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 queue service tree for idle class
  40  */
  41 #define CFQ_IDLE_DELAY          (NSEC_PER_SEC / 5)
  42 /* offset from end of group service tree under time slice mode */
  43 #define CFQ_SLICE_MODE_GROUP_DELAY (NSEC_PER_SEC / 5)
  44 /* offset from end of group service under IOPS mode */
  45 #define CFQ_IOPS_MODE_GROUP_DELAY (HZ / 5)
  46
  47 /*
  48  * below this threshold, we consider thinktime immediate
  49  */
  50 #define CFQ_MIN_TT              (2 * NSEC_PER_SEC / HZ)
  51
  52 #define CFQ_SLICE_SCALE         (5)
  53 #define CFQ_HW_QUEUE_MIN        (5)
  54 #define CFQ_SERVICE_SHIFT       12
  55
  56 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  57 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  58 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  59 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  60
  61 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  62 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  63 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  64
  65 static struct kmem_cache *cfq_pool;
  66
  67 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  68 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  69 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  70
  71 #define sample_valid(samples)   ((samples) > 80)
  72 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  73
  74 /* blkio-related constants */
  75 #define CFQ_WEIGHT_LEGACY_MIN   10
  76 #define CFQ_WEIGHT_LEGACY_DFL   500
  77 #define CFQ_WEIGHT_LEGACY_MAX   1000
  78
  79 struct cfq_ttime {
  80         u64 last_end_request;
  81
  82         u64 ttime_total;
  83         u64 ttime_mean;
  84         unsigned long ttime_samples;
  85 };
  86
  87 /*
  88  * Most of our rbtree usage is for sorting with min extraction, so
  89  * if we cache the leftmost node we don't have to walk down the tree
  90  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  91  * move this into the elevator for the rq sorting as well.
  92  */
  93 struct cfq_rb_root {
  94         struct rb_root rb;
  95         struct rb_node *left;
  96         unsigned count;
  97         u64 min_vdisktime;
  98         struct cfq_ttime ttime;
  99 };
 100 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
 101                         .ttime = {.last_end_request = ktime_get_ns(),},}
 102
 103 /*
 104  * Per process-grouping structure
 105  */
 106 struct cfq_queue {
 107         /* reference count */
 108         int ref;
 109         /* various state flags, see below */
 110         unsigned int flags;
 111         /* parent cfq_data */
 112         struct cfq_data *cfqd;
 113         /* service_tree member */
 114         struct rb_node rb_node;
 115         /* service_tree key */
 116         u64 rb_key;
 117         /* prio tree member */
 118         struct rb_node p_node;
 119         /* prio tree root we belong to, if any */
 120         struct rb_root *p_root;
 121         /* sorted list of pending requests */
 122         struct rb_root sort_list;
 123         /* if fifo isn't expired, next request to serve */
 124         struct request *next_rq;
 125         /* requests queued in sort_list */
 126         int queued[2];
 127         /* currently allocated requests */
 128         int allocated[2];
 129         /* fifo list of requests in sort_list */
 130         struct list_head fifo;
 131
 132         /* time when queue got scheduled in to dispatch first request. */
 133         u64 dispatch_start;
 134         u64 allocated_slice;
 135         u64 slice_dispatch;
 136         /* time when first request from queue completed and slice started. */
 137         u64 slice_start;
 138         u64 slice_end;
 139         s64 slice_resid;
 140
 141         /* pending priority requests */
 142         int prio_pending;
 143         /* number of requests that are on the dispatch list or inside driver */
 144         int dispatched;
 145
 146         /* io prio of this group */
 147         unsigned short ioprio, org_ioprio;
 148         unsigned short ioprio_class, org_ioprio_class;
 149
 150         pid_t pid;
 151
 152         u32 seek_history;
 153         sector_t last_request_pos;
 154
 155         struct cfq_rb_root *service_tree;
 156         struct cfq_queue *new_cfqq;
 157         struct cfq_group *cfqg;
 158         /* Number of sectors dispatched from queue in single dispatch round */
 159         unsigned long nr_sectors;
 160 };
 161
 162 /*
 163  * First index in the service_trees.
 164  * IDLE is handled separately, so it has negative index
 165  */
 166 enum wl_class_t {
 167         BE_WORKLOAD = 0,
 168         RT_WORKLOAD = 1,
 169         IDLE_WORKLOAD = 2,
 170         CFQ_PRIO_NR,
 171 };
 172
 173 /*
 174  * Second index in the service_trees.
 175  */
 176 enum wl_type_t {
 177         ASYNC_WORKLOAD = 0,
 178         SYNC_NOIDLE_WORKLOAD = 1,
 179         SYNC_WORKLOAD = 2
 180 };
 181
 182 struct cfqg_stats {
 183 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 184         /* number of ios merged */
 185         struct blkg_rwstat              merged;
 186         /* total time spent on device in ns, may not be accurate w/ queueing */
 187         struct blkg_rwstat              service_time;
 188         /* total time spent waiting in scheduler queue in ns */
 189         struct blkg_rwstat              wait_time;
 190         /* number of IOs queued up */
 191         struct blkg_rwstat              queued;
 192         /* total disk time and nr sectors dispatched by this group */
 193         struct blkg_stat                time;
 194 #ifdef CONFIG_DEBUG_BLK_CGROUP
 195         /* time not charged to this cgroup */
 196         struct blkg_stat                unaccounted_time;
 197         /* sum of number of ios queued across all samples */
 198         struct blkg_stat                avg_queue_size_sum;
 199         /* count of samples taken for average */
 200         struct blkg_stat                avg_queue_size_samples;
 201         /* how many times this group has been removed from service tree */
 202         struct blkg_stat                dequeue;
 203         /* total time spent waiting for it to be assigned a timeslice. */
 204         struct blkg_stat                group_wait_time;
 205         /* time spent idling for this blkcg_gq */
 206         struct blkg_stat                idle_time;
 207         /* total time with empty current active q with other requests queued */
 208         struct blkg_stat                empty_time;
 209         /* fields after this shouldn't be cleared on stat reset */
 210         uint64_t                        start_group_wait_time;
 211         uint64_t                        start_idle_time;
 212         uint64_t                        start_empty_time;
 213         uint16_t                        flags;
 214 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 215 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 216 };
 217
 218 /* Per-cgroup data */
 219 struct cfq_group_data {
 220         /* must be the first member */
 221         struct blkcg_policy_data cpd;
 222
 223         unsigned int weight;
 224         unsigned int leaf_weight;
 225 };
 226
 227 /* This is per cgroup per device grouping structure */
 228 struct cfq_group {
 229         /* must be the first member */
 230         struct blkg_policy_data pd;
 231
 232         /* group service_tree member */
 233         struct rb_node rb_node;
 234
 235         /* group service_tree key */
 236         u64 vdisktime;
 237
 238         /*
 239          * The number of active cfqgs and sum of their weights under this
 240          * cfqg.  This covers this cfqg's leaf_weight and all children's
 241          * weights, but does not cover weights of further descendants.
 242          *
 243          * If a cfqg is on the service tree, it's active.  An active cfqg
 244          * also activates its parent and contributes to the children_weight
 245          * of the parent.
 246          */
 247         int nr_active;
 248         unsigned int children_weight;
 249
 250         /*
 251          * vfraction is the fraction of vdisktime that the tasks in this
 252          * cfqg are entitled to.  This is determined by compounding the
 253          * ratios walking up from this cfqg to the root.
 254          *
 255          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 256          * vfractions on a service tree is approximately 1.  The sum may
 257          * deviate a bit due to rounding errors and fluctuations caused by
 258          * cfqgs entering and leaving the service tree.
 259          */
 260         unsigned int vfraction;
 261
 262         /*
 263          * There are two weights - (internal) weight is the weight of this
 264          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 265          * this cfqg against the child cfqgs.  For the root cfqg, both
 266          * weights are kept in sync for backward compatibility.
 267          */
 268         unsigned int weight;
 269         unsigned int new_weight;
 270         unsigned int dev_weight;
 271
 272         unsigned int leaf_weight;
 273         unsigned int new_leaf_weight;
 274         unsigned int dev_leaf_weight;
 275
 276         /* number of cfqq currently on this group */
 277         int nr_cfqq;
 278
 279         /*
 280          * Per group busy queues average. Useful for workload slice calc. We
 281          * create the array for each prio class but at run time it is used
 282          * only for RT and BE class and slot for IDLE class remains unused.
 283          * This is primarily done to avoid confusion and a gcc warning.
 284          */
 285         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 286         /*
 287          * rr lists of queues with requests. We maintain service trees for
 288          * RT and BE classes. These trees are subdivided in subclasses
 289          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 290          * class there is no subclassification and all the cfq queues go on
 291          * a single tree service_tree_idle.
 292          * Counts are embedded in the cfq_rb_root
 293          */
 294         struct cfq_rb_root service_trees[2][3];
 295         struct cfq_rb_root service_tree_idle;
 296
 297         u64 saved_wl_slice;
 298         enum wl_type_t saved_wl_type;
 299         enum wl_class_t saved_wl_class;
 300
 301         /* number of requests that are on the dispatch list or inside driver */
 302         int dispatched;
 303         struct cfq_ttime ttime;
 304         struct cfqg_stats stats;        /* stats for this cfqg */
 305
 306         /* async queue for each priority case */
 307         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 308         struct cfq_queue *async_idle_cfqq;
 309
 310 };
 311
 312 struct cfq_io_cq {
 313         struct io_cq            icq;            /* must be the first member */
 314         struct cfq_queue        *cfqq[2];
 315         struct cfq_ttime        ttime;
 316         int                     ioprio;         /* the current ioprio */
 317 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 318         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 319 #endif
 320 };
 321
 322 /*
 323  * Per block device queue structure
 324  */
 325 struct cfq_data {
 326         struct request_queue *queue;
 327         /* Root service tree for cfq_groups */
 328         struct cfq_rb_root grp_service_tree;
 329         struct cfq_group *root_group;
 330
 331         /*
 332          * The priority currently being served
 333          */
 334         enum wl_class_t serving_wl_class;
 335         enum wl_type_t serving_wl_type;
 336         u64 workload_expires;
 337         struct cfq_group *serving_group;
 338
 339         /*
 340          * Each priority tree is sorted by next_request position.  These
 341          * trees are used when determining if two or more queues are
 342          * interleaving requests (see cfq_close_cooperator).
