
3020         elv_dispatch_sort(q, rq);
3021
3022         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
3023         cfqq->nr_sectors += blk_rq_sectors(rq);
3024 }
3025
3026 /*
3027  * return expired entry, or NULL to just start from scratch in rbtree
3028  */
3029 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3030 {
3031         struct request *rq = NULL;
3032
3033         if (cfq_cfqq_fifo_expire(cfqq))
3034                 return NULL;
3035
3036         cfq_mark_cfqq_fifo_expire(cfqq);
3037
3038         if (list_empty(&cfqq->fifo))
3039                 return NULL;
3040
3041         rq = rq_entry_fifo(cfqq->fifo.next);
3042         if (ktime_get_ns() < rq->fifo_time)
3043                 rq = NULL;
3044
3045         return rq;
3046 }
3047
3048 static inline int
3049 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3050 {
3051         const int base_rq = cfqd->cfq_slice_async_rq;
3052
3053         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3054
3055         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3056 }
3057
3058 /*
3059  * Must be called with the queue_lock held.
3060  */
3061 static int cfqq_process_refs(struct cfq_queue *cfqq)
3062 {
3063         int process_refs, io_refs;
3064
3065         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3066         process_refs = cfqq->ref - io_refs;
3067         BUG_ON(process_refs < 0);
3068         return process_refs;
3069 }
3070
3071 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3072 {
3073         int process_refs, new_process_refs;
3074         struct cfq_queue *__cfqq;
3075
3076         /*
3077          * If there are no process references on the new_cfqq, then it is
3078          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3079          * chain may have dropped their last reference (not just their
3080          * last process reference).
3081          */
3082         if (!cfqq_process_refs(new_cfqq))
3083                 return;
3084
3085         /* Avoid a circular list and skip interim queue merges */
3086         while ((__cfqq = new_cfqq->new_cfqq)) {
3087                 if (__cfqq == cfqq)
3088                         return;
3089                 new_cfqq = __cfqq;
3090         }
3091
3092         process_refs = cfqq_process_refs(cfqq);
3093         new_process_refs = cfqq_process_refs(new_cfqq);
3094         /*
3095          * If the process for the cfqq has gone away, there is no
3096          * sense in merging the queues.
3097          */
3098         if (process_refs == 0 || new_process_refs == 0)
3099                 return;
3100
3101         /*
3102          * Merge in the direction of the lesser amount of work.
3103          */
3104         if (new_process_refs >= process_refs) {
3105                 cfqq->new_cfqq = new_cfqq;
3106                 new_cfqq->ref += process_refs;
3107         } else {
3108                 new_cfqq->new_cfqq = cfqq;
3109                 cfqq->ref += new_process_refs;
3110         }
3111 }
3112
3113 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3114                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3115 {
3116         struct cfq_queue *queue;
3117         int i;
3118         bool key_valid = false;
3119         u64 lowest_key = 0;
3120         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3121
3122         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3123                 /* select the one with lowest rb_key */
3124                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3125                 if (queue &&
3126                     (!key_valid || queue->rb_key < lowest_key)) {
3127                         lowest_key = queue->rb_key;
3128                         cur_best = i;
3129                         key_valid = true;
3130                 }
3131         }
3132
3133         return cur_best;
3134 }
3135
3136 static void
3137 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3138 {
3139         u64 slice;
3140         unsigned count;
3141         struct cfq_rb_root *st;
3142         u64 group_slice;
3143         enum wl_class_t original_class = cfqd->serving_wl_class;
3144         u64 now = ktime_get_ns();
3145
3146         /* Choose next priority. RT > BE > IDLE */
3147         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3148                 cfqd->serving_wl_class = RT_WORKLOAD;
3149         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3150                 cfqd->serving_wl_class = BE_WORKLOAD;
3151         else {
3152                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3153                 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3154                 return;
3155         }
3156
3157         if (original_class != cfqd->serving_wl_class)
3158                 goto new_workload;
3159
3160         /*
3161          * For RT and BE, we have to choose also the type
3162          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3163          * expiration time
3164          */
3165         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3166         count = st->count;
3167
3168         /*
3169          * check workload expiration, and that we still have other queues ready
3170          */
3171         if (count && !(now > cfqd->workload_expires))
3172                 return;
3173
3174 new_workload:
3175         /* otherwise select new workload type */
3176         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3177                                         cfqd->serving_wl_class);
3178         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3179         count = st->count;
3180
3181         /*
3182          * the workload slice is computed as a fraction of target latency
3183          * proportional to the number of queues in that workload, over
3184          * all the queues in the same priority class
3185          */
3186         group_slice = cfq_group_slice(cfqd, cfqg);
3187
3188         slice = div_u64(group_slice * count,
3189                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3190                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3191                                         cfqg)));
3192
3193         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3194                 u64 tmp;
3195
3196                 /*
3197                  * Async queues are currently system wide. Just taking
3198                  * proportion of queues with-in same group will lead to higher
3199                  * async ratio system wide as generally root group is going
3200                  * to have higher weight. A more accurate thing would be to
3201                  * calculate system wide asnc/sync ratio.
3202                  */
3203                 tmp = cfqd->cfq_target_latency *
3204                         cfqg_busy_async_queues(cfqd, cfqg);
3205                 tmp = div_u64(tmp, cfqd->busy_queues);
3206                 slice = min_t(u64, slice, tmp);
3207
3208                 /* async workload slice is scaled down according to
3209                  * the sync/async slice ratio. */
3210                 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3211         } else
3212                 /* sync workload slice is at least 2 * cfq_slice_idle */
3213                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3214
3215         slice = max_t(u64, slice, CFQ_MIN_TT);
3216         cfq_log(cfqd, "workload slice:%llu", slice);
3217         cfqd->workload_expires = now + slice;
3218 }
3219
3220 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3221 {
3222         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3223         struct cfq_group *cfqg;
3224
3225         if (RB_EMPTY_ROOT(&st->rb))
3226                 return NULL;
3227         cfqg = cfq_rb_first_group(st);
3228         update_min_vdisktime(st);
3229         return cfqg;
3230 }
3231
3232 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3233 {
3234         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3235         u64 now = ktime_get_ns();
3236
3237         cfqd->serving_group = cfqg;
3238
3239         /* Restore the workload type data */
3240         if (cfqg->saved_wl_slice) {
3241                 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3242                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3243                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3244         } else
3245                 cfqd->workload_expires = now - 1;
3246
3247         choose_wl_class_and_type(cfqd, cfqg);
3248 }
3249
3250 /*
3251  * Select a queue for service. If we have a current active queue,
3252  * check whether to continue servicing it, or retrieve and set a new one.
3253  */
3254 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3255 {
3256         struct cfq_queue *cfqq, *new_cfqq = NULL;
3257         u64 now = ktime_get_ns();
3258
3259         cfqq = cfqd->active_queue;
3260         if (!cfqq)
3261                 goto new_queue;
3262
3263         if (!cfqd->rq_queued)
3264                 return NULL;
3265
3266         /*
3267          * We were waiting for group to get backlogged. Expire the queue
3268          */
3269         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3270                 goto expire;
3271
3272         /*
3273          * The active queue has run out of time, expire it and select new.
3274          */
3275         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3276                 /*
3277                  * If slice had not expired at the completion of last request
3278                  * we might not have turned on wait_busy flag. Don't expire
3279                  * the queue yet. Allow the group to get backlogged.