 343          */
 344         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 345
 346         unsigned int busy_queues;
 347         unsigned int busy_sync_queues;
 348
 349         int rq_in_driver;
 350         int rq_in_flight[2];
 351
 352         /*
 353          * queue-depth detection
 354          */
 355         int rq_queued;
 356         int hw_tag;
 357         /*
 358          * hw_tag can be
 359          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 360          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 361          *  0 => no NCQ
 362          */
 363         int hw_tag_est_depth;
 364         unsigned int hw_tag_samples;
 365
 366         /*
 367          * idle window management
 368          */
 369         struct hrtimer idle_slice_timer;
 370         struct work_struct unplug_work;
 371
 372         struct cfq_queue *active_queue;
 373         struct cfq_io_cq *active_cic;
 374
 375         sector_t last_position;
 376
 377         /*
 378          * tunables, see top of file
 379          */
 380         unsigned int cfq_quantum;
 381         unsigned int cfq_back_penalty;
 382         unsigned int cfq_back_max;
 383         unsigned int cfq_slice_async_rq;
 384         unsigned int cfq_latency;
 385         u64 cfq_fifo_expire[2];
 386         u64 cfq_slice[2];
 387         u64 cfq_slice_idle;
 388         u64 cfq_group_idle;
 389         u64 cfq_target_latency;
 390
 391         /*
 392          * Fallback dummy cfqq for extreme OOM conditions
 393          */
 394         struct cfq_queue oom_cfqq;
 395
 396         u64 last_delayed_sync;
 397 };
 398
 399 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 400 static void cfq_put_queue(struct cfq_queue *cfqq);
 401
 402 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 403                                             enum wl_class_t class,
 404                                             enum wl_type_t type)
 405 {
 406         if (!cfqg)
 407                 return NULL;
 408
 409         if (class == IDLE_WORKLOAD)
 410                 return &cfqg->service_tree_idle;
 411
 412         return &cfqg->service_trees[class][type];
 413 }
 414
 415 enum cfqq_state_flags {
 416         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 417         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 418         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 419         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 420         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 421         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 422         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 423         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 424         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 425         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 426         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 427         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 428         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 429 };
 430
 431 #define CFQ_CFQQ_FNS(name)                                              \
 432 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 433 {                                                                       \
 434         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 435 }                                                                       \
 436 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 437 {                                                                       \
 438         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 439 }                                                                       \
 440 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 441 {                                                                       \
 442         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 443 }
 444
 445 CFQ_CFQQ_FNS(on_rr);
 446 CFQ_CFQQ_FNS(wait_request);
 447 CFQ_CFQQ_FNS(must_dispatch);
 448 CFQ_CFQQ_FNS(must_alloc_slice);
 449 CFQ_CFQQ_FNS(fifo_expire);
 450 CFQ_CFQQ_FNS(idle_window);
 451 CFQ_CFQQ_FNS(prio_changed);
 452 CFQ_CFQQ_FNS(slice_new);
 453 CFQ_CFQQ_FNS(sync);
 454 CFQ_CFQQ_FNS(coop);
 455 CFQ_CFQQ_FNS(split_coop);
 456 CFQ_CFQQ_FNS(deep);
 457 CFQ_CFQQ_FNS(wait_busy);
 458 #undef CFQ_CFQQ_FNS
 459
 460 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 461
 462 /* cfqg stats flags */
 463 enum cfqg_stats_flags {
 464         CFQG_stats_waiting = 0,
 465         CFQG_stats_idling,
 466         CFQG_stats_empty,
 467 };
 468
 469 #define CFQG_FLAG_FNS(name)                                             \
 470 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 471 {                                                                       \
 472         stats->flags |= (1 << CFQG_stats_##name);                       \
 473 }                                                                       \
 474 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 475 {                                                                       \
 476         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 477 }                                                                       \
 478 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 479 {                                                                       \
 480         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 481 }                                                                       \
 482
 483 CFQG_FLAG_FNS(waiting)
 484 CFQG_FLAG_FNS(idling)
 485 CFQG_FLAG_FNS(empty)
 486 #undef CFQG_FLAG_FNS
 487
 488 /* This should be called with the queue_lock held. */
 489 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 490 {
 491         unsigned long long now;
 492
 493         if (!cfqg_stats_waiting(stats))
 494                 return;
 495
 496         now = sched_clock();
 497         if (time_after64(now, stats->start_group_wait_time))
 498                 blkg_stat_add(&stats->group_wait_time,
 499                               now - stats->start_group_wait_time);
 500         cfqg_stats_clear_waiting(stats);
 501 }
 502
 503 /* This should be called with the queue_lock held. */
 504 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 505                                                  struct cfq_group *curr_cfqg)
 506 {
 507         struct cfqg_stats *stats = &cfqg->stats;
 508
 509         if (cfqg_stats_waiting(stats))
 510                 return;
 511         if (cfqg == curr_cfqg)
 512                 return;
 513         stats->start_group_wait_time = sched_clock();
 514         cfqg_stats_mark_waiting(stats);
 515 }
 516
 517 /* This should be called with the queue_lock held. */
 518 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 519 {
 520         unsigned long long now;
 521
 522         if (!cfqg_stats_empty(stats))
 523                 return;
 524
 525         now = sched_clock();
 526         if (time_after64(now, stats->start_empty_time))
 527                 blkg_stat_add(&stats->empty_time,
 528                               now - stats->start_empty_time);
 529         cfqg_stats_clear_empty(stats);
 530 }
 531
 532 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 533 {
 534         blkg_stat_add(&cfqg->stats.dequeue, 1);
 535 }
 536
 537 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 538 {
 539         struct cfqg_stats *stats = &cfqg->stats;
 540
 541         if (blkg_rwstat_total(&stats->queued))
 542                 return;
 543
 544         /*
 545          * group is already marked empty. This can happen if cfqq got new
 546          * request in parent group and moved to this group while being added
 547          * to service tree. Just ignore the event and move on.
 548          */
 549         if (cfqg_stats_empty(stats))
 550                 return;
 551
 552         stats->start_empty_time = sched_clock();
 553         cfqg_stats_mark_empty(stats);
 554 }
 555
 556 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 557 {
 558         struct cfqg_stats *stats = &cfqg->stats;
 559
 560         if (cfqg_stats_idling(stats)) {
 561                 unsigned long long now = sched_clock();
 562
 563                 if (time_after64(now, stats->start_idle_time))
 564                         blkg_stat_add(&stats->idle_time,
 565                                       now - stats->start_idle_time);
 566                 cfqg_stats_clear_idling(stats);
 567         }
 568 }
 569
 570 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 571 {
 572         struct cfqg_stats *stats = &cfqg->stats;
 573
 574         BUG_ON(cfqg_stats_idling(stats));
 575
 576         stats->start_idle_time = sched_clock();
 577         cfqg_stats_mark_idling(stats);
 578 }
 579
 580 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 581 {
 582         struct cfqg_stats *stats = &cfqg->stats;
 583
 584         blkg_stat_add(&stats->avg_queue_size_sum,
 585                       blkg_rwstat_total(&stats->queued));
 586         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 587         cfqg_stats_update_group_wait_time(stats);
 588 }
 589
 590 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 591
 592 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 593 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 594 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 595 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 596 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 597 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 598 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 599
 600 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 601
 602 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 603
 604 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 605 {
 606         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 607 }
 608
 609 static struct cfq_group_data
 610 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 611 {
 612         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 613 }
 614
 615 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 616 {
 617         return pd_to_blkg(&cfqg->pd);
 618 }
 619
 620 static struct blkcg_policy blkcg_policy_cfq;
 621
 622 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 623 {
 624         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 625 }
 626
 627 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 628 {
 629         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 630 }
 631
 632 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 633 {
 634         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 635
 636         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 637 }
 638
 639 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 640                                       struct cfq_group *ancestor)
 641 {
 642         return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
 643                                     cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
 644 }
 645
 646 static inline void cfqg_get(struct cfq_group *cfqg)
 647 {
 648         return blkg_get(cfqg_to_blkg(cfqg));
 649 }
 650
 651 static inline void cfqg_put(struct cfq_group *cfqg)
 652 {
 653         return blkg_put(cfqg_to_blkg(cfqg));
 654 }
 655
 656 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 657         char __pbuf[128];                                               \
 658                                                                         \
 659         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 660         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 661                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 662                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 663                           __pbuf, ##args);                              \
 664 } while (0)
 665
 666 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 667         char __pbuf[128];                                               \
 668                                                                         \
 669         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 670         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 671 } while (0)
 672
 673 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 674                                             struct cfq_group *curr_cfqg, int op,
 675                                             int op_flags)
 676 {
 677         blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, 1);
 678         cfqg_stats_end_empty_time(&cfqg->stats);
 679         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 680 }
 681
 682 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 683                         uint64_t time, unsigned long unaccounted_time)
 684 {
 685         blkg_stat_add(&cfqg->stats.time, time);
 686 #ifdef CONFIG_DEBUG_BLK_CGROUP
 687         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 688 #endif
 689 }
 690
 691 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
 692                                                int op_flags)
 693 {
 694         blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, -1);
 695 }
 696
 697 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
 698                                                int op_flags)
 699 {
 700         blkg_rwstat_add(&cfqg->stats.merged, op, op_flags, 1);
 701 }
 702
 703 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 704                         uint64_t start_time, uint64_t io_start_time, int op,
 705                         int op_flags)
 706 {
 707         struct cfqg_stats *stats = &cfqg->stats;
 708         unsigned long long now = sched_clock();
 709
 710         if (time_after64(now, io_start_time))
 711                 blkg_rwstat_add(&stats->service_time, op, op_flags,
 712                                 now - io_start_time);
 713         if (time_after64(io_start_time, start_time))
 714                 blkg_rwstat_add(&stats->wait_time, op, op_flags,
 715                                 io_start_time - start_time);
 716 }
 717
 718 /* @stats = 0 */
 719 static void cfqg_stats_reset(struct cfqg_stats *stats)
 720 {
 721         /* queued stats shouldn't be cleared */
 722         blkg_rwstat_reset(&stats->merged);
 723         blkg_rwstat_reset(&stats->service_time);
 724         blkg_rwstat_reset(&stats->wait_time);
 725         blkg_stat_reset(&stats->time);
 726 #ifdef CONFIG_DEBUG_BLK_CGROUP
 727         blkg_stat_reset(&stats->unaccounted_time);
 728         blkg_stat_reset(&stats->avg_queue_size_sum);
 729         blkg_stat_reset(&stats->avg_queue_size_samples);
 730         blkg_stat_reset(&stats->dequeue);
 731         blkg_stat_reset(&stats->group_wait_time);
 732         blkg_stat_reset(&stats->idle_time);
 733         blkg_stat_reset(&stats->empty_time);
 734 #endif
 735 }
 736
 737 /* @to += @from */
 738 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 739 {
 740         /* queued stats shouldn't be cleared */
 741         blkg_rwstat_add_aux(&to->merged, &from->merged);
 742         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 743         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 744         blkg_stat_add_aux(&from->time, &from->time);
 745 #ifdef CONFIG_DEBUG_BLK_CGROUP
 746         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 747         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 748         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 749         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 750         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 751         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 752         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 753 #endif
 754 }
 755
 756 /*
 757  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 758  * recursive stats can still account for the amount used by this cfqg after
 759  * it's gone.
 760  */
 761 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 762 {
 763         struct cfq_group *parent = cfqg_parent(cfqg);
 764
 765         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 766
 767         if (unlikely(!parent))
 768                 return;
 769
 770         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 771         cfqg_stats_reset(&cfqg->stats);
 772 }
 773
 774 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 775
 776 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 777 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 778                                       struct cfq_group *ancestor)
 779 {
 780         return true;
 781 }
 782 static inline void cfqg_get(struct cfq_group *cfqg) { }
 783 static inline void cfqg_put(struct cfq_group *cfqg) { }
 784
 785 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 786         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 787                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 788                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 789                                 ##args)
 790 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 791
 792 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 793                         struct cfq_group *curr_cfqg, int op, int op_flags) { }
 794 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 795                         uint64_t time, unsigned long unaccounted_time) { }
 796 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op,
 797                         int op_flags) { }
 798 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op,
 799                         int op_flags) { }
 800 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 801                         uint64_t start_time, uint64_t io_start_time, int op,
 802                         int op_flags) { }
 803
 804 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 805
 806 #define cfq_log(cfqd, fmt, args...)     \
 807         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 808
 809 /* Traverses through cfq group service trees */
 810 #define for_each_cfqg_st(cfqg, i, j, st) \
 811         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 812                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 813                         : &cfqg->service_tree_idle; \
 814                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 815                         (i == IDLE_WORKLOAD && j == 0); \
 816                         j++, st = i < IDLE_WORKLOAD ? \
 817                         &cfqg->service_trees[i][j]: NULL) \
 818
 819 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 820         struct cfq_ttime *ttime, bool group_idle)
 821 {
 822         u64 slice;
 823         if (!sample_valid(ttime->ttime_samples))
 824                 return false;
 825         if (group_idle)
 826                 slice = cfqd->cfq_group_idle;
 827         else
 828                 slice = cfqd->cfq_slice_idle;
 829         return ttime->ttime_mean > slice;
 830 }
 831
 832 static inline bool iops_mode(struct cfq_data *cfqd)
 833 {
 834         /*
 835          * If we are not idling on queues and it is a NCQ drive, parallel
 836          * execution of requests is on and measuring time is not possible
 837          * in most of the cases until and unless we drive shallower queue
 838          * depths and that becomes a performance bottleneck. In such cases
 839          * switch to start providing fairness in terms of number of IOs.
 840          */
 841         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 842                 return true;
 843         else
 844                 return false;
 845 }
 846
 847 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 848 {
 849         if (cfq_class_idle(cfqq))
 850                 return IDLE_WORKLOAD;
 851         if (cfq_class_rt(cfqq))
 852                 return RT_WORKLOAD;
 853         return BE_WORKLOAD;
 854 }
 855
 856
 857 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 858 {
 859         if (!cfq_cfqq_sync(cfqq))
 860                 return ASYNC_WORKLOAD;
 861         if (!cfq_cfqq_idle_window(cfqq))
 862                 return SYNC_NOIDLE_WORKLOAD;
 863         return SYNC_WORKLOAD;
 864 }
 865
 866 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 867                                         struct cfq_data *cfqd,
 868                                         struct cfq_group *cfqg)
 869 {
 870         if (wl_class == IDLE_WORKLOAD)
 871                 return cfqg->service_tree_idle.count;
 872
 873         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 874                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 875                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 876 }
 877
 878 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 879                                         struct cfq_group *cfqg)
 880 {
 881         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 882                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 883 }
 884
 885 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 886 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 887                                        struct cfq_io_cq *cic, struct bio *bio);
 888
 889 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 890 {
 891         /* cic->icq is the first member, %NULL will convert to %NULL */
 892         return container_of(icq, struct cfq_io_cq, icq);
 893 }
 894
 895 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 896                                                struct io_context *ioc)
 897 {
 898         if (ioc)
 899                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 900         return NULL;
 901 }
 902
 903 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 904 {
 905         return cic->cfqq[is_sync];
 906 }
 907
 908 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 909                                 bool is_sync)
 910 {
 911         cic->cfqq[is_sync] = cfqq;
 912 }
 913
 914 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 915 {
 916         return cic->icq.q->elevator->elevator_data;
 917 }
 918
 919 /*
 920  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 921  * set (in which case it could also be direct WRITE).