3280                  *
3281                  * The very fact that we have used the slice, that means we
3282                  * have been idling all along on this queue and it should be
3283                  * ok to wait for this request to complete.
3284                  */
3285                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3286                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3287                         cfqq = NULL;
3288                         goto keep_queue;
3289                 } else
3290                         goto check_group_idle;
3291         }
3292
3293         /*
3294          * The active queue has requests and isn't expired, allow it to
3295          * dispatch.
3296          */
3297         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3298                 goto keep_queue;
3299
3300         /*
3301          * If another queue has a request waiting within our mean seek
3302          * distance, let it run.  The expire code will check for close
3303          * cooperators and put the close queue at the front of the service
3304          * tree.  If possible, merge the expiring queue with the new cfqq.
3305          */
3306         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3307         if (new_cfqq) {
3308                 if (!cfqq->new_cfqq)
3309                         cfq_setup_merge(cfqq, new_cfqq);
3310                 goto expire;
3311         }
3312
3313         /*
3314          * No requests pending. If the active queue still has requests in
3315          * flight or is idling for a new request, allow either of these
3316          * conditions to happen (or time out) before selecting a new queue.
3317          */
3318         if (hrtimer_active(&cfqd->idle_slice_timer)) {
3319                 cfqq = NULL;
3320                 goto keep_queue;
3321         }
3322
3323         /*
3324          * This is a deep seek queue, but the device is much faster than
3325          * the queue can deliver, don't idle
3326          **/
3327         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3328             (cfq_cfqq_slice_new(cfqq) ||
3329             (cfqq->slice_end - now > now - cfqq->slice_start))) {
3330                 cfq_clear_cfqq_deep(cfqq);
3331                 cfq_clear_cfqq_idle_window(cfqq);
3332         }
3333
3334         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3335                 cfqq = NULL;
3336                 goto keep_queue;
3337         }
3338
3339         /*
3340          * If group idle is enabled and there are requests dispatched from
3341          * this group, wait for requests to complete.
3342          */
3343 check_group_idle:
3344         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3345             cfqq->cfqg->dispatched &&
3346             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3347                 cfqq = NULL;
3348                 goto keep_queue;
3349         }
3350
3351 expire:
3352         cfq_slice_expired(cfqd, 0);
3353 new_queue:
3354         /*
3355          * Current queue expired. Check if we have to switch to a new
3356          * service tree
3357          */
3358         if (!new_cfqq)
3359                 cfq_choose_cfqg(cfqd);
3360
3361         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3362 keep_queue:
3363         return cfqq;
3364 }
3365
3366 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3367 {
3368         int dispatched = 0;
3369
3370         while (cfqq->next_rq) {
3371                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3372                 dispatched++;
3373         }
3374
3375         BUG_ON(!list_empty(&cfqq->fifo));
3376
3377         /* By default cfqq is not expired if it is empty. Do it explicitly */
3378         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3379         return dispatched;
3380 }
3381
3382 /*
3383  * Drain our current requests. Used for barriers and when switching
3384  * io schedulers on-the-fly.
3385  */
3386 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3387 {
3388         struct cfq_queue *cfqq;
3389         int dispatched = 0;
3390
3391         /* Expire the timeslice of the current active queue first */
3392         cfq_slice_expired(cfqd, 0);
3393         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3394                 __cfq_set_active_queue(cfqd, cfqq);
3395                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3396         }
3397
3398         BUG_ON(cfqd->busy_queues);
3399
3400         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3401         return dispatched;
3402 }
3403
3404 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3405         struct cfq_queue *cfqq)
3406 {
3407         u64 now = ktime_get_ns();
3408
3409         /* the queue hasn't finished any request, can't estimate */
3410         if (cfq_cfqq_slice_new(cfqq))
3411                 return true;
3412         if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end)
3413                 return true;
3414
3415         return false;
3416 }
3417
3418 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3419 {
3420         unsigned int max_dispatch;
3421
3422         if (cfq_cfqq_must_dispatch(cfqq))
3423                 return true;
3424
3425         /*
3426          * Drain async requests before we start sync IO
3427          */
3428         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3429                 return false;
3430
3431         /*
3432          * If this is an async queue and we have sync IO in flight, let it wait
3433          */
3434         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3435                 return false;
3436
3437         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3438         if (cfq_class_idle(cfqq))
3439                 max_dispatch = 1;
3440
3441         /*
3442          * Does this cfqq already have too much IO in flight?
3443          */
3444         if (cfqq->dispatched >= max_dispatch) {
3445                 bool promote_sync = false;
3446                 /*
3447                  * idle queue must always only have a single IO in flight
3448                  */
3449                 if (cfq_class_idle(cfqq))
3450                         return false;
3451
3452                 /*
3453                  * If there is only one sync queue
3454                  * we can ignore async queue here and give the sync
3455                  * queue no dispatch limit. The reason is a sync queue can
3456                  * preempt async queue, limiting the sync queue doesn't make
3457                  * sense. This is useful for aiostress test.
3458                  */
3459                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3460                         promote_sync = true;
3461
3462                 /*
3463                  * We have other queues, don't allow more IO from this one
3464                  */
3465                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3466                                 !promote_sync)
3467                         return false;
3468
3469                 /*
3470                  * Sole queue user, no limit
3471                  */
3472                 if (cfqd->busy_queues == 1 || promote_sync)
3473                         max_dispatch = -1;
3474                 else
3475                         /*
3476                          * Normally we start throttling cfqq when cfq_quantum/2
3477                          * requests have been dispatched. But we can drive
3478                          * deeper queue depths at the beginning of slice
3479                          * subjected to upper limit of cfq_quantum.
3480                          * */
3481                         max_dispatch = cfqd->cfq_quantum;
3482         }
3483
3484         /*
3485          * Async queues must wait a bit before being allowed dispatch.
3486          * We also ramp up the dispatch depth gradually for async IO,
3487          * based on the last sync IO we serviced
3488          */
3489         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3490                 u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync;
3491                 unsigned int depth;
3492
3493                 depth = div64_u64(last_sync, cfqd->cfq_slice[1]);
3494                 if (!depth && !cfqq->dispatched)
3495                         depth = 1;
3496                 if (depth < max_dispatch)
3497                         max_dispatch = depth;
3498         }
3499
3500         /*
3501          * If we're below the current max, allow a dispatch
3502          */
3503         return cfqq->dispatched < max_dispatch;
3504 }
3505
3506 /*
3507  * Dispatch a request from cfqq, moving them to the request queue
3508  * dispatch list.
3509  */
3510 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3511 {
3512         struct request *rq;
3513
3514         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3515
3516         rq = cfq_check_fifo(cfqq);
3517         if (rq)
3518                 cfq_mark_cfqq_must_dispatch(cfqq);
3519
3520         if (!cfq_may_dispatch(cfqd, cfqq))
3521                 return false;
3522
3523         /*
3524          * follow expired path, else get first next available
3525          */
3526         if (!rq)
3527                 rq = cfqq->next_rq;
3528         else
3529                 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
3530
3531         /*
3532          * insert request into driver dispatch list
3533          */
3534         cfq_dispatch_insert(cfqd->queue, rq);
3535
3536         if (!cfqd->active_cic) {
3537                 struct cfq_io_cq *cic = RQ_CIC(rq);
3538
3539                 atomic_long_inc(&cic->icq.ioc->refcount);
3540                 cfqd->active_cic = cic;
3541         }
3542
3543         return true;
3544 }
3545
3546 /*
3547  * Find the cfqq that we need to service and move a request from that to the
3548  * dispatch list
3549  */
3550 static int cfq_dispatch_requests(struct request_queue *q, int force)
3551 {
3552         struct cfq_data *cfqd = q->elevator->elevator_data;
3553         struct cfq_queue *cfqq;
3554
3555         if (!cfqd->busy_queues)
3556                 return 0;
3557
3558         if (unlikely(force))
3559                 return cfq_forced_dispatch(cfqd);
3560
3561         cfqq = cfq_select_queue(cfqd);
3562         if (!cfqq)
3563                 return 0;
3564
3565         /*
3566          * Dispatch a request from this cfqq, if it is allowed
3567          */
3568         if (!cfq_dispatch_request(cfqd, cfqq))
3569                 return 0;
3570
3571         cfqq->slice_dispatch++;
3572         cfq_clear_cfqq_must_dispatch(cfqq);
3573
3574         /*
3575          * expire an async queue immediately if it has used up its slice. idle
3576          * queue always expire after 1 dispatch round.
3577          */
3578         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3579             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3580             cfq_class_idle(cfqq))) {
3581                 cfqq->slice_end = ktime_get_ns() + 1;
3582                 cfq_slice_expired(cfqd, 0);
3583         }
3584
3585         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3586         return 1;
3587 }
3588
3589 /*
3590  * task holds one reference to the queue, dropped when task exits. each rq
3591  * in-flight on this queue also holds a reference, dropped when rq is freed.
3592  *
3593  * Each cfq queue took a reference on the parent group. Drop it now.
3594  * queue lock must be held here.
3595  */
3596 static void cfq_put_queue(struct cfq_queue *cfqq)
3597 {
3598         struct cfq_data *cfqd = cfqq->cfqd;
3599         struct cfq_group *cfqg;
3600
3601         BUG_ON(cfqq->ref <= 0);
3602
3603         cfqq->ref--;
3604         if (cfqq->ref)
3605                 return;
3606
3607         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3608         BUG_ON(rb_first(&cfqq->sort_list));
3609         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3610         cfqg = cfqq->cfqg;
3611
3612         if (unlikely(cfqd->active_queue == cfqq)) {
3613                 __cfq_slice_expired(cfqd, cfqq, 0);
3614                 cfq_schedule_dispatch(cfqd);
3615         }
3616
3617         BUG_ON(cfq_cfqq_on_rr(cfqq));
3618         kmem_cache_free(cfq_pool, cfqq);
3619         cfqg_put(cfqg);
3620 }
3621
3622 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3623 {
3624         struct cfq_queue *__cfqq, *next;
3625
3626         /*
3627          * If this queue was scheduled to merge with another queue, be
3628          * sure to drop the reference taken on that queue (and others in
3629          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3630          */
3631         __cfqq = cfqq->new_cfqq;
3632         while (__cfqq) {
3633                 if (__cfqq == cfqq) {
3634                         WARN(1, "cfqq->new_cfqq loop detected\n");
3635                         break;
3636                 }
3637                 next = __cfqq->new_cfqq;
3638                 cfq_put_queue(__cfqq);
3639                 __cfqq = next;
3640         }
3641 }
3642
3643 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3644 {
3645         if (unlikely(cfqq == cfqd->active_queue)) {
3646                 __cfq_slice_expired(cfqd, cfqq, 0);
3647                 cfq_schedule_dispatch(cfqd);
3648         }
3649
3650         cfq_put_cooperator(cfqq);
3651
3652         cfq_put_queue(cfqq);
3653 }
3654
3655 static void cfq_init_icq(struct io_cq *icq)
3656 {
3657         struct cfq_io_cq *cic = icq_to_cic(icq);
3658
3659         cic->ttime.last_end_request = ktime_get_ns();
3660 }
3661
3662 static void cfq_exit_icq(struct io_cq *icq)
3663 {
3664         struct cfq_io_cq *cic = icq_to_cic(icq);
3665         struct cfq_data *cfqd = cic_to_cfqd(cic);
3666
3667         if (cic_to_cfqq(cic, false)) {
3668                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3669                 cic_set_cfqq(cic, NULL, false);
3670         }
3671
3672         if (cic_to_cfqq(cic, true)) {
3673                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3674                 cic_set_cfqq(cic, NULL, true);
3675         }
3676 }
3677
3678 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3679 {
3680         struct task_struct *tsk = current;
3681         int ioprio_class;
3682
3683         if (!cfq_cfqq_prio_changed(cfqq))
3684                 return;
3685
3686         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3687         switch (ioprio_class) {
3688         default:
3689                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3690         case IOPRIO_CLASS_NONE:
3691                 /*
3692                  * no prio set, inherit CPU scheduling settings
3693                  */
3694                 cfqq->ioprio = task_nice_ioprio(tsk);
3695                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3696                 break;
3697         case IOPRIO_CLASS_RT:
3698                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3699                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3700                 break;
3701         case IOPRIO_CLASS_BE:
3702                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3703                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3704                 break;
3705         case IOPRIO_CLASS_IDLE:
3706                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3707                 cfqq->ioprio = 7;
3708                 cfq_clear_cfqq_idle_window(cfqq);
3709                 break;
3710         }
3711
3712         /*
3713          * keep track of original prio settings in case we have to temporarily
3714          * elevate the priority of this queue
3715          */
3716         cfqq->org_ioprio = cfqq->ioprio;
3717         cfqq->org_ioprio_class = cfqq->ioprio_class;
3718         cfq_clear_cfqq_prio_changed(cfqq);
3719 }
3720
3721 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3722 {
3723         int ioprio = cic->icq.ioc->ioprio;
3724         struct cfq_data *cfqd = cic_to_cfqd(cic);
3725         struct cfq_queue *cfqq;
3726
3727         /*
3728          * Check whether ioprio has changed.  The condition may trigger
3729          * spuriously on a newly created cic but there's no harm.
3730          */
3731         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3732                 return;
3733
3734         cfqq = cic_to_cfqq(cic, false);
3735         if (cfqq) {
3736                 cfq_put_queue(cfqq);
3737                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3738                 cic_set_cfqq(cic, cfqq, false);
3739         }
3740
3741         cfqq = cic_to_cfqq(cic, true);
3742         if (cfqq)
3743                 cfq_mark_cfqq_prio_changed(cfqq);
3744
3745         cic->ioprio = ioprio;
3746 }
3747
3748 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3749                           pid_t pid, bool is_sync)
3750 {
3751         RB_CLEAR_NODE(&cfqq->rb_node);
3752         RB_CLEAR_NODE(&cfqq->p_node);
3753         INIT_LIST_HEAD(&cfqq->fifo);
3754
3755         cfqq->ref = 0;
3756         cfqq->cfqd = cfqd;
3757
3758         cfq_mark_cfqq_prio_changed(cfqq);
3759
3760         if (is_sync) {
3761                 if (!cfq_class_idle(cfqq))
3762                         cfq_mark_cfqq_idle_window(cfqq);
3763                 cfq_mark_cfqq_sync(cfqq);
3764         }
3765         cfqq->pid = pid;
3766 }
3767
3768 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3769 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3770 {
3771         struct cfq_data *cfqd = cic_to_cfqd(cic);
3772         struct cfq_queue *cfqq;
3773         uint64_t serial_nr;
3774
3775         rcu_read_lock();
3776         serial_nr = bio_blkcg(bio)->css.serial_nr;
3777         rcu_read_unlock();
3778
3779         /*
3780          * Check whether blkcg has changed.  The condition may trigger
3781          * spuriously on a newly created cic but there's no harm.