 922  */
 923 static inline bool cfq_bio_sync(struct bio *bio)
 924 {
 925         return bio_data_dir(bio) == READ || (bio->bi_opf & REQ_SYNC);
 926 }
 927
 928 /*
 929  * scheduler run of queue, if there are requests pending and no one in the
 930  * driver that will restart queueing
 931  */
 932 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 933 {
 934         if (cfqd->busy_queues) {
 935                 cfq_log(cfqd, "schedule dispatch");
 936                 kblockd_schedule_work(&cfqd->unplug_work);
 937         }
 938 }
 939
 940 /*
 941  * Scale schedule slice based on io priority. Use the sync time slice only
 942  * if a queue is marked sync and has sync io queued. A sync queue with async
 943  * io only, should not get full sync slice length.
 944  */
 945 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 946                                  unsigned short prio)
 947 {
 948         u64 base_slice = cfqd->cfq_slice[sync];
 949         u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
 950
 951         WARN_ON(prio >= IOPRIO_BE_NR);
 952
 953         return base_slice + (slice * (4 - prio));
 954 }
 955
 956 static inline u64
 957 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 958 {
 959         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 960 }
 961
 962 /**
 963  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 964  * @charge: disk time being charged
 965  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 966  *
 967  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 968  * scaling is inversely proportional.
 969  *
 970  * scaled = charge / vfraction
 971  *
 972  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 973  */
 974 static inline u64 cfqg_scale_charge(u64 charge,
 975                                     unsigned int vfraction)
 976 {
 977         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 978
 979         /* charge / vfraction */
 980         c <<= CFQ_SERVICE_SHIFT;
 981         return div_u64(c, vfraction);
 982 }
 983
 984 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 985 {
 986         s64 delta = (s64)(vdisktime - min_vdisktime);
 987         if (delta > 0)
 988                 min_vdisktime = vdisktime;
 989
 990         return min_vdisktime;
 991 }
 992
 993 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 994 {
 995         s64 delta = (s64)(vdisktime - min_vdisktime);
 996         if (delta < 0)
 997                 min_vdisktime = vdisktime;
 998
 999         return min_vdisktime;
1000 }
1001
1002 static void update_min_vdisktime(struct cfq_rb_root *st)
1003 {
1004         struct cfq_group *cfqg;
1005
1006         if (st->left) {
1007                 cfqg = rb_entry_cfqg(st->left);
1008                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1009                                                   cfqg->vdisktime);
1010         }
1011 }
1012
1013 /*
1014  * get averaged number of queues of RT/BE priority.
1015  * average is updated, with a formula that gives more weight to higher numbers,
1016  * to quickly follows sudden increases and decrease slowly
1017  */
1018
1019 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1020                                         struct cfq_group *cfqg, bool rt)
1021 {
1022         unsigned min_q, max_q;
1023         unsigned mult  = cfq_hist_divisor - 1;
1024         unsigned round = cfq_hist_divisor / 2;
1025         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1026
1027         min_q = min(cfqg->busy_queues_avg[rt], busy);
1028         max_q = max(cfqg->busy_queues_avg[rt], busy);
1029         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1030                 cfq_hist_divisor;
1031         return cfqg->busy_queues_avg[rt];
1032 }
1033
1034 static inline u64
1035 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1036 {
1037         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1038 }
1039
1040 static inline u64
1041 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1042 {
1043         u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1044         if (cfqd->cfq_latency) {
1045                 /*
1046                  * interested queues (we consider only the ones with the same
1047                  * priority class in the cfq group)
1048                  */
1049                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1050                                                 cfq_class_rt(cfqq));
1051                 u64 sync_slice = cfqd->cfq_slice[1];
1052                 u64 expect_latency = sync_slice * iq;
1053                 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1054
1055                 if (expect_latency > group_slice) {
1056                         u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1057                         u64 low_slice;
1058
1059                         /* scale low_slice according to IO priority
1060                          * and sync vs async */
1061                         low_slice = div64_u64(base_low_slice*slice, sync_slice);
1062                         low_slice = min(slice, low_slice);
1063                         /* the adapted slice value is scaled to fit all iqs
1064                          * into the target latency */
1065                         slice = div64_u64(slice*group_slice, expect_latency);
1066                         slice = max(slice, low_slice);
1067                 }
1068         }
1069         return slice;
1070 }
1071
1072 static inline void
1073 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1074 {
1075         u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1076         u64 now = ktime_get_ns();
1077
1078         cfqq->slice_start = now;
1079         cfqq->slice_end = now + slice;
1080         cfqq->allocated_slice = slice;
1081         cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1082 }
1083
1084 /*
1085  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1086  * isn't valid until the first request from the dispatch is activated
1087  * and the slice time set.
1088  */
1089 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1090 {
1091         if (cfq_cfqq_slice_new(cfqq))
1092                 return false;
1093         if (ktime_get_ns() < cfqq->slice_end)
1094                 return false;
1095
1096         return true;
1097 }
1098
1099 /*
1100  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1101  * We choose the request that is closest to the head right now. Distance
1102  * behind the head is penalized and only allowed to a certain extent.
1103  */
1104 static struct request *
1105 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1106 {
1107         sector_t s1, s2, d1 = 0, d2 = 0;
1108         unsigned long back_max;
1109 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1110 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1111         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1112
1113         if (rq1 == NULL || rq1 == rq2)
1114                 return rq2;
1115         if (rq2 == NULL)
1116                 return rq1;
1117
1118         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1119                 return rq_is_sync(rq1) ? rq1 : rq2;
1120
1121         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1122                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1123
1124         s1 = blk_rq_pos(rq1);
1125         s2 = blk_rq_pos(rq2);
1126
1127         /*
1128          * by definition, 1KiB is 2 sectors
1129          */
1130         back_max = cfqd->cfq_back_max * 2;
1131
1132         /*
1133          * Strict one way elevator _except_ in the case where we allow
1134          * short backward seeks which are biased as twice the cost of a
1135          * similar forward seek.
1136          */
1137         if (s1 >= last)
1138                 d1 = s1 - last;
1139         else if (s1 + back_max >= last)
1140                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1141         else
1142                 wrap |= CFQ_RQ1_WRAP;
1143
1144         if (s2 >= last)
1145                 d2 = s2 - last;
1146         else if (s2 + back_max >= last)
1147                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1148         else
1149                 wrap |= CFQ_RQ2_WRAP;
1150
1151         /* Found required data */
1152
1153         /*
1154          * By doing switch() on the bit mask "wrap" we avoid having to
1155          * check two variables for all permutations: --> faster!
1156          */
1157         switch (wrap) {
1158         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1159                 if (d1 < d2)
1160                         return rq1;
1161                 else if (d2 < d1)
1162                         return rq2;
1163                 else {
1164                         if (s1 >= s2)
1165                                 return rq1;
1166                         else
1167                                 return rq2;
1168                 }
1169
1170         case CFQ_RQ2_WRAP:
1171                 return rq1;
1172         case CFQ_RQ1_WRAP:
1173                 return rq2;
1174         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1175         default:
1176                 /*
1177                  * Since both rqs are wrapped,
1178                  * start with the one that's further behind head
1179                  * (--> only *one* back seek required),
1180                  * since back seek takes more time than forward.
1181                  */
1182                 if (s1 <= s2)
1183                         return rq1;
1184                 else
1185                         return rq2;
1186         }
1187 }
1188
1189 /*
1190  * The below is leftmost cache rbtree addon
1191  */
1192 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1193 {
1194         /* Service tree is empty */
1195         if (!root->count)
1196                 return NULL;
1197
1198         if (!root->left)
1199                 root->left = rb_first(&root->rb);
1200
1201         if (root->left)
1202                 return rb_entry(root->left, struct cfq_queue, rb_node);
1203
1204         return NULL;
1205 }
1206
1207 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1208 {
1209         if (!root->left)
1210                 root->left = rb_first(&root->rb);
1211
1212         if (root->left)
1213                 return rb_entry_cfqg(root->left);
1214
1215         return NULL;
1216 }
1217
1218 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1219 {
1220         rb_erase(n, root);
1221         RB_CLEAR_NODE(n);
1222 }
1223
1224 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1225 {
1226         if (root->left == n)
1227                 root->left = NULL;
1228         rb_erase_init(n, &root->rb);
1229         --root->count;
1230 }
1231
1232 /*
1233  * would be nice to take fifo expire time into account as well
1234  */
1235 static struct request *
1236 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1237                   struct request *last)
1238 {
1239         struct rb_node *rbnext = rb_next(&last->rb_node);
1240         struct rb_node *rbprev = rb_prev(&last->rb_node);
1241         struct request *next = NULL, *prev = NULL;
1242
1243         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1244
1245         if (rbprev)
1246                 prev = rb_entry_rq(rbprev);
1247
1248         if (rbnext)
1249                 next = rb_entry_rq(rbnext);
1250         else {
1251                 rbnext = rb_first(&cfqq->sort_list);
1252                 if (rbnext && rbnext != &last->rb_node)
1253                         next = rb_entry_rq(rbnext);
1254         }
1255
1256         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1257 }
1258
1259 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1260                             struct cfq_queue *cfqq)
1261 {
1262         /*
1263          * just an approximation, should be ok.
1264          */
1265         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1266                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1267 }
1268
1269 static inline s64
1270 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1271 {
1272         return cfqg->vdisktime - st->min_vdisktime;
1273 }
1274
1275 static void
1276 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1277 {
1278         struct rb_node **node = &st->rb.rb_node;
1279         struct rb_node *parent = NULL;
1280         struct cfq_group *__cfqg;
1281         s64 key = cfqg_key(st, cfqg);
1282         int left = 1;
1283
1284         while (*node != NULL) {
1285                 parent = *node;
1286                 __cfqg = rb_entry_cfqg(parent);
1287
1288                 if (key < cfqg_key(st, __cfqg))
1289                         node = &parent->rb_left;
1290                 else {
1291                         node = &parent->rb_right;
1292                         left = 0;
1293                 }
1294         }
1295
1296         if (left)
1297                 st->left = &cfqg->rb_node;
1298
1299         rb_link_node(&cfqg->rb_node, parent, node);
1300         rb_insert_color(&cfqg->rb_node, &st->rb);
1301 }
1302
1303 /*
1304  * This has to be called only on activation of cfqg
1305  */
1306 static void
1307 cfq_update_group_weight(struct cfq_group *cfqg)
1308 {
1309         if (cfqg->new_weight) {
1310                 cfqg->weight = cfqg->new_weight;
1311                 cfqg->new_weight = 0;
1312         }
1313 }
1314
1315 static void
1316 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1317 {
1318         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1319
1320         if (cfqg->new_leaf_weight) {
1321                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1322                 cfqg->new_leaf_weight = 0;
1323         }
1324 }
1325
1326 static void
1327 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1328 {
1329         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1330         struct cfq_group *pos = cfqg;
1331         struct cfq_group *parent;
1332         bool propagate;
1333
1334         /* add to the service tree */
1335         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1336
1337         /*
1338          * Update leaf_weight.  We cannot update weight at this point
1339          * because cfqg might already have been activated and is
1340          * contributing its current weight to the parent's child_weight.
1341          */
1342         cfq_update_group_leaf_weight(cfqg);
1343         __cfq_group_service_tree_add(st, cfqg);
1344
1345         /*
1346          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1347          * entitled to.  vfraction is calculated by walking the tree
1348          * towards the root calculating the fraction it has at each level.
1349          * The compounded ratio is how much vfraction @cfqg owns.
1350          *
1351          * Start with the proportion tasks in this cfqg has against active
1352          * children cfqgs - its leaf_weight against children_weight.
1353          */
1354         propagate = !pos->nr_active++;
1355         pos->children_weight += pos->leaf_weight;
1356         vfr = vfr * pos->leaf_weight / pos->children_weight;
1357
1358         /*
1359          * Compound ->weight walking up the tree.  Both activation and
1360          * vfraction calculation are done in the same loop.  Propagation
1361          * stops once an already activated node is met.  vfraction
1362          * calculation should always continue to the root.