3782          */
3783         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3784                 return;
3785
3786         /*
3787          * Drop reference to queues.  New queues will be assigned in new
3788          * group upon arrival of fresh requests.
3789          */
3790         cfqq = cic_to_cfqq(cic, false);
3791         if (cfqq) {
3792                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3793                 cic_set_cfqq(cic, NULL, false);
3794                 cfq_put_queue(cfqq);
3795         }
3796
3797         cfqq = cic_to_cfqq(cic, true);
3798         if (cfqq) {
3799                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3800                 cic_set_cfqq(cic, NULL, true);
3801                 cfq_put_queue(cfqq);
3802         }
3803
3804         cic->blkcg_serial_nr = serial_nr;
3805 }
3806 #else
3807 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3808 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3809
3810 static struct cfq_queue **
3811 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3812 {
3813         switch (ioprio_class) {
3814         case IOPRIO_CLASS_RT:
3815                 return &cfqg->async_cfqq[0][ioprio];
3816         case IOPRIO_CLASS_NONE:
3817                 ioprio = IOPRIO_NORM;
3818                 /* fall through */
3819         case IOPRIO_CLASS_BE:
3820                 return &cfqg->async_cfqq[1][ioprio];
3821         case IOPRIO_CLASS_IDLE:
3822                 return &cfqg->async_idle_cfqq;
3823         default:
3824                 BUG();
3825         }
3826 }
3827
3828 static struct cfq_queue *
3829 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3830               struct bio *bio)
3831 {
3832         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3833         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3834         struct cfq_queue **async_cfqq = NULL;
3835         struct cfq_queue *cfqq;
3836         struct cfq_group *cfqg;
3837
3838         rcu_read_lock();
3839         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3840         if (!cfqg) {
3841                 cfqq = &cfqd->oom_cfqq;
3842                 goto out;
3843         }
3844
3845         if (!is_sync) {
3846                 if (!ioprio_valid(cic->ioprio)) {
3847                         struct task_struct *tsk = current;
3848                         ioprio = task_nice_ioprio(tsk);
3849                         ioprio_class = task_nice_ioclass(tsk);
3850                 }
3851                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3852                 cfqq = *async_cfqq;
3853                 if (cfqq)
3854                         goto out;
3855         }
3856
3857         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3858                                      cfqd->queue->node);
3859         if (!cfqq) {
3860                 cfqq = &cfqd->oom_cfqq;
3861                 goto out;
3862         }
3863
3864         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3865         cfq_init_prio_data(cfqq, cic);
3866         cfq_link_cfqq_cfqg(cfqq, cfqg);
3867         cfq_log_cfqq(cfqd, cfqq, "alloced");
3868
3869         if (async_cfqq) {
3870                 /* a new async queue is created, pin and remember */
3871                 cfqq->ref++;
3872                 *async_cfqq = cfqq;
3873         }
3874 out:
3875         cfqq->ref++;
3876         rcu_read_unlock();
3877         return cfqq;
3878 }
3879
3880 static void
3881 __cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle)
3882 {
3883         u64 elapsed = ktime_get_ns() - ttime->last_end_request;
3884         elapsed = min(elapsed, 2UL * slice_idle);
3885
3886         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3887         ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
3888         ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3889                                      ttime->ttime_samples);
3890 }
3891
3892 static void
3893 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3894                         struct cfq_io_cq *cic)
3895 {
3896         if (cfq_cfqq_sync(cfqq)) {
3897                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3898                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3899                         cfqd->cfq_slice_idle);
3900         }
3901 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3902         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3903 #endif
3904 }
3905
3906 static void
3907 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3908                        struct request *rq)
3909 {
3910         sector_t sdist = 0;
3911         sector_t n_sec = blk_rq_sectors(rq);
3912         if (cfqq->last_request_pos) {
3913                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3914                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3915                 else
3916                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3917         }
3918
3919         cfqq->seek_history <<= 1;
3920         if (blk_queue_nonrot(cfqd->queue))
3921                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3922         else
3923                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3924 }
3925
3926 /*
3927  * Disable idle window if the process thinks too long or seeks so much that
3928  * it doesn't matter
3929  */
3930 static void
3931 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3932                        struct cfq_io_cq *cic)
3933 {
3934         int old_idle, enable_idle;
3935
3936         /*
3937          * Don't idle for async or idle io prio class
3938          */
3939         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3940                 return;
3941
3942         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3943
3944         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3945                 cfq_mark_cfqq_deep(cfqq);
3946
3947         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3948                 enable_idle = 0;
3949         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3950                  !cfqd->cfq_slice_idle ||
3951                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3952                 enable_idle = 0;
3953         else if (sample_valid(cic->ttime.ttime_samples)) {
3954                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3955                         enable_idle = 0;
3956                 else
3957                         enable_idle = 1;
3958         }
3959
3960         if (old_idle != enable_idle) {
3961                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3962                 if (enable_idle)
3963                         cfq_mark_cfqq_idle_window(cfqq);
3964                 else
3965                         cfq_clear_cfqq_idle_window(cfqq);
3966         }
3967 }
3968
3969 /*
3970  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3971  * no or if we aren't sure, a 1 will cause a preempt.
3972  */
3973 static bool
3974 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3975                    struct request *rq)
3976 {
3977         struct cfq_queue *cfqq;
3978
3979         cfqq = cfqd->active_queue;
3980         if (!cfqq)
3981                 return false;
3982
3983         if (cfq_class_idle(new_cfqq))
3984                 return false;
3985
3986         if (cfq_class_idle(cfqq))
3987                 return true;
3988
3989         /*
3990          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3991          */
3992         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3993                 return false;
3994
3995         /*
3996          * if the new request is sync, but the currently running queue is
3997          * not, let the sync request have priority.
3998          */
3999         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
4000                 return true;
4001
4002         /*
4003          * Treat ancestors of current cgroup the same way as current cgroup.
4004          * For anybody else we disallow preemption to guarantee service
4005          * fairness among cgroups.
4006          */
4007         if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
4008                 return false;
4009
4010         if (cfq_slice_used(cfqq))
4011                 return true;
4012
4013         /*
4014          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
4015          */
4016         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
4017                 return true;
4018
4019         WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
4020         /* Allow preemption only if we are idling on sync-noidle tree */
4021         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
4022             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
4023             RB_EMPTY_ROOT(&cfqq->sort_list))
4024                 return true;
4025
4026         /*
4027          * So both queues are sync. Let the new request get disk time if
4028          * it's a metadata request and the current queue is doing regular IO.
4029          */
4030         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
4031                 return true;
4032
4033         /* An idle queue should not be idle now for some reason */
4034         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4035                 return true;
4036
4037         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
4038                 return false;
4039
4040         /*
4041          * if this request is as-good as one we would expect from the
4042          * current cfqq, let it preempt
4043          */
4044         if (cfq_rq_close(cfqd, cfqq, rq))
4045                 return true;
4046
4047         return false;
4048 }
4049
4050 /*
4051  * cfqq preempts the active queue. if we allowed preempt with no slice left,
4052  * let it have half of its nominal slice.