1363          */
1364         while ((parent = cfqg_parent(pos))) {
1365                 if (propagate) {
1366                         cfq_update_group_weight(pos);
1367                         propagate = !parent->nr_active++;
1368                         parent->children_weight += pos->weight;
1369                 }
1370                 vfr = vfr * pos->weight / parent->children_weight;
1371                 pos = parent;
1372         }
1373
1374         cfqg->vfraction = max_t(unsigned, vfr, 1);
1375 }
1376
1377 static inline u64 cfq_get_cfqg_vdisktime_delay(struct cfq_data *cfqd)
1378 {
1379         if (!iops_mode(cfqd))
1380                 return CFQ_SLICE_MODE_GROUP_DELAY;
1381         else
1382                 return CFQ_IOPS_MODE_GROUP_DELAY;
1383 }
1384
1385 static void
1386 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1387 {
1388         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1389         struct cfq_group *__cfqg;
1390         struct rb_node *n;
1391
1392         cfqg->nr_cfqq++;
1393         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1394                 return;
1395
1396         /*
1397          * Currently put the group at the end. Later implement something
1398          * so that groups get lesser vtime based on their weights, so that
1399          * if group does not loose all if it was not continuously backlogged.
1400          */
1401         n = rb_last(&st->rb);
1402         if (n) {
1403                 __cfqg = rb_entry_cfqg(n);
1404                 cfqg->vdisktime = __cfqg->vdisktime +
1405                         cfq_get_cfqg_vdisktime_delay(cfqd);
1406         } else
1407                 cfqg->vdisktime = st->min_vdisktime;
1408         cfq_group_service_tree_add(st, cfqg);
1409 }
1410
1411 static void
1412 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1413 {
1414         struct cfq_group *pos = cfqg;
1415         bool propagate;
1416
1417         /*
1418          * Undo activation from cfq_group_service_tree_add().  Deactivate
1419          * @cfqg and propagate deactivation upwards.
1420          */
1421         propagate = !--pos->nr_active;
1422         pos->children_weight -= pos->leaf_weight;
1423
1424         while (propagate) {
1425                 struct cfq_group *parent = cfqg_parent(pos);
1426
1427                 /* @pos has 0 nr_active at this point */
1428                 WARN_ON_ONCE(pos->children_weight);
1429                 pos->vfraction = 0;
1430
1431                 if (!parent)
1432                         break;
1433
1434                 propagate = !--parent->nr_active;
1435                 parent->children_weight -= pos->weight;
1436                 pos = parent;
1437         }
1438
1439         /* remove from the service tree */
1440         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1441                 cfq_rb_erase(&cfqg->rb_node, st);
1442 }
1443
1444 static void
1445 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1446 {
1447         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1448
1449         BUG_ON(cfqg->nr_cfqq < 1);
1450         cfqg->nr_cfqq--;
1451
1452         /* If there are other cfq queues under this group, don't delete it */
1453         if (cfqg->nr_cfqq)
1454                 return;
1455
1456         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1457         cfq_group_service_tree_del(st, cfqg);
1458         cfqg->saved_wl_slice = 0;
1459         cfqg_stats_update_dequeue(cfqg);
1460 }
1461
1462 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1463                                        u64 *unaccounted_time)
1464 {
1465         u64 slice_used;
1466         u64 now = ktime_get_ns();
1467
1468         /*
1469          * Queue got expired before even a single request completed or
1470          * got expired immediately after first request completion.
1471          */
1472         if (!cfqq->slice_start || cfqq->slice_start == now) {
1473                 /*
1474                  * Also charge the seek time incurred to the group, otherwise
1475                  * if there are mutiple queues in the group, each can dispatch
1476                  * a single request on seeky media and cause lots of seek time
1477                  * and group will never know it.
1478                  */
1479                 slice_used = max_t(u64, (now - cfqq->dispatch_start),
1480                                         jiffies_to_nsecs(1));
1481         } else {
1482                 slice_used = now - cfqq->slice_start;
1483                 if (slice_used > cfqq->allocated_slice) {
1484                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1485                         slice_used = cfqq->allocated_slice;
1486                 }
1487                 if (cfqq->slice_start > cfqq->dispatch_start)
1488                         *unaccounted_time += cfqq->slice_start -
1489                                         cfqq->dispatch_start;
1490         }
1491
1492         return slice_used;
1493 }
1494
1495 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1496                                 struct cfq_queue *cfqq)
1497 {
1498         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1499         u64 used_sl, charge, unaccounted_sl = 0;
1500         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1501                         - cfqg->service_tree_idle.count;
1502         unsigned int vfr;
1503         u64 now = ktime_get_ns();
1504
1505         BUG_ON(nr_sync < 0);
1506         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1507
1508         if (iops_mode(cfqd))
1509                 charge = cfqq->slice_dispatch;
1510         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1511                 charge = cfqq->allocated_slice;
1512
1513         /*
1514          * Can't update vdisktime while on service tree and cfqg->vfraction
1515          * is valid only while on it.  Cache vfr, leave the service tree,
1516          * update vdisktime and go back on.  The re-addition to the tree
1517          * will also update the weights as necessary.
1518          */
1519         vfr = cfqg->vfraction;
1520         cfq_group_service_tree_del(st, cfqg);
1521         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1522         cfq_group_service_tree_add(st, cfqg);
1523
1524         /* This group is being expired. Save the context */
1525         if (cfqd->workload_expires > now) {
1526                 cfqg->saved_wl_slice = cfqd->workload_expires - now;
1527                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1528                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1529         } else
1530                 cfqg->saved_wl_slice = 0;
1531
1532         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1533                                         st->min_vdisktime);
1534         cfq_log_cfqq(cfqq->cfqd, cfqq,
1535                      "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
1536                      used_sl, cfqq->slice_dispatch, charge,
1537                      iops_mode(cfqd), cfqq->nr_sectors);
1538         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1539         cfqg_stats_set_start_empty_time(cfqg);
1540 }
1541
1542 /**
1543  * cfq_init_cfqg_base - initialize base part of a cfq_group
1544  * @cfqg: cfq_group to initialize
1545  *
1546  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1547  * is enabled or not.
1548  */
1549 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1550 {
1551         struct cfq_rb_root *st;
1552         int i, j;
1553
1554         for_each_cfqg_st(cfqg, i, j, st)
1555                 *st = CFQ_RB_ROOT;
1556         RB_CLEAR_NODE(&cfqg->rb_node);
1557
1558         cfqg->ttime.last_end_request = ktime_get_ns();
1559 }
1560
1561 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1562 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1563                             bool on_dfl, bool reset_dev, bool is_leaf_weight);
1564
1565 static void cfqg_stats_exit(struct cfqg_stats *stats)
1566 {
1567         blkg_rwstat_exit(&stats->merged);
1568         blkg_rwstat_exit(&stats->service_time);
1569         blkg_rwstat_exit(&stats->wait_time);
1570         blkg_rwstat_exit(&stats->queued);
1571         blkg_stat_exit(&stats->time);
1572 #ifdef CONFIG_DEBUG_BLK_CGROUP
1573         blkg_stat_exit(&stats->unaccounted_time);
1574         blkg_stat_exit(&stats->avg_queue_size_sum);
1575         blkg_stat_exit(&stats->avg_queue_size_samples);
1576         blkg_stat_exit(&stats->dequeue);
1577         blkg_stat_exit(&stats->group_wait_time);
1578         blkg_stat_exit(&stats->idle_time);
1579         blkg_stat_exit(&stats->empty_time);
1580 #endif
1581 }
1582
1583 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1584 {
1585         if (blkg_rwstat_init(&stats->merged, gfp) ||
1586             blkg_rwstat_init(&stats->service_time, gfp) ||
1587             blkg_rwstat_init(&stats->wait_time, gfp) ||
1588             blkg_rwstat_init(&stats->queued, gfp) ||
1589             blkg_stat_init(&stats->time, gfp))
1590                 goto err;
1591
1592 #ifdef CONFIG_DEBUG_BLK_CGROUP
1593         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1594             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1595             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1596             blkg_stat_init(&stats->dequeue, gfp) ||
1597             blkg_stat_init(&stats->group_wait_time, gfp) ||
1598             blkg_stat_init(&stats->idle_time, gfp) ||
1599             blkg_stat_init(&stats->empty_time, gfp))
1600                 goto err;
1601 #endif
1602         return 0;
1603 err:
1604         cfqg_stats_exit(stats);
1605         return -ENOMEM;
1606 }
1607
1608 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1609 {
1610         struct cfq_group_data *cgd;
1611
1612         cgd = kzalloc(sizeof(*cgd), gfp);
1613         if (!cgd)
1614                 return NULL;
1615         return &cgd->cpd;
1616 }
1617
1618 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1619 {
1620         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1621         unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1622                               CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1623
1624         if (cpd_to_blkcg(cpd) == &blkcg_root)
1625                 weight *= 2;
1626
1627         cgd->weight = weight;
1628         cgd->leaf_weight = weight;
1629 }
1630
1631 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1632 {
1633         kfree(cpd_to_cfqgd(cpd));
1634 }
1635
1636 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1637 {
1638         struct blkcg *blkcg = cpd_to_blkcg(cpd);
1639         bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1640         unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1641
1642         if (blkcg == &blkcg_root)
1643                 weight *= 2;
1644
1645         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1646         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1647 }
1648
1649 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1650 {
1651         struct cfq_group *cfqg;
1652
1653         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1654         if (!cfqg)
1655                 return NULL;
1656
1657         cfq_init_cfqg_base(cfqg);
1658         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1659                 kfree(cfqg);
1660                 return NULL;
1661         }
1662
1663         return &cfqg->pd;
1664 }
1665
1666 static void cfq_pd_init(struct blkg_policy_data *pd)
1667 {
1668         struct cfq_group *cfqg = pd_to_cfqg(pd);
1669         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1670
1671         cfqg->weight = cgd->weight;
1672         cfqg->leaf_weight = cgd->leaf_weight;
1673 }
1674
1675 static void cfq_pd_offline(struct blkg_policy_data *pd)
1676 {
1677         struct cfq_group *cfqg = pd_to_cfqg(pd);
1678         int i;
1679
1680         for (i = 0; i < IOPRIO_BE_NR; i++) {
1681                 if (cfqg->async_cfqq[0][i])
1682                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1683                 if (cfqg->async_cfqq[1][i])
1684                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1685         }
1686
1687         if (cfqg->async_idle_cfqq)
1688                 cfq_put_queue(cfqg->async_idle_cfqq);
1689
1690         /*
1691          * @blkg is going offline and will be ignored by
1692          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1693          * that they don't get lost.  If IOs complete after this point, the
1694          * stats for them will be lost.  Oh well...