4053  */
4054 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4055 {
4056         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4057
4058         cfq_log_cfqq(cfqd, cfqq, "preempt");
4059         cfq_slice_expired(cfqd, 1);
4060
4061         /*
4062          * workload type is changed, don't save slice, otherwise preempt
4063          * doesn't happen
4064          */
4065         if (old_type != cfqq_type(cfqq))
4066                 cfqq->cfqg->saved_wl_slice = 0;
4067
4068         /*
4069          * Put the new queue at the front of the of the current list,
4070          * so we know that it will be selected next.
4071          */
4072         BUG_ON(!cfq_cfqq_on_rr(cfqq));
4073
4074         cfq_service_tree_add(cfqd, cfqq, 1);
4075
4076         cfqq->slice_end = 0;
4077         cfq_mark_cfqq_slice_new(cfqq);
4078 }
4079
4080 /*
4081  * Called when a new fs request (rq) is added (to cfqq). Check if there's
4082  * something we should do about it
4083  */
4084 static void
4085 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4086                 struct request *rq)
4087 {
4088         struct cfq_io_cq *cic = RQ_CIC(rq);
4089
4090         cfqd->rq_queued++;
4091         if (rq->cmd_flags & REQ_PRIO)
4092                 cfqq->prio_pending++;
4093
4094         cfq_update_io_thinktime(cfqd, cfqq, cic);
4095         cfq_update_io_seektime(cfqd, cfqq, rq);
4096         cfq_update_idle_window(cfqd, cfqq, cic);
4097
4098         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4099
4100         if (cfqq == cfqd->active_queue) {
4101                 /*
4102                  * Remember that we saw a request from this process, but
4103                  * don't start queuing just yet. Otherwise we risk seeing lots
4104                  * of tiny requests, because we disrupt the normal plugging
4105                  * and merging. If the request is already larger than a single
4106                  * page, let it rip immediately. For that case we assume that
4107                  * merging is already done. Ditto for a busy system that
4108                  * has other work pending, don't risk delaying until the
4109                  * idle timer unplug to continue working.
4110                  */
4111                 if (cfq_cfqq_wait_request(cfqq)) {
4112                         if (blk_rq_bytes(rq) > PAGE_SIZE ||
4113                             cfqd->busy_queues > 1) {
4114                                 cfq_del_timer(cfqd, cfqq);
4115                                 cfq_clear_cfqq_wait_request(cfqq);
4116                                 __blk_run_queue(cfqd->queue);
4117                         } else {
4118                                 cfqg_stats_update_idle_time(cfqq->cfqg);
4119                                 cfq_mark_cfqq_must_dispatch(cfqq);
4120                         }
4121                 }
4122         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
4123                 /*
4124                  * not the active queue - expire current slice if it is
4125                  * idle and has expired it's mean thinktime or this new queue
4126                  * has some old slice time left and is of higher priority or
4127                  * this new queue is RT and the current one is BE
4128                  */
4129                 cfq_preempt_queue(cfqd, cfqq);
4130                 __blk_run_queue(cfqd->queue);
4131         }
4132 }
4133
4134 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4135 {
4136         struct cfq_data *cfqd = q->elevator->elevator_data;
4137         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4138
4139         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4140         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4141
4142         rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4143         list_add_tail(&rq->queuelist, &cfqq->fifo);
4144         cfq_add_rq_rb(rq);
4145         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group, req_op(rq),
4146                                  rq->cmd_flags);
4147         cfq_rq_enqueued(cfqd, cfqq, rq);
4148 }
4149
4150 /*
4151  * Update hw_tag based on peak queue depth over 50 samples under
4152  * sufficient load.
4153  */
4154 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4155 {
4156         struct cfq_queue *cfqq = cfqd->active_queue;
4157
4158         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4159                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4160
4161         if (cfqd->hw_tag == 1)
4162                 return;
4163
4164         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4165             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4166                 return;
4167
4168         /*
4169          * If active queue hasn't enough requests and can idle, cfq might not
4170          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4171          * case
4172          */
4173         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4174             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4175             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4176                 return;
4177
4178         if (cfqd->hw_tag_samples++ < 50)
4179                 return;
4180
4181         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4182                 cfqd->hw_tag = 1;
4183         else
4184                 cfqd->hw_tag = 0;
4185 }
4186
4187 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4188 {
4189         struct cfq_io_cq *cic = cfqd->active_cic;
4190         u64 now = ktime_get_ns();
4191
4192         /* If the queue already has requests, don't wait */
4193         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4194                 return false;
4195
4196         /* If there are other queues in the group, don't wait */
4197         if (cfqq->cfqg->nr_cfqq > 1)
4198                 return false;
4199
4200         /* the only queue in the group, but think time is big */
4201         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4202                 return false;
4203
4204         if (cfq_slice_used(cfqq))
4205                 return true;
4206
4207         /* if slice left is less than think time, wait busy */
4208         if (cic && sample_valid(cic->ttime.ttime_samples)
4209             && (cfqq->slice_end - now < cic->ttime.ttime_mean))
4210                 return true;
4211
4212         /*
4213          * If think times is less than a jiffy than ttime_mean=0 and above
4214          * will not be true. It might happen that slice has not expired yet
4215          * but will expire soon (4-5 ns) during select_queue(). To cover the
4216          * case where think time is less than a jiffy, mark the queue wait
4217          * busy if only 1 jiffy is left in the slice.
4218          */
4219         if (cfqq->slice_end - now <= jiffies_to_nsecs(1))
4220                 return true;
4221
4222         return false;
4223 }
4224
4225 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4226 {
4227         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4228         struct cfq_data *cfqd = cfqq->cfqd;
4229         const int sync = rq_is_sync(rq);
4230         u64 now = ktime_get_ns();
4231
4232         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4233                      !!(rq->cmd_flags & REQ_NOIDLE));
4234
4235         cfq_update_hw_tag(cfqd);
4236
4237         WARN_ON(!cfqd->rq_in_driver);
4238         WARN_ON(!cfqq->dispatched);
4239         cfqd->rq_in_driver--;
4240         cfqq->dispatched--;
4241         (RQ_CFQG(rq))->dispatched--;
4242         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4243                                      rq_io_start_time_ns(rq), req_op(rq),
4244                                      rq->cmd_flags);
4245
4246         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4247
4248         if (sync) {
4249                 struct cfq_rb_root *st;
4250
4251                 RQ_CIC(rq)->ttime.last_end_request = now;
4252
4253                 if (cfq_cfqq_on_rr(cfqq))
4254                         st = cfqq->service_tree;
4255                 else
4256                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4257                                         cfqq_type(cfqq));
4258
4259                 st->ttime.last_end_request = now;
4260                 /*
4261                  * We have to do this check in jiffies since start_time is in
4262                  * jiffies and it is not trivial to convert to ns. If
4263                  * cfq_fifo_expire[1] ever comes close to 1 jiffie, this test
4264                  * will become problematic but so far we are fine (the default
4265                  * is 128 ms).
4266                  */
4267                 if (!time_after(rq->start_time +
4268                                   nsecs_to_jiffies(cfqd->cfq_fifo_expire[1]),
4269                                 jiffies))
4270                         cfqd->last_delayed_sync = now;
4271         }
4272
4273 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4274         cfqq->cfqg->ttime.last_end_request = now;
4275 #endif
4276
4277         /*
4278          * If this is the active queue, check if it needs to be expired,
4279          * or if we want to idle in case it has no pending requests.