1695          */
1696         cfqg_stats_xfer_dead(cfqg);
1697 }
1698
1699 static void cfq_pd_free(struct blkg_policy_data *pd)
1700 {
1701         struct cfq_group *cfqg = pd_to_cfqg(pd);
1702
1703         cfqg_stats_exit(&cfqg->stats);
1704         return kfree(cfqg);
1705 }
1706
1707 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1708 {
1709         struct cfq_group *cfqg = pd_to_cfqg(pd);
1710
1711         cfqg_stats_reset(&cfqg->stats);
1712 }
1713
1714 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1715                                          struct blkcg *blkcg)
1716 {
1717         struct blkcg_gq *blkg;
1718
1719         blkg = blkg_lookup(blkcg, cfqd->queue);
1720         if (likely(blkg))
1721                 return blkg_to_cfqg(blkg);
1722         return NULL;
1723 }
1724
1725 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1726 {
1727         cfqq->cfqg = cfqg;
1728         /* cfqq reference on cfqg */
1729         cfqg_get(cfqg);
1730 }
1731
1732 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1733                                      struct blkg_policy_data *pd, int off)
1734 {
1735         struct cfq_group *cfqg = pd_to_cfqg(pd);
1736
1737         if (!cfqg->dev_weight)
1738                 return 0;
1739         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1740 }
1741
1742 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1743 {
1744         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1745                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1746                           0, false);
1747         return 0;
1748 }
1749
1750 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1751                                           struct blkg_policy_data *pd, int off)
1752 {
1753         struct cfq_group *cfqg = pd_to_cfqg(pd);
1754
1755         if (!cfqg->dev_leaf_weight)
1756                 return 0;
1757         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1758 }
1759
1760 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1761 {
1762         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1763                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1764                           0, false);
1765         return 0;
1766 }
1767
1768 static int cfq_print_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->weight;
1776
1777         seq_printf(sf, "%u\n", val);
1778         return 0;
1779 }
1780
1781 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1782 {
1783         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1784         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1785         unsigned int val = 0;
1786
1787         if (cgd)
1788                 val = cgd->leaf_weight;
1789
1790         seq_printf(sf, "%u\n", val);
1791         return 0;
1792 }
1793
1794 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1795                                         char *buf, size_t nbytes, loff_t off,
1796                                         bool on_dfl, bool is_leaf_weight)
1797 {
1798         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1799         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1800         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1801         struct blkg_conf_ctx ctx;
1802         struct cfq_group *cfqg;
1803         struct cfq_group_data *cfqgd;
1804         int ret;
1805         u64 v;
1806
1807         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1808         if (ret)
1809                 return ret;
1810
1811         if (sscanf(ctx.body, "%llu", &v) == 1) {
1812                 /* require "default" on dfl */
1813                 ret = -ERANGE;
1814                 if (!v && on_dfl)
1815                         goto out_finish;
1816         } else if (!strcmp(strim(ctx.body), "default")) {
1817                 v = 0;
1818         } else {
1819                 ret = -EINVAL;
1820                 goto out_finish;
1821         }
1822
1823         cfqg = blkg_to_cfqg(ctx.blkg);
1824         cfqgd = blkcg_to_cfqgd(blkcg);
1825
1826         ret = -ERANGE;
1827         if (!v || (v >= min && v <= max)) {
1828                 if (!is_leaf_weight) {
1829                         cfqg->dev_weight = v;
1830                         cfqg->new_weight = v ?: cfqgd->weight;
1831                 } else {
1832                         cfqg->dev_leaf_weight = v;
1833                         cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1834                 }
1835                 ret = 0;
1836         }
1837 out_finish:
1838         blkg_conf_finish(&ctx);
1839         return ret ?: nbytes;
1840 }
1841
1842 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1843                                       char *buf, size_t nbytes, loff_t off)
1844 {
1845         return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1846 }
1847
1848 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1849                                            char *buf, size_t nbytes, loff_t off)
1850 {
1851         return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1852 }
1853
1854 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1855                             bool on_dfl, bool reset_dev, bool is_leaf_weight)
1856 {
1857         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1858         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1859         struct blkcg *blkcg = css_to_blkcg(css);
1860         struct blkcg_gq *blkg;
1861         struct cfq_group_data *cfqgd;
1862         int ret = 0;
1863
1864         if (val < min || val > max)
1865                 return -ERANGE;
1866
1867         spin_lock_irq(&blkcg->lock);
1868         cfqgd = blkcg_to_cfqgd(blkcg);
1869         if (!cfqgd) {
1870                 ret = -EINVAL;
1871                 goto out;
1872         }
1873
1874         if (!is_leaf_weight)
1875                 cfqgd->weight = val;
1876         else
1877                 cfqgd->leaf_weight = val;
1878
1879         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1880                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1881
1882                 if (!cfqg)
1883                         continue;
1884
1885                 if (!is_leaf_weight) {
1886                         if (reset_dev)
1887                                 cfqg->dev_weight = 0;
1888                         if (!cfqg->dev_weight)
1889                                 cfqg->new_weight = cfqgd->weight;
1890                 } else {
1891                         if (reset_dev)
1892                                 cfqg->dev_leaf_weight = 0;
1893                         if (!cfqg->dev_leaf_weight)
1894                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1895                 }
1896         }
1897
1898 out:
1899         spin_unlock_irq(&blkcg->lock);
1900         return ret;
1901 }
1902
1903 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1904                           u64 val)
1905 {
1906         return __cfq_set_weight(css, val, false, false, false);
1907 }
1908
1909 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1910                                struct cftype *cft, u64 val)
1911 {
1912         return __cfq_set_weight(css, val, false, false, true);
1913 }
1914
1915 static int cfqg_print_stat(struct seq_file *sf, void *v)
1916 {
1917         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1918                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1919         return 0;
1920 }
1921
1922 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1923 {
1924         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1925                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1926         return 0;
1927 }
1928
1929 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1930                                       struct blkg_policy_data *pd, int off)
1931 {
1932         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1933                                           &blkcg_policy_cfq, off);
1934         return __blkg_prfill_u64(sf, pd, sum);
1935 }
1936
1937 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1938                                         struct blkg_policy_data *pd, int off)
1939 {
1940         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1941                                                         &blkcg_policy_cfq, off);
1942         return __blkg_prfill_rwstat(sf, pd, &sum);
1943 }
1944
1945 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1946 {
1947         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1948                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1949                           seq_cft(sf)->private, false);
1950         return 0;
1951 }
1952
1953 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1954 {
1955         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1956                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1957                           seq_cft(sf)->private, true);
1958         return 0;
1959 }
1960
1961 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1962                                int off)
1963 {
1964         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1965
1966         return __blkg_prfill_u64(sf, pd, sum >> 9);
1967 }
1968
1969 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1970 {
1971         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1972                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1973         return 0;
1974 }
1975
1976 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1977                                          struct blkg_policy_data *pd, int off)
1978 {
1979         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1980                                         offsetof(struct blkcg_gq, stat_bytes));
1981         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1982                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1983
1984         return __blkg_prfill_u64(sf, pd, sum >> 9);
1985 }
1986
1987 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1988 {
1989         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1990                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1991                           false);
1992         return 0;
1993 }
1994
1995 #ifdef CONFIG_DEBUG_BLK_CGROUP
1996 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1997                                       struct blkg_policy_data *pd, int off)
1998 {
1999         struct cfq_group *cfqg = pd_to_cfqg(pd);
2000         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
2001         u64 v = 0;
2002
2003         if (samples) {
2004                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
2005                 v = div64_u64(v, samples);
2006         }
2007         __blkg_prfill_u64(sf, pd, v);
2008         return 0;
2009 }
2010
2011 /* print avg_queue_size */
2012 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
2013 {
2014         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
2015                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
2016                           0, false);
2017         return 0;
2018 }
2019 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2020
2021 static struct cftype cfq_blkcg_legacy_files[] = {
2022         /* on root, weight is mapped to leaf_weight */
2023         {
2024                 .name = "weight_device",
2025                 .flags = CFTYPE_ONLY_ON_ROOT,
2026                 .seq_show = cfqg_print_leaf_weight_device,
2027                 .write = cfqg_set_leaf_weight_device,
2028         },
2029         {
2030                 .name = "weight",
2031                 .flags = CFTYPE_ONLY_ON_ROOT,
2032                 .seq_show = cfq_print_leaf_weight,
2033                 .write_u64 = cfq_set_leaf_weight,
2034         },
2035
2036         /* no such mapping necessary for !roots */
2037         {
2038                 .name = "weight_device",
2039                 .flags = CFTYPE_NOT_ON_ROOT,
2040                 .seq_show = cfqg_print_weight_device,
2041                 .write = cfqg_set_weight_device,
2042         },
2043         {
2044                 .name = "weight",
2045                 .flags = CFTYPE_NOT_ON_ROOT,
2046                 .seq_show = cfq_print_weight,
2047                 .write_u64 = cfq_set_weight,
2048         },
2049
2050         {
2051                 .name = "leaf_weight_device",
2052                 .seq_show = cfqg_print_leaf_weight_device,
2053                 .write = cfqg_set_leaf_weight_device,
2054         },
2055         {
2056                 .name = "leaf_weight",
2057                 .seq_show = cfq_print_leaf_weight,
2058                 .write_u64 = cfq_set_leaf_weight,
2059         },
2060
2061         /* statistics, covers only the tasks in the cfqg */
2062         {
2063                 .name = "time",
2064                 .private = offsetof(struct cfq_group, stats.time),
2065                 .seq_show = cfqg_print_stat,
2066         },
2067         {
2068                 .name = "sectors",
2069                 .seq_show = cfqg_print_stat_sectors,
2070         },
2071         {
2072                 .name = "io_service_bytes",
2073                 .private = (unsigned long)&blkcg_policy_cfq,
2074                 .seq_show = blkg_print_stat_bytes,
2075         },
2076         {
2077                 .name = "io_serviced",
2078                 .private = (unsigned long)&blkcg_policy_cfq,
2079                 .seq_show = blkg_print_stat_ios,
2080         },
2081         {
2082                 .name = "io_service_time",
2083                 .private = offsetof(struct cfq_group, stats.service_time),
2084                 .seq_show = cfqg_print_rwstat,
2085         },
2086         {
2087                 .name = "io_wait_time",
2088                 .private = offsetof(struct cfq_group, stats.wait_time),
2089                 .seq_show = cfqg_print_rwstat,
2090         },
2091         {
2092                 .name = "io_merged",
2093                 .private = offsetof(struct cfq_group, stats.merged),
2094                 .seq_show = cfqg_print_rwstat,
2095         },
2096         {
2097                 .name = "io_queued",
2098                 .private = offsetof(struct cfq_group, stats.queued),
2099                 .seq_show = cfqg_print_rwstat,
2100         },
2101
2102         /* the same statictics which cover the cfqg and its descendants */
2103         {
2104                 .name = "time_recursive",
2105                 .private = offsetof(struct cfq_group, stats.time),
2106                 .seq_show = cfqg_print_stat_recursive,
2107         },
2108         {
2109                 .name = "sectors_recursive",
2110                 .seq_show = cfqg_print_stat_sectors_recursive,
2111         },
2112         {
2113                 .name = "io_service_bytes_recursive",
2114                 .private = (unsigned long)&blkcg_policy_cfq,
2115                 .seq_show = blkg_print_stat_bytes_recursive,
2116         },
2117         {
2118                 .name = "io_serviced_recursive",
2119                 .private = (unsigned long)&blkcg_policy_cfq,
2120                 .seq_show = blkg_print_stat_ios_recursive,
2121         },
2122         {
2123                 .name = "io_service_time_recursive",
2124                 .private = offsetof(struct cfq_group, stats.service_time),
2125                 .seq_show = cfqg_print_rwstat_recursive,
2126         },
2127         {
2128                 .name = "io_wait_time_recursive",
2129                 .private = offsetof(struct cfq_group, stats.wait_time),
2130                 .seq_show = cfqg_print_rwstat_recursive,
2131         },
2132         {
2133                 .name = "io_merged_recursive",
2134                 .private = offsetof(struct cfq_group, stats.merged),
2135                 .seq_show = cfqg_print_rwstat_recursive,
2136         },
2137         {
2138                 .name = "io_queued_recursive",
2139                 .private = offsetof(struct cfq_group, stats.queued),
2140                 .seq_show = cfqg_print_rwstat_recursive,
2141         },
2142 #ifdef CONFIG_DEBUG_BLK_CGROUP
2143         {
2144                 .name = "avg_queue_size",
2145                 .seq_show = cfqg_print_avg_queue_size,
2146         },
2147         {
2148                 .name = "group_wait_time",
2149                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2150                 .seq_show = cfqg_print_stat,
2151         },
2152         {
2153                 .name = "idle_time",
2154                 .private = offsetof(struct cfq_group, stats.idle_time),
2155                 .seq_show = cfqg_print_stat,
2156         },
2157         {
2158                 .name = "empty_time",
2159                 .private = offsetof(struct cfq_group, stats.empty_time),
2160                 .seq_show = cfqg_print_stat,
2161         },
2162         {
2163                 .name = "dequeue",
2164                 .private = offsetof(struct cfq_group, stats.dequeue),
2165                 .seq_show = cfqg_print_stat,
2166         },
2167         {
2168                 .name = "unaccounted_time",
2169                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2170                 .seq_show = cfqg_print_stat,
2171         },
2172 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2173         { }     /* terminate */
2174 };
2175
2176 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2177 {
2178         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2179         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2180
2181         seq_printf(sf, "default %u\n", cgd->weight);
2182         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2183                           &blkcg_policy_cfq, 0, false);
2184         return 0;
2185 }
2186
2187 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2188                                      char *buf, size_t nbytes, loff_t off)
2189 {
2190         char *endp;
2191         int ret;
2192         u64 v;
2193
2194         buf = strim(buf);
2195
2196         /* "WEIGHT" or "default WEIGHT" sets the default weight */
2197         v = simple_strtoull(buf, &endp, 0);
2198         if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2199                 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2200                 return ret ?: nbytes;
2201         }
2202
2203         /* "MAJ:MIN WEIGHT" */
2204         return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2205 }
2206
2207 static struct cftype cfq_blkcg_files[] = {
2208         {
2209                 .name = "weight",
2210                 .flags = CFTYPE_NOT_ON_ROOT,
2211                 .seq_show = cfq_print_weight_on_dfl,
2212                 .write = cfq_set_weight_on_dfl,
2213         },
2214         { }     /* terminate */
2215 };
2216
2217 #else /* GROUP_IOSCHED */
2218 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2219                                          struct blkcg *blkcg)
2220 {
2221         return cfqd->root_group;
2222 }
2223
2224 static inline void
2225 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2226         cfqq->cfqg = cfqg;
2227 }
2228
2229 #endif /* GROUP_IOSCHED */
2230
2231 /*
2232  * The cfqd->service_trees holds all pending cfq_queue's that have
2233  * requests waiting to be processed. It is sorted in the order that
2234  * we will service the queues.