4280          */
4281         if (cfqd->active_queue == cfqq) {
4282                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4283
4284                 if (cfq_cfqq_slice_new(cfqq)) {
4285                         cfq_set_prio_slice(cfqd, cfqq);
4286                         cfq_clear_cfqq_slice_new(cfqq);
4287                 }
4288
4289                 /*
4290                  * Should we wait for next request to come in before we expire
4291                  * the queue.
4292                  */
4293                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4294                         u64 extend_sl = cfqd->cfq_slice_idle;
4295                         if (!cfqd->cfq_slice_idle)
4296                                 extend_sl = cfqd->cfq_group_idle;
4297                         cfqq->slice_end = now + extend_sl;
4298                         cfq_mark_cfqq_wait_busy(cfqq);
4299                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4300                 }
4301
4302                 /*
4303                  * Idling is not enabled on:
4304                  * - expired queues
4305                  * - idle-priority queues
4306                  * - async queues
4307                  * - queues with still some requests queued
4308                  * - when there is a close cooperator
4309                  */
4310                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4311                         cfq_slice_expired(cfqd, 1);
4312                 else if (sync && cfqq_empty &&
4313                          !cfq_close_cooperator(cfqd, cfqq)) {
4314                         cfq_arm_slice_timer(cfqd);
4315                 }
4316         }
4317
4318         if (!cfqd->rq_in_driver)
4319                 cfq_schedule_dispatch(cfqd);
4320 }
4321
4322 static void cfqq_boost_on_prio(struct cfq_queue *cfqq, int op_flags)
4323 {
4324         /*
4325          * If REQ_PRIO is set, boost class and prio level, if it's below
4326          * BE/NORM. If prio is not set, restore the potentially boosted
4327          * class/prio level.
4328          */
4329         if (!(op_flags & REQ_PRIO)) {
4330                 cfqq->ioprio_class = cfqq->org_ioprio_class;
4331                 cfqq->ioprio = cfqq->org_ioprio;
4332         } else {
4333                 if (cfq_class_idle(cfqq))
4334                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
4335                 if (cfqq->ioprio > IOPRIO_NORM)
4336                         cfqq->ioprio = IOPRIO_NORM;
4337         }
4338 }
4339
4340 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4341 {
4342         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4343                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4344                 return ELV_MQUEUE_MUST;
4345         }
4346
4347         return ELV_MQUEUE_MAY;
4348 }
4349
4350 static int cfq_may_queue(struct request_queue *q, int op, int op_flags)
4351 {
4352         struct cfq_data *cfqd = q->elevator->elevator_data;
4353         struct task_struct *tsk = current;
4354         struct cfq_io_cq *cic;
4355         struct cfq_queue *cfqq;
4356
4357         /*
4358          * don't force setup of a queue from here, as a call to may_queue
4359          * does not necessarily imply that a request actually will be queued.
4360          * so just lookup a possibly existing queue, or return 'may queue'
4361          * if that fails
4362          */
4363         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4364         if (!cic)
4365                 return ELV_MQUEUE_MAY;
4366
4367         cfqq = cic_to_cfqq(cic, rw_is_sync(op, op_flags));
4368         if (cfqq) {
4369                 cfq_init_prio_data(cfqq, cic);
4370                 cfqq_boost_on_prio(cfqq, op_flags);
4371
4372                 return __cfq_may_queue(cfqq);
4373         }
4374
4375         return ELV_MQUEUE_MAY;
4376 }
4377
4378 /*
4379  * queue lock held here
4380  */
4381 static void cfq_put_request(struct request *rq)
4382 {
4383         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4384
4385         if (cfqq) {
4386                 const int rw = rq_data_dir(rq);
4387
4388                 BUG_ON(!cfqq->allocated[rw]);
4389                 cfqq->allocated[rw]--;
4390
4391                 /* Put down rq reference on cfqg */
4392                 cfqg_put(RQ_CFQG(rq));
4393                 rq->elv.priv[0] = NULL;
4394                 rq->elv.priv[1] = NULL;
4395
4396                 cfq_put_queue(cfqq);
4397         }
4398 }
4399
4400 static struct cfq_queue *
4401 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4402                 struct cfq_queue *cfqq)
4403 {
4404         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4405         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4406         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4407         cfq_put_queue(cfqq);
4408         return cic_to_cfqq(cic, 1);
4409 }
4410
4411 /*
4412  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4413  * was the last process referring to said cfqq.
4414  */
4415 static struct cfq_queue *
4416 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4417 {
4418         if (cfqq_process_refs(cfqq) == 1) {
4419                 cfqq->pid = current->pid;
4420                 cfq_clear_cfqq_coop(cfqq);
4421                 cfq_clear_cfqq_split_coop(cfqq);
4422                 return cfqq;
4423         }
4424
4425         cic_set_cfqq(cic, NULL, 1);
4426
4427         cfq_put_cooperator(cfqq);
4428
4429         cfq_put_queue(cfqq);
4430         return NULL;
4431 }
4432 /*
4433  * Allocate cfq data structures associated with this request.
4434  */
4435 static int
4436 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4437                 gfp_t gfp_mask)
4438 {
4439         struct cfq_data *cfqd = q->elevator->elevator_data;
4440         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4441         const int rw = rq_data_dir(rq);
4442         const bool is_sync = rq_is_sync(rq);
4443         struct cfq_queue *cfqq;
4444
4445         spin_lock_irq(q->queue_lock);
4446
4447         check_ioprio_changed(cic, bio);
4448         check_blkcg_changed(cic, bio);
4449 new_queue:
4450         cfqq = cic_to_cfqq(cic, is_sync);
4451         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4452                 if (cfqq)
4453                         cfq_put_queue(cfqq);
4454                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4455                 cic_set_cfqq(cic, cfqq, is_sync);
4456         } else {
4457                 /*
4458                  * If the queue was seeky for too long, break it apart.
4459                  */
4460                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4461                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4462                         cfqq = split_cfqq(cic, cfqq);
4463                         if (!cfqq)
4464                                 goto new_queue;
4465                 }
4466
4467                 /*
4468                  * Check to see if this queue is scheduled to merge with
4469                  * another, closely cooperating queue.  The merging of
4470                  * queues happens here as it must be done in process context.
4471                  * The reference on new_cfqq was taken in merge_cfqqs.