2235  */
2236 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2237                                  bool add_front)
2238 {
2239         struct rb_node **p, *parent;
2240         struct cfq_queue *__cfqq;
2241         u64 rb_key;
2242         struct cfq_rb_root *st;
2243         int left;
2244         int new_cfqq = 1;
2245         u64 now = ktime_get_ns();
2246
2247         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2248         if (cfq_class_idle(cfqq)) {
2249                 rb_key = CFQ_IDLE_DELAY;
2250                 parent = rb_last(&st->rb);
2251                 if (parent && parent != &cfqq->rb_node) {
2252                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2253                         rb_key += __cfqq->rb_key;
2254                 } else
2255                         rb_key += now;
2256         } else if (!add_front) {
2257                 /*
2258                  * Get our rb key offset. Subtract any residual slice
2259                  * value carried from last service. A negative resid
2260                  * count indicates slice overrun, and this should position
2261                  * the next service time further away in the tree.
2262                  */
2263                 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2264                 rb_key -= cfqq->slice_resid;
2265                 cfqq->slice_resid = 0;
2266         } else {
2267                 rb_key = -NSEC_PER_SEC;
2268                 __cfqq = cfq_rb_first(st);
2269                 rb_key += __cfqq ? __cfqq->rb_key : now;
2270         }
2271
2272         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2273                 new_cfqq = 0;
2274                 /*
2275                  * same position, nothing more to do
2276                  */
2277                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2278                         return;
2279
2280                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2281                 cfqq->service_tree = NULL;
2282         }
2283
2284         left = 1;
2285         parent = NULL;
2286         cfqq->service_tree = st;
2287         p = &st->rb.rb_node;
2288         while (*p) {
2289                 parent = *p;
2290                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2291
2292                 /*
2293                  * sort by key, that represents service time.
2294                  */
2295                 if (rb_key < __cfqq->rb_key)
2296                         p = &parent->rb_left;
2297                 else {
2298                         p = &parent->rb_right;
2299                         left = 0;
2300                 }
2301         }
2302
2303         if (left)
2304                 st->left = &cfqq->rb_node;
2305
2306         cfqq->rb_key = rb_key;
2307         rb_link_node(&cfqq->rb_node, parent, p);
2308         rb_insert_color(&cfqq->rb_node, &st->rb);
2309         st->count++;
2310         if (add_front || !new_cfqq)
2311                 return;
2312         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2313 }
2314
2315 static struct cfq_queue *
2316 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2317                      sector_t sector, struct rb_node **ret_parent,
2318                      struct rb_node ***rb_link)
2319 {
2320         struct rb_node **p, *parent;
2321         struct cfq_queue *cfqq = NULL;
2322
2323         parent = NULL;
2324         p = &root->rb_node;
2325         while (*p) {
2326                 struct rb_node **n;
2327
2328                 parent = *p;
2329                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2330
2331                 /*
2332                  * Sort strictly based on sector.  Smallest to the left,
2333                  * largest to the right.
2334                  */
2335                 if (sector > blk_rq_pos(cfqq->next_rq))
2336                         n = &(*p)->rb_right;
2337                 else if (sector < blk_rq_pos(cfqq->next_rq))
2338                         n = &(*p)->rb_left;
2339                 else
2340                         break;
2341                 p = n;
2342                 cfqq = NULL;
2343         }
2344
2345         *ret_parent = parent;
2346         if (rb_link)
2347                 *rb_link = p;
2348         return cfqq;
2349 }
2350
2351 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2352 {
2353         struct rb_node **p, *parent;
2354         struct cfq_queue *__cfqq;
2355
2356         if (cfqq->p_root) {
2357                 rb_erase(&cfqq->p_node, cfqq->p_root);
2358                 cfqq->p_root = NULL;
2359         }
2360
2361         if (cfq_class_idle(cfqq))
2362                 return;
2363         if (!cfqq->next_rq)
2364                 return;
2365
2366         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2367         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2368                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2369         if (!__cfqq) {
2370                 rb_link_node(&cfqq->p_node, parent, p);
2371                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2372         } else
2373                 cfqq->p_root = NULL;
2374 }
2375
2376 /*
2377  * Update cfqq's position in the service tree.
2378  */
2379 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2380 {
2381         /*
2382          * Resorting requires the cfqq to be on the RR list already.
2383          */
2384         if (cfq_cfqq_on_rr(cfqq)) {
2385                 cfq_service_tree_add(cfqd, cfqq, 0);
2386                 cfq_prio_tree_add(cfqd, cfqq);
2387         }
2388 }
2389
2390 /*
2391  * add to busy list of queues for service, trying to be fair in ordering
2392  * the pending list according to last request service
2393  */
2394 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2395 {
2396         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2397         BUG_ON(cfq_cfqq_on_rr(cfqq));
2398         cfq_mark_cfqq_on_rr(cfqq);
2399         cfqd->busy_queues++;
2400         if (cfq_cfqq_sync(cfqq))
2401                 cfqd->busy_sync_queues++;
2402
2403         cfq_resort_rr_list(cfqd, cfqq);
2404 }
2405
2406 /*
2407  * Called when the cfqq no longer has requests pending, remove it from
2408  * the service tree.
2409  */
2410 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2411 {
2412         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2413         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2414         cfq_clear_cfqq_on_rr(cfqq);
2415
2416         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2417                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2418                 cfqq->service_tree = NULL;
2419         }
2420         if (cfqq->p_root) {
2421                 rb_erase(&cfqq->p_node, cfqq->p_root);
2422                 cfqq->p_root = NULL;
2423         }
2424
2425         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2426         BUG_ON(!cfqd->busy_queues);
2427         cfqd->busy_queues--;
2428         if (cfq_cfqq_sync(cfqq))
2429                 cfqd->busy_sync_queues--;
2430 }
2431
2432 /*
2433  * rb tree support functions
2434  */
2435 static void cfq_del_rq_rb(struct request *rq)
2436 {
2437         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2438         const int sync = rq_is_sync(rq);
2439
2440         BUG_ON(!cfqq->queued[sync]);
2441         cfqq->queued[sync]--;
2442
2443         elv_rb_del(&cfqq->sort_list, rq);
2444
2445         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2446                 /*
2447                  * Queue will be deleted from service tree when we actually
2448                  * expire it later. Right now just remove it from prio tree
2449                  * as it is empty.
2450                  */
2451                 if (cfqq->p_root) {
2452                         rb_erase(&cfqq->p_node, cfqq->p_root);
2453                         cfqq->p_root = NULL;
2454                 }
2455         }
2456 }
2457
2458 static void cfq_add_rq_rb(struct request *rq)
2459 {
2460         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2461         struct cfq_data *cfqd = cfqq->cfqd;
2462         struct request *prev;
2463
2464         cfqq->queued[rq_is_sync(rq)]++;
2465
2466         elv_rb_add(&cfqq->sort_list, rq);
2467
2468         if (!cfq_cfqq_on_rr(cfqq))
2469                 cfq_add_cfqq_rr(cfqd, cfqq);
2470
2471         /*
2472          * check if this request is a better next-serve candidate
2473          */
2474         prev = cfqq->next_rq;
2475         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2476
2477         /*
2478          * adjust priority tree position, if ->next_rq changes
2479          */
2480         if (prev != cfqq->next_rq)
2481                 cfq_prio_tree_add(cfqd, cfqq);
2482
2483         BUG_ON(!cfqq->next_rq);
2484 }
2485
2486 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2487 {
2488         elv_rb_del(&cfqq->sort_list, rq);
2489         cfqq->queued[rq_is_sync(rq)]--;
2490         cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
2491         cfq_add_rq_rb(rq);
2492         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2493                                  req_op(rq), rq->cmd_flags);
2494 }
2495
2496 static struct request *
2497 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2498 {
2499         struct task_struct *tsk = current;
2500         struct cfq_io_cq *cic;
2501         struct cfq_queue *cfqq;
2502
2503         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2504         if (!cic)
2505                 return NULL;
2506
2507         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2508         if (cfqq)
2509                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2510
2511         return NULL;
2512 }
2513
2514 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2515 {
2516         struct cfq_data *cfqd = q->elevator->elevator_data;
2517
2518         cfqd->rq_in_driver++;
2519         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2520                                                 cfqd->rq_in_driver);
2521
2522         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2523 }
2524
2525 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2526 {
2527         struct cfq_data *cfqd = q->elevator->elevator_data;
2528
2529         WARN_ON(!cfqd->rq_in_driver);
2530         cfqd->rq_in_driver--;
2531         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2532                                                 cfqd->rq_in_driver);
2533 }
2534
2535 static void cfq_remove_request(struct request *rq)
2536 {
2537         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2538
2539         if (cfqq->next_rq == rq)
2540                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2541
2542         list_del_init(&rq->queuelist);
2543         cfq_del_rq_rb(rq);
2544
2545         cfqq->cfqd->rq_queued--;
2546         cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags);
2547         if (rq->cmd_flags & REQ_PRIO) {
2548                 WARN_ON(!cfqq->prio_pending);
2549                 cfqq->prio_pending--;
2550         }
2551 }
2552
2553 static int cfq_merge(struct request_queue *q, struct request **req,
2554                      struct bio *bio)
2555 {
2556         struct cfq_data *cfqd = q->elevator->elevator_data;
2557         struct request *__rq;
2558
2559         __rq = cfq_find_rq_fmerge(cfqd, bio);
2560         if (__rq && elv_bio_merge_ok(__rq, bio)) {
2561                 *req = __rq;
2562                 return ELEVATOR_FRONT_MERGE;
2563         }
2564
2565         return ELEVATOR_NO_MERGE;
2566 }
2567
2568 static void cfq_merged_request(struct request_queue *q, struct request *req,
2569                                int type)
2570 {
2571         if (type == ELEVATOR_FRONT_MERGE) {
2572                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2573
2574                 cfq_reposition_rq_rb(cfqq, req);
2575         }
2576 }
2577
2578 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2579                                 struct bio *bio)
2580 {
2581         cfqg_stats_update_io_merged(RQ_CFQG(req), bio_op(bio), bio->bi_opf);
2582 }
2583
2584 static void
2585 cfq_merged_requests(struct request_queue *q, struct request *rq,
2586                     struct request *next)
2587 {
2588         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2589         struct cfq_data *cfqd = q->elevator->elevator_data;
2590
2591         /*
2592          * reposition in fifo if next is older than rq
2593          */
2594         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2595             next->fifo_time < rq->fifo_time &&
2596             cfqq == RQ_CFQQ(next)) {
2597                 list_move(&rq->queuelist, &next->queuelist);
2598                 rq->fifo_time = next->fifo_time;
2599         }
2600
2601         if (cfqq->next_rq == next)
2602                 cfqq->next_rq = rq;
2603         cfq_remove_request(next);
2604         cfqg_stats_update_io_merged(RQ_CFQG(rq), req_op(next), next->cmd_flags);
2605
2606         cfqq = RQ_CFQQ(next);
2607         /*
2608          * all requests of this queue are merged to other queues, delete it
2609          * from the service tree. If it's the active_queue,
2610          * cfq_dispatch_requests() will choose to expire it or do idle
2611          */
2612         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2613             cfqq != cfqd->active_queue)
2614                 cfq_del_cfqq_rr(cfqd, cfqq);
2615 }
2616
2617 static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2618                                struct bio *bio)
2619 {
2620         struct cfq_data *cfqd = q->elevator->elevator_data;
2621         struct cfq_io_cq *cic;
2622         struct cfq_queue *cfqq;
2623
2624         /*
2625          * Disallow merge of a sync bio into an async request.
2626          */
2627         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2628                 return false;
2629
2630         /*
2631          * Lookup the cfqq that this bio will be queued with and allow
2632          * merge only if rq is queued there.
2633          */
2634         cic = cfq_cic_lookup(cfqd, current->io_context);
2635         if (!cic)
2636                 return false;
2637
2638         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2639         return cfqq == RQ_CFQQ(rq);
2640 }
2641
2642 static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2643                               struct request *next)
2644 {
2645         return RQ_CFQQ(rq) == RQ_CFQQ(next);
2646 }
2647
2648 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2649 {
2650         hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2651         cfqg_stats_update_idle_time(cfqq->cfqg);
2652 }
2653
2654 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2655                                    struct cfq_queue *cfqq)
2656 {
2657         if (cfqq) {
2658                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2659                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2660                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2661                 cfqq->slice_start = 0;
2662                 cfqq->dispatch_start = ktime_get_ns();
2663                 cfqq->allocated_slice = 0;
2664                 cfqq->slice_end = 0;
2665                 cfqq->slice_dispatch = 0;
2666                 cfqq->nr_sectors = 0;
2667
2668                 cfq_clear_cfqq_wait_request(cfqq);
2669                 cfq_clear_cfqq_must_dispatch(cfqq);
2670                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2671                 cfq_clear_cfqq_fifo_expire(cfqq);
2672                 cfq_mark_cfqq_slice_new(cfqq);
2673
2674                 cfq_del_timer(cfqd, cfqq);
2675         }
2676
2677         cfqd->active_queue = cfqq;
2678 }
2679
2680 /*
2681  * current cfqq expired its slice (or was too idle), select new one
2682  */
2683 static void
2684 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2685                     bool timed_out)
2686 {
2687         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2688
2689         if (cfq_cfqq_wait_request(cfqq))
2690                 cfq_del_timer(cfqd, cfqq);
2691
2692         cfq_clear_cfqq_wait_request(cfqq);
2693         cfq_clear_cfqq_wait_busy(cfqq);
2694
2695         /*
2696          * If this cfqq is shared between multiple processes, check to
2697          * make sure that those processes are still issuing I/Os within
2698          * the mean seek distance.  If not, it may be time to break the
2699          * queues apart again.
2700          */
2701         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2702                 cfq_mark_cfqq_split_coop(cfqq);
2703
2704         /*
2705          * store what was left of this slice, if the queue idled/timed out
2706          */
2707         if (timed_out) {
2708                 if (cfq_cfqq_slice_new(cfqq))
2709                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2710                 else
2711                         cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2712                 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2713         }
2714
2715         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2716
2717         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2718                 cfq_del_cfqq_rr(cfqd, cfqq);
2719
2720         cfq_resort_rr_list(cfqd, cfqq);
2721
2722         if (cfqq == cfqd->active_queue)
2723                 cfqd->active_queue = NULL;
2724
2725         if (cfqd->active_cic) {
2726                 put_io_context(cfqd->active_cic->icq.ioc);
2727                 cfqd->active_cic = NULL;
2728         }
2729 }
2730
2731 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2732 {
2733         struct cfq_queue *cfqq = cfqd->active_queue;
2734
2735         if (cfqq)
2736                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2737 }
2738
2739 /*
2740  * Get next queue for service. Unless we have a queue preemption,
2741  * we'll simply select the first cfqq in the service tree.
2742  */
2743 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2744 {
2745         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2746                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2747
2748         if (!cfqd->rq_queued)
2749                 return NULL;
2750
2751         /* There is nothing to dispatch */
2752         if (!st)
2753                 return NULL;
2754         if (RB_EMPTY_ROOT(&st->rb))
2755                 return NULL;
2756         return cfq_rb_first(st);
2757 }
2758
2759 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2760 {
2761         struct cfq_group *cfqg;
2762         struct cfq_queue *cfqq;
2763         int i, j;
2764         struct cfq_rb_root *st;
2765
2766         if (!cfqd->rq_queued)
2767                 return NULL;
2768
2769         cfqg = cfq_get_next_cfqg(cfqd);
2770         if (!cfqg)
2771                 return NULL;
2772
2773         for_each_cfqg_st(cfqg, i, j, st)
2774                 if ((cfqq = cfq_rb_first(st)) != NULL)
2775                         return cfqq;
2776         return NULL;
2777 }
2778
2779 /*
2780  * Get and set a new active queue for service.
2781  */
2782 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2783                                               struct cfq_queue *cfqq)
2784 {
2785         if (!cfqq)
2786                 cfqq = cfq_get_next_queue(cfqd);
2787
2788         __cfq_set_active_queue(cfqd, cfqq);
2789         return cfqq;
2790 }
2791
2792 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2793                                           struct request *rq)
2794 {
2795         if (blk_rq_pos(rq) >= cfqd->last_position)
2796                 return blk_rq_pos(rq) - cfqd->last_position;
2797         else
2798                 return cfqd->last_position - blk_rq_pos(rq);
2799 }
2800
2801 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2802                                struct request *rq)
2803 {
2804         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2805 }
2806
2807 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2808                                     struct cfq_queue *cur_cfqq)
2809 {
2810         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2811         struct rb_node *parent, *node;
2812         struct cfq_queue *__cfqq;
2813         sector_t sector = cfqd->last_position;
2814
2815         if (RB_EMPTY_ROOT(root))
2816                 return NULL;
2817
2818         /*
2819          * First, if we find a request starting at the end of the last
2820          * request, choose it.
2821          */
2822         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2823         if (__cfqq)
2824                 return __cfqq;
2825
2826         /*
2827          * If the exact sector wasn't found, the parent of the NULL leaf
2828          * will contain the closest sector.
2829          */
2830         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2831         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2832                 return __cfqq;
2833
2834         if (blk_rq_pos(__cfqq->next_rq) < sector)
2835                 node = rb_next(&__cfqq->p_node);
2836         else
2837                 node = rb_prev(&__cfqq->p_node);
2838         if (!node)
2839                 return NULL;
2840
2841         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2842         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2843                 return __cfqq;
2844
2845         return NULL;
2846 }
2847
2848 /*
2849  * cfqd - obvious
2850  * cur_cfqq - passed in so that we don't decide that the current queue is
2851  *            closely cooperating with itself.
2852  *
2853  * So, basically we're assuming that that cur_cfqq has dispatched at least
2854  * one request, and that cfqd->last_position reflects a position on the disk
2855  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2856  * assumption.
2857  */
2858 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2859                                               struct cfq_queue *cur_cfqq)
2860 {
2861         struct cfq_queue *cfqq;
2862
2863         if (cfq_class_idle(cur_cfqq))
2864                 return NULL;
2865         if (!cfq_cfqq_sync(cur_cfqq))
2866                 return NULL;
2867         if (CFQQ_SEEKY(cur_cfqq))
2868                 return NULL;
2869
2870         /*
2871          * Don't search priority tree if it's the only queue in the group.
2872          */
2873         if (cur_cfqq->cfqg->nr_cfqq == 1)
2874                 return NULL;
2875
2876         /*
2877          * We should notice if some of the queues are cooperating, eg
2878          * working closely on the same area of the disk. In that case,
2879          * we can group them together and don't waste time idling.
2880          */
2881         cfqq = cfqq_close(cfqd, cur_cfqq);
2882         if (!cfqq)
2883                 return NULL;
2884
2885         /* If new queue belongs to different cfq_group, don't choose it */
2886         if (cur_cfqq->cfqg != cfqq->cfqg)
2887                 return NULL;
2888
2889         /*
2890          * It only makes sense to merge sync queues.
2891          */
2892         if (!cfq_cfqq_sync(cfqq))
2893                 return NULL;
2894         if (CFQQ_SEEKY(cfqq))
2895                 return NULL;
2896
2897         /*
2898          * Do not merge queues of different priority classes
2899          */
2900         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2901                 return NULL;
2902
2903         return cfqq;
2904 }
2905
2906 /*
2907  * Determine whether we should enforce idle window for this queue.
2908  */
2909
2910 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2911 {
2912         enum wl_class_t wl_class = cfqq_class(cfqq);
2913         struct cfq_rb_root *st = cfqq->service_tree;
2914
2915         BUG_ON(!st);
2916         BUG_ON(!st->count);
2917
2918         if (!cfqd->cfq_slice_idle)
2919                 return false;
2920
2921         /* We never do for idle class queues. */
2922         if (wl_class == IDLE_WORKLOAD)
2923                 return false;
2924
2925         /* We do for queues that were marked with idle window flag. */
2926         if (cfq_cfqq_idle_window(cfqq) &&
2927            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2928                 return true;
2929
2930         /*
2931          * Otherwise, we do only if they are the last ones
2932          * in their service tree.
2933          */
2934         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2935            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2936                 return true;
2937         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2938         return false;
2939 }
2940
2941 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2942 {
2943         struct cfq_queue *cfqq = cfqd->active_queue;
2944         struct cfq_rb_root *st = cfqq->service_tree;
2945         struct cfq_io_cq *cic;
2946         u64 sl, group_idle = 0;
2947         u64 now = ktime_get_ns();
2948
2949         /*
2950          * SSD device without seek penalty, disable idling. But only do so
2951          * for devices that support queuing, otherwise we still have a problem
2952          * with sync vs async workloads.
2953          */
2954         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag &&
2955                 !cfqd->cfq_group_idle)
2956                 return;
2957
2958         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2959         WARN_ON(cfq_cfqq_slice_new(cfqq));
2960
2961         /*
2962          * idle is disabled, either manually or by past process history
2963          */
2964         if (!cfq_should_idle(cfqd, cfqq)) {
2965                 /* no queue idling. Check for group idling */
2966                 if (cfqd->cfq_group_idle)
2967                         group_idle = cfqd->cfq_group_idle;
2968                 else
2969                         return;
2970         }
2971
2972         /*
2973          * still active requests from this queue, don't idle
2974          */
2975         if (cfqq->dispatched)
2976                 return;
2977
2978         /*
2979          * task has exited, don't wait
2980          */
2981         cic = cfqd->active_cic;
2982         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2983                 return;
2984
2985         /*
2986          * If our average think time is larger than the remaining time
2987          * slice, then don't idle. This avoids overrunning the allotted
2988          * time slice.
2989          */
2990         if (sample_valid(cic->ttime.ttime_samples) &&
2991             (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2992                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2993                              cic->ttime.ttime_mean);
2994                 return;
2995         }
2996
2997         /*
2998          * There are other queues in the group or this is the only group and
2999          * it has too big thinktime, don't do group idle.
3000          */
3001         if (group_idle &&
3002             (cfqq->cfqg->nr_cfqq > 1 ||
3003              cfq_io_thinktime_big(cfqd, &st->ttime, true)))
3004                 return;
3005
3006         cfq_mark_cfqq_wait_request(cfqq);
3007
3008         if (group_idle)
3009                 sl = cfqd->cfq_group_idle;
3010         else
3011                 sl = cfqd->cfq_slice_idle;
3012
3013         hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
3014                       HRTIMER_MODE_REL);
3015         cfqg_stats_set_start_idle_time(cfqq->cfqg);
3016         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
3017                         group_idle ? 1 : 0);
3018 }
3019
3020 /*
3021  * Move request from internal lists to the request queue dispatch list.
3022  */
3023 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
3024 {
3025         struct cfq_data *cfqd = q->elevator->elevator_data;
3026         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3027
3028         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3029
3030         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
3031         cfq_remove_request(rq);
3032         cfqq->dispatched++;
3033         (RQ_CFQG(rq))->dispatched++;
3034         elv_dispatch_sort(q, rq);
3035
3036         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
3037         cfqq->nr_sectors += blk_rq_sectors(rq);
3038 }
3039
3040 /*
3041  * return expired entry, or NULL to just start from scratch in rbtree
3042  */
3043 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3044 {
3045         struct request *rq = NULL;
3046
3047         if (cfq_cfqq_fifo_expire(cfqq))
3048                 return NULL;
3049
3050         cfq_mark_cfqq_fifo_expire(cfqq);
3051
3052         if (list_empty(&cfqq->fifo))
3053                 return NULL;
3054
3055         rq = rq_entry_fifo(cfqq->fifo.next);
3056         if (ktime_get_ns() < rq->fifo_time)
3057                 rq = NULL;
3058
3059         return rq;
3060 }
3061
3062 static inline int
3063 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3064 {
3065         const int base_rq = cfqd->cfq_slice_async_rq;
3066
3067         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3068
3069         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3070 }
3071
3072 /*
3073  * Must be called with the queue_lock held.
3074  */
3075 static int cfqq_process_refs(struct cfq_queue *cfqq)
3076 {
3077         int process_refs, io_refs;
3078
3079         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3080         process_refs = cfqq->ref - io_refs;
3081         BUG_ON(process_refs < 0);
3082         return process_refs;
3083 }
3084
3085 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3086 {
3087         int process_refs, new_process_refs;
3088         struct cfq_queue *__cfqq;
3089
3090         /*
3091          * If there are no process references on the new_cfqq, then it is
3092          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3093          * chain may have dropped their last reference (not just their
3094          * last process reference).
3095          */
3096         if (!cfqq_process_refs(new_cfqq))
3097                 return;
3098
3099         /* Avoid a circular list and skip interim queue merges */
3100         while ((__cfqq = new_cfqq->new_cfqq)) {
3101                 if (__cfqq == cfqq)
3102                         return;
3103                 new_cfqq = __cfqq;
3104         }
3105
3106         process_refs = cfqq_process_refs(cfqq);
3107         new_process_refs = cfqq_process_refs(new_cfqq);
3108         /*
3109          * If the process for the cfqq has gone away, there is no
3110          * sense in merging the queues.
3111          */
3112         if (process_refs == 0 || new_process_refs == 0)
3113                 return;
3114
3115         /*
3116          * Merge in the direction of the lesser amount of work.