4472                  */
4473                 if (cfqq->new_cfqq)
4474                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4475         }
4476
4477         cfqq->allocated[rw]++;
4478
4479         cfqq->ref++;
4480         cfqg_get(cfqq->cfqg);
4481         rq->elv.priv[0] = cfqq;
4482         rq->elv.priv[1] = cfqq->cfqg;
4483         spin_unlock_irq(q->queue_lock);
4484         return 0;
4485 }
4486
4487 static void cfq_kick_queue(struct work_struct *work)
4488 {
4489         struct cfq_data *cfqd =
4490                 container_of(work, struct cfq_data, unplug_work);
4491         struct request_queue *q = cfqd->queue;
4492
4493         spin_lock_irq(q->queue_lock);
4494         __blk_run_queue(cfqd->queue);
4495         spin_unlock_irq(q->queue_lock);
4496 }
4497
4498 /*
4499  * Timer running if the active_queue is currently idling inside its time slice
4500  */
4501 static enum hrtimer_restart cfq_idle_slice_timer(struct hrtimer *timer)
4502 {
4503         struct cfq_data *cfqd = container_of(timer, struct cfq_data,
4504                                              idle_slice_timer);
4505         struct cfq_queue *cfqq;
4506         unsigned long flags;
4507         int timed_out = 1;
4508
4509         cfq_log(cfqd, "idle timer fired");
4510
4511         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4512
4513         cfqq = cfqd->active_queue;
4514         if (cfqq) {
4515                 timed_out = 0;
4516
4517                 /*
4518                  * We saw a request before the queue expired, let it through
4519                  */
4520                 if (cfq_cfqq_must_dispatch(cfqq))
4521                         goto out_kick;
4522
4523                 /*
4524                  * expired
4525                  */
4526                 if (cfq_slice_used(cfqq))
4527                         goto expire;
4528
4529                 /*
4530                  * only expire and reinvoke request handler, if there are
4531                  * other queues with pending requests
4532                  */
4533                 if (!cfqd->busy_queues)
4534                         goto out_cont;
4535
4536                 /*
4537                  * not expired and it has a request pending, let it dispatch
4538                  */
4539                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4540                         goto out_kick;
4541
4542                 /*
4543                  * Queue depth flag is reset only when the idle didn't succeed
4544                  */
4545                 cfq_clear_cfqq_deep(cfqq);
4546         }
4547 expire:
4548         cfq_slice_expired(cfqd, timed_out);
4549 out_kick:
4550         cfq_schedule_dispatch(cfqd);
4551 out_cont:
4552         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4553         return HRTIMER_NORESTART;
4554 }
4555
4556 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4557 {
4558         hrtimer_cancel(&cfqd->idle_slice_timer);
4559         cancel_work_sync(&cfqd->unplug_work);
4560 }
4561
4562 static void cfq_exit_queue(struct elevator_queue *e)
4563 {
4564         struct cfq_data *cfqd = e->elevator_data;
4565         struct request_queue *q = cfqd->queue;
4566
4567         cfq_shutdown_timer_wq(cfqd);
4568
4569         spin_lock_irq(q->queue_lock);
4570
4571         if (cfqd->active_queue)
4572                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4573
4574         spin_unlock_irq(q->queue_lock);
4575
4576         cfq_shutdown_timer_wq(cfqd);
4577
4578 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4579         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4580 #else
4581         kfree(cfqd->root_group);
4582 #endif
4583         kfree(cfqd);
4584 }
4585
4586 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4587 {
4588         struct cfq_data *cfqd;
4589         struct blkcg_gq *blkg __maybe_unused;
4590         int i, ret;
4591         struct elevator_queue *eq;
4592
4593         eq = elevator_alloc(q, e);
4594         if (!eq)
4595                 return -ENOMEM;
4596
4597         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4598         if (!cfqd) {
4599                 kobject_put(&eq->kobj);
4600                 return -ENOMEM;
4601         }
4602         eq->elevator_data = cfqd;
4603
4604         cfqd->queue = q;
4605         spin_lock_irq(q->queue_lock);
4606         q->elevator = eq;
4607         spin_unlock_irq(q->queue_lock);
4608
4609         /* Init root service tree */
4610         cfqd->grp_service_tree = CFQ_RB_ROOT;
4611
4612         /* Init root group and prefer root group over other groups by default */
4613 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4614         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4615         if (ret)
4616                 goto out_free;
4617
4618         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4619 #else
4620         ret = -ENOMEM;
4621         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4622                                         GFP_KERNEL, cfqd->queue->node);
4623         if (!cfqd->root_group)
4624                 goto out_free;
4625
4626         cfq_init_cfqg_base(cfqd->root_group);
4627         cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4628         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4629 #endif
4630
4631         /*
4632          * Not strictly needed (since RB_ROOT just clears the node and we
4633          * zeroed cfqd on alloc), but better be safe in case someone decides
4634          * to add magic to the rb code
4635          */
4636         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4637                 cfqd->prio_trees[i] = RB_ROOT;
4638
4639         /*
4640          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4641          * Grab a permanent reference to it, so that the normal code flow
4642          * will not attempt to free it.  oom_cfqq is linked to root_group
4643          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4644          * the reference from linking right away.
4645          */
4646         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4647         cfqd->oom_cfqq.ref++;
4648
4649         spin_lock_irq(q->queue_lock);
4650         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4651         cfqg_put(cfqd->root_group);
4652         spin_unlock_irq(q->queue_lock);
4653
4654         hrtimer_init(&cfqd->idle_slice_timer, CLOCK_MONOTONIC,
4655                      HRTIMER_MODE_REL);
4656         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4657
4658         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4659
4660         cfqd->cfq_quantum = cfq_quantum;
4661         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4662         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4663         cfqd->cfq_back_max = cfq_back_max;
4664         cfqd->cfq_back_penalty = cfq_back_penalty;
4665         cfqd->cfq_slice[0] = cfq_slice_async;
4666         cfqd->cfq_slice[1] = cfq_slice_sync;
4667         cfqd->cfq_target_latency = cfq_target_latency;
4668         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4669         cfqd->cfq_slice_idle = cfq_slice_idle;
4670         cfqd->cfq_group_idle = cfq_group_idle;
4671         cfqd->cfq_latency = 1;
4672         cfqd->hw_tag = -1;
4673         /*
4674          * we optimistically start assuming sync ops weren't delayed in last
4675          * second, in order to have larger depth for async operations.
4676          */
4677         cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC;
4678         return 0;
4679
4680 out_free:
4681         kfree(cfqd);
4682         kobject_put(&eq->kobj);
4683         return ret;
4684 }
4685
4686 static void cfq_registered_queue(struct request_queue *q)
4687 {
4688         struct elevator_queue *e = q->elevator;
4689         struct cfq_data *cfqd = e->elevator_data;
4690
4691         /*
4692          * Default to IOPS mode with no idling for SSDs
4693          */
4694         if (blk_queue_nonrot(q))
4695                 cfqd->cfq_slice_idle = 0;
4696 }
4697
4698 /*
4699  * sysfs parts below -->
4700  */
4701 static ssize_t
4702 cfq_var_show(unsigned int var, char *page)
4703 {
4704         return sprintf(page, "%u\n", var);
4705 }
4706
4707 static ssize_t
4708 cfq_var_store(unsigned int *var, const char *page, size_t count)
4709 {
4710         char *p = (char *) page;
4711
4712         *var = simple_strtoul(p, &p, 10);
4713         return count;
4714 }
4715
4716 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4717 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4718 {                                                                       \
4719         struct cfq_data *cfqd = e->elevator_data;                       \
4720         u64 __data = __VAR;                                             \
4721         if (__CONV)                                                     \
4722                 __data = div_u64(__data, NSEC_PER_MSEC);                        \
4723         return cfq_var_show(__data, (page));                            \
4724 }
4725 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4726 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4727 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4728 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4729 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4730 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4731 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4732 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4733 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4734 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4735 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4736 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4737 #undef SHOW_FUNCTION
4738
4739 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)                               \
4740 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4741 {                                                                       \
4742         struct cfq_data *cfqd = e->elevator_data;                       \
4743         u64 __data = __VAR;                                             \
4744         __data = div_u64(__data, NSEC_PER_USEC);                        \
4745         return cfq_var_show(__data, (page));                            \
4746 }
4747 USEC_SHOW_FUNCTION(cfq_slice_idle_us_show, cfqd->cfq_slice_idle);
4748 USEC_SHOW_FUNCTION(cfq_group_idle_us_show, cfqd->cfq_group_idle);
4749 USEC_SHOW_FUNCTION(cfq_slice_sync_us_show, cfqd->cfq_slice[1]);
4750 USEC_SHOW_FUNCTION(cfq_slice_async_us_show, cfqd->cfq_slice[0]);
4751 USEC_SHOW_FUNCTION(cfq_target_latency_us_show, cfqd->cfq_target_latency);
4752 #undef USEC_SHOW_FUNCTION
4753
4754 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4755 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4756 {                                                                       \
4757         struct cfq_data *cfqd = e->elevator_data;                       \
4758         unsigned int __data;                                            \
4759         int ret = cfq_var_store(&__data, (page), count);                \
4760         if (__data < (MIN))                                             \
4761                 __data = (MIN);                                         \
4762         else if (__data > (MAX))                                        \
4763                 __data = (MAX);                                         \
4764         if (__CONV)                                                     \
4765                 *(__PTR) = (u64)__data * NSEC_PER_MSEC;                 \
4766         else                                                            \
4767                 *(__PTR) = __data;                                      \
4768         return ret;                                                     \
4769 }
4770 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4771 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4772                 UINT_MAX, 1);
4773 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4774                 UINT_MAX, 1);
4775 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4776 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4777                 UINT_MAX, 0);
4778 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4779 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4780 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4781 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4782 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4783                 UINT_MAX, 0);
4784 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4785 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4786 #undef STORE_FUNCTION
4787
4788 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)                    \
4789 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4790 {                                                                       \
4791         struct cfq_data *cfqd = e->elevator_data;                       \
4792         unsigned int __data;                                            \
4793         int ret = cfq_var_store(&__data, (page), count);                \
4794         if (__data < (MIN))                                             \
4795                 __data = (MIN);                                         \
4796         else if (__data > (MAX))                                        \
4797                 __data = (MAX);                                         \
4798         *(__PTR) = (u64)__data * NSEC_PER_USEC;                         \
4799         return ret;                                                     \
4800 }
4801 USEC_STORE_FUNCTION(cfq_slice_idle_us_store, &cfqd->cfq_slice_idle, 0, UINT_MAX);
4802 USEC_STORE_FUNCTION(cfq_group_idle_us_store, &cfqd->cfq_group_idle, 0, UINT_MAX);
4803 USEC_STORE_FUNCTION(cfq_slice_sync_us_store, &cfqd->cfq_slice[1], 1, UINT_MAX);
4804 USEC_STORE_FUNCTION(cfq_slice_async_us_store, &cfqd->cfq_slice[0], 1, UINT_MAX);
4805 USEC_STORE_FUNCTION(cfq_target_latency_us_store, &cfqd->cfq_target_latency, 1, UINT_MAX);
4806 #undef USEC_STORE_FUNCTION
4807
4808 #define CFQ_ATTR(name) \
4809         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4810
4811 static struct elv_fs_entry cfq_attrs[] = {
4812         CFQ_ATTR(quantum),
4813         CFQ_ATTR(fifo_expire_sync),
4814         CFQ_ATTR(fifo_expire_async),
4815         CFQ_ATTR(back_seek_max),
4816         CFQ_ATTR(back_seek_penalty),
4817         CFQ_ATTR(slice_sync),
4818         CFQ_ATTR(slice_sync_us),
4819         CFQ_ATTR(slice_async),
4820         CFQ_ATTR(slice_async_us),
4821         CFQ_ATTR(slice_async_rq),
4822         CFQ_ATTR(slice_idle),
4823         CFQ_ATTR(slice_idle_us),
4824         CFQ_ATTR(group_idle),
4825         CFQ_ATTR(group_idle_us),
4826         CFQ_ATTR(low_latency),
4827         CFQ_ATTR(target_latency),
4828         CFQ_ATTR(target_latency_us),
4829         __ATTR_NULL
4830 };
4831
4832 static struct elevator_type iosched_cfq = {
4833         .ops = {
4834                 .elevator_merge_fn =            cfq_merge,
4835                 .elevator_merged_fn =           cfq_merged_request,
4836                 .elevator_merge_req_fn =        cfq_merged_requests,
4837                 .elevator_allow_bio_merge_fn =  cfq_allow_bio_merge,
4838                 .elevator_allow_rq_merge_fn =   cfq_allow_rq_merge,
4839                 .elevator_bio_merged_fn =       cfq_bio_merged,
4840                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4841                 .elevator_add_req_fn =          cfq_insert_request,
4842                 .elevator_activate_req_fn =     cfq_activate_request,
4843                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4844                 .elevator_completed_req_fn =    cfq_completed_request,
4845                 .elevator_former_req_fn =       elv_rb_former_request,
4846                 .elevator_latter_req_fn =       elv_rb_latter_request,
4847                 .elevator_init_icq_fn =         cfq_init_icq,
4848                 .elevator_exit_icq_fn =         cfq_exit_icq,
4849                 .elevator_set_req_fn =          cfq_set_request,
4850                 .elevator_put_req_fn =          cfq_put_request,
4851                 .elevator_may_queue_fn =        cfq_may_queue,
4852                 .elevator_init_fn =             cfq_init_queue,
4853                 .elevator_exit_fn =             cfq_exit_queue,
4854                 .elevator_registered_fn =       cfq_registered_queue,
4855         },
4856         .icq_size       =       sizeof(struct cfq_io_cq),
4857         .icq_align      =       __alignof__(struct cfq_io_cq),
4858         .elevator_attrs =       cfq_attrs,
4859         .elevator_name  =       "cfq",
4860         .elevator_owner =       THIS_MODULE,
4861 };
4862
4863 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4864 static struct blkcg_policy blkcg_policy_cfq = {
4865         .dfl_cftypes            = cfq_blkcg_files,
4866         .legacy_cftypes         = cfq_blkcg_legacy_files,
4867
4868         .cpd_alloc_fn           = cfq_cpd_alloc,
4869         .cpd_init_fn            = cfq_cpd_init,
4870         .cpd_free_fn            = cfq_cpd_free,
4871         .cpd_bind_fn            = cfq_cpd_bind,
4872
4873         .pd_alloc_fn            = cfq_pd_alloc,
4874         .pd_init_fn             = cfq_pd_init,
4875         .pd_offline_fn          = cfq_pd_offline,
4876         .pd_free_fn             = cfq_pd_free,
4877         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4878 };
4879 #endif
4880
4881 static int __init cfq_init(void)
4882 {
4883         int ret;
4884
4885 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4886         ret = blkcg_policy_register(&blkcg_policy_cfq);
4887         if (ret)
4888                 return ret;
4889 #else
4890         cfq_group_idle = 0;
4891 #endif
4892
4893         ret = -ENOMEM;
4894         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4895         if (!cfq_pool)
4896                 goto err_pol_unreg;
4897
4898         ret = elv_register(&iosched_cfq);
4899         if (ret)
4900                 goto err_free_pool;
4901
4902         return 0;
4903
4904 err_free_pool:
4905         kmem_cache_destroy(cfq_pool);
4906 err_pol_unreg:
4907 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4908         blkcg_policy_unregister(&blkcg_policy_cfq);
4909 #endif
4910         return ret;
4911 }
4912
4913 static void __exit cfq_exit(void)
4914 {
4915 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4916         blkcg_policy_unregister(&blkcg_policy_cfq);
4917 #endif
4918         elv_unregister(&iosched_cfq);
4919         kmem_cache_destroy(cfq_pool);
4920 }
4921
4922 module_init(cfq_init);
4923 module_exit(cfq_exit);
4924
4925 MODULE_AUTHOR("Jens Axboe");
4926 MODULE_LICENSE("GPL");
4927 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");