3117          */
3118         if (new_process_refs >= process_refs) {
3119                 cfqq->new_cfqq = new_cfqq;
3120                 new_cfqq->ref += process_refs;
3121         } else {
3122                 new_cfqq->new_cfqq = cfqq;
3123                 cfqq->ref += new_process_refs;
3124         }
3125 }
3126
3127 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3128                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3129 {
3130         struct cfq_queue *queue;
3131         int i;
3132         bool key_valid = false;
3133         u64 lowest_key = 0;
3134         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3135
3136         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3137                 /* select the one with lowest rb_key */
3138                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3139                 if (queue &&
3140                     (!key_valid || queue->rb_key < lowest_key)) {
3141                         lowest_key = queue->rb_key;
3142                         cur_best = i;
3143                         key_valid = true;
3144                 }
3145         }
3146
3147         return cur_best;
3148 }
3149
3150 static void
3151 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3152 {
3153         u64 slice;
3154         unsigned count;
3155         struct cfq_rb_root *st;
3156         u64 group_slice;
3157         enum wl_class_t original_class = cfqd->serving_wl_class;
3158         u64 now = ktime_get_ns();
3159
3160         /* Choose next priority. RT > BE > IDLE */
3161         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3162                 cfqd->serving_wl_class = RT_WORKLOAD;
3163         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3164                 cfqd->serving_wl_class = BE_WORKLOAD;
3165         else {
3166                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3167                 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3168                 return;
3169         }
3170
3171         if (original_class != cfqd->serving_wl_class)
3172                 goto new_workload;
3173
3174         /*
3175          * For RT and BE, we have to choose also the type
3176          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3177          * expiration time
3178          */
3179         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3180         count = st->count;
3181
3182         /*
3183          * check workload expiration, and that we still have other queues ready
3184          */
3185         if (count && !(now > cfqd->workload_expires))
3186                 return;
3187
3188 new_workload:
3189         /* otherwise select new workload type */
3190         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3191                                         cfqd->serving_wl_class);
3192         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3193         count = st->count;
3194
3195         /*
3196          * the workload slice is computed as a fraction of target latency
3197          * proportional to the number of queues in that workload, over
3198          * all the queues in the same priority class
3199          */
3200         group_slice = cfq_group_slice(cfqd, cfqg);
3201
3202         slice = div_u64(group_slice * count,
3203                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3204                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3205                                         cfqg)));
3206
3207         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3208                 u64 tmp;
3209
3210                 /*
3211                  * Async queues are currently system wide. Just taking
3212                  * proportion of queues with-in same group will lead to higher
3213                  * async ratio system wide as generally root group is going
3214                  * to have higher weight. A more accurate thing would be to
3215                  * calculate system wide asnc/sync ratio.
3216                  */
3217                 tmp = cfqd->cfq_target_latency *
3218                         cfqg_busy_async_queues(cfqd, cfqg);
3219                 tmp = div_u64(tmp, cfqd->busy_queues);
3220                 slice = min_t(u64, slice, tmp);
3221
3222                 /* async workload slice is scaled down according to
3223                  * the sync/async slice ratio. */
3224                 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3225         } else
3226                 /* sync workload slice is at least 2 * cfq_slice_idle */
3227                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3228
3229         slice = max_t(u64, slice, CFQ_MIN_TT);
3230         cfq_log(cfqd, "workload slice:%llu", slice);
3231         cfqd->workload_expires = now + slice;
3232 }
3233
3234 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3235 {
3236         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3237         struct cfq_group *cfqg;
3238
3239         if (RB_EMPTY_ROOT(&st->rb))
3240                 return NULL;
3241         cfqg = cfq_rb_first_group(st);
3242         update_min_vdisktime(st);
3243         return cfqg;
3244 }
3245
3246 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3247 {
3248         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3249         u64 now = ktime_get_ns();
3250
3251         cfqd->serving_group = cfqg;
3252
3253         /* Restore the workload type data */
3254         if (cfqg->saved_wl_slice) {
3255                 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3256                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3257                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3258         } else
3259                 cfqd->workload_expires = now - 1;
3260
3261         choose_wl_class_and_type(cfqd, cfqg);
3262 }
3263
3264 /*
3265  * Select a queue for service. If we have a current active queue,
3266  * check whether to continue servicing it, or retrieve and set a new one.
3267  */
3268 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3269 {
3270         struct cfq_queue *cfqq, *new_cfqq = NULL;
3271         u64 now = ktime_get_ns();
3272
3273         cfqq = cfqd->active_queue;
3274         if (!cfqq)
3275                 goto new_queue;
3276
3277         if (!cfqd->rq_queued)
3278                 return NULL;
3279
3280         /*
3281          * We were waiting for group to get backlogged. Expire the queue
3282          */
3283         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3284                 goto expire;
3285
3286         /*
3287          * The active queue has run out of time, expire it and select new.
3288          */
3289         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3290                 /*
3291                  * If slice had not expired at the completion of last request
3292                  * we might not have turned on wait_busy flag. Don't expire
3293                  * the queue yet. Allow the group to get backlogged.
3294                  *
3295                  * The very fact that we have used the slice, that means we
3296                  * have been idling all along on this queue and it should be
3297                  * ok to wait for this request to complete.
3298                  */
3299                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3300                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3301                         cfqq = NULL;
3302                         goto keep_queue;
3303                 } else
3304                         goto check_group_idle;
3305         }
3306
3307         /*
3308          * The active queue has requests and isn't expired, allow it to
3309          * dispatch.
3310          */
3311         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3312                 goto keep_queue;
3313
3314         /*
3315          * If another queue has a request waiting within our mean seek
3316          * distance, let it run.  The expire code will check for close
3317          * cooperators and put the close queue at the front of the service
3318          * tree.  If possible, merge the expiring queue with the new cfqq.
3319          */
3320         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3321         if (new_cfqq) {
3322                 if (!cfqq->new_cfqq)
3323                         cfq_setup_merge(cfqq, new_cfqq);
3324                 goto expire;
3325         }
3326
3327         /*
3328          * No requests pending. If the active queue still has requests in
3329          * flight or is idling for a new request, allow either of these
3330          * conditions to happen (or time out) before selecting a new queue.
3331          */
3332         if (hrtimer_active(&cfqd->idle_slice_timer)) {
3333                 cfqq = NULL;
3334                 goto keep_queue;
3335         }
3336
3337         /*
3338          * This is a deep seek queue, but the device is much faster than
3339          * the queue can deliver, don't idle
3340          **/
3341         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3342             (cfq_cfqq_slice_new(cfqq) ||
3343             (cfqq->slice_end - now > now - cfqq->slice_start))) {
3344                 cfq_clear_cfqq_deep(cfqq);
3345                 cfq_clear_cfqq_idle_window(cfqq);
3346         }
3347
3348         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3349                 cfqq = NULL;
3350                 goto keep_queue;
3351         }
3352
3353         /*
3354          * If group idle is enabled and there are requests dispatched from
3355          * this group, wait for requests to complete.
3356          */
3357 check_group_idle:
3358         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3359             cfqq->cfqg->dispatched &&
3360             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3361                 cfqq = NULL;
3362                 goto keep_queue;
3363         }
3364
3365 expire:
3366         cfq_slice_expired(cfqd, 0);
3367 new_queue:
3368         /*
3369          * Current queue expired. Check if we have to switch to a new
3370          * service tree
3371          */
3372         if (!new_cfqq)
3373                 cfq_choose_cfqg(cfqd);
3374
3375         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3376 keep_queue:
3377         return cfqq;
3378 }
3379
3380 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3381 {
3382         int dispatched = 0;
3383
3384         while (cfqq->next_rq) {
3385                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3386                 dispatched++;
3387         }
3388
3389         BUG_ON(!list_empty(&cfqq->fifo));
3390
3391         /* By default cfqq is not expired if it is empty. Do it explicitly */
3392         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3393         return dispatched;
3394 }
3395
3396 /*
3397  * Drain our current requests. Used for barriers and when switching
3398  * io schedulers on-the-fly.
3399  */
3400 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3401 {
3402         struct cfq_queue *cfqq;
3403         int dispatched = 0;
3404
3405         /* Expire the timeslice of the current active queue first */
3406         cfq_slice_expired(cfqd, 0);
3407         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3408                 __cfq_set_active_queue(cfqd, cfqq);
3409                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3410         }
3411
3412         BUG_ON(cfqd->busy_queues);
3413
3414         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3415         return dispatched;
3416 }
3417
3418 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3419         struct cfq_queue *cfqq)
3420 {
3421         u64 now = ktime_get_ns();
3422
3423         /* the queue hasn't finished any request, can't estimate */
3424         if (cfq_cfqq_slice_new(cfqq))
3425                 return true;
3426         if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end)
3427                 return true;
3428
3429         return false;
3430 }
3431
3432 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3433 {
3434         unsigned int max_dispatch;
3435
3436         if (cfq_cfqq_must_dispatch(cfqq))
3437                 return true;
3438
3439         /*
3440          * Drain async requests before we start sync IO
3441          */
3442         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3443                 return false;
3444
3445         /*
3446          * If this is an async queue and we have sync IO in flight, let it wait
3447          */
3448         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3449                 return false;
3450
3451         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3452         if (cfq_class_idle(cfqq))
3453                 max_dispatch = 1;
3454
3455         /*
3456          * Does this cfqq already have too much IO in flight?
3457          */
3458         if (cfqq->dispatched >= max_dispatch) {
3459                 bool promote_sync = false;
3460                 /*
3461                  * idle queue must always only have a single IO in flight
3462                  */
3463                 if (cfq_class_idle(cfqq))
3464                         return false;
3465
3466                 /*
3467                  * If there is only one sync queue
3468                  * we can ignore async queue here and give the sync
3469                  * queue no dispatch limit. The reason is a sync queue can
3470                  * preempt async queue, limiting the sync queue doesn't make
3471                  * sense. This is useful for aiostress test.
3472                  */
3473                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3474                         promote_sync = true;
3475
3476                 /*
3477                  * We have other queues, don't allow more IO from this one
3478                  */
3479                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3480                                 !promote_sync)
3481                         return false;
3482
3483                 /*
3484                  * Sole queue user, no limit
3485                  */
3486                 if (cfqd->busy_queues == 1 || promote_sync)
3487                         max_dispatch = -1;
3488                 else
3489                         /*
3490                          * Normally we start throttling cfqq when cfq_quantum/2
3491                          * requests have been dispatched. But we can drive
3492                          * deeper queue depths at the beginning of slice
3493                          * subjected to upper limit of cfq_quantum.
3494                          * */
3495                         max_dispatch = cfqd->cfq_quantum;
3496         }
3497
3498         /*
3499          * Async queues must wait a bit before being allowed dispatch.
3500          * We also ramp up the dispatch depth gradually for async IO,
3501          * based on the last sync IO we serviced
3502          */
3503         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3504                 u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync;
3505                 unsigned int depth;
3506
3507                 depth = div64_u64(last_sync, cfqd->cfq_slice[1]);
3508                 if (!depth && !cfqq->dispatched)
3509                         depth = 1;
3510                 if (depth < max_dispatch)
3511                         max_dispatch = depth;
3512         }
3513
3514         /*
3515          * If we're below the current max, allow a dispatch
3516          */
3517         return cfqq->dispatched < max_dispatch;
3518 }
3519
3520 /*
3521  * Dispatch a request from cfqq, moving them to the request queue
3522  * dispatch list.
3523  */
3524 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3525 {
3526         struct request *rq;
3527
3528         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3529
3530         rq = cfq_check_fifo(cfqq);
3531         if (rq)
3532                 cfq_mark_cfqq_must_dispatch(cfqq);
3533
3534         if (!cfq_may_dispatch(cfqd, cfqq))
3535                 return false;
3536
3537         /*
3538          * follow expired path, else get first next available
3539          */
3540         if (!rq)
3541                 rq = cfqq->next_rq;
3542         else
3543                 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
3544
3545         /*
3546          * insert request into driver dispatch list
3547          */
3548         cfq_dispatch_insert(cfqd->queue, rq);
3549
3550         if (!cfqd->active_cic) {
3551                 struct cfq_io_cq *cic = RQ_CIC(rq);
3552
3553                 atomic_long_inc(&cic->icq.ioc->refcount);
3554                 cfqd->active_cic = cic;
3555         }
3556
3557         return true;
3558 }
3559
3560 /*
3561  * Find the cfqq that we need to service and move a request from that to the
3562  * dispatch list
3563  */
3564 static int cfq_dispatch_requests(struct request_queue *q, int force)
3565 {
3566         struct cfq_data *cfqd = q->elevator->elevator_data;
3567         struct cfq_queue *cfqq;
3568
3569         if (!cfqd->busy_queues)
3570                 return 0;
3571
3572         if (unlikely(force))
3573                 return cfq_forced_dispatch(cfqd);
3574
3575         cfqq = cfq_select_queue(cfqd);
3576         if (!cfqq)
3577                 return 0;
3578
3579         /*
3580          * Dispatch a request from this cfqq, if it is allowed
3581          */
3582         if (!cfq_dispatch_request(cfqd, cfqq))
3583                 return 0;
3584
3585         cfqq->slice_dispatch++;
3586         cfq_clear_cfqq_must_dispatch(cfqq);
3587
3588         /*
3589          * expire an async queue immediately if it has used up its slice. idle
3590          * queue always expire after 1 dispatch round.
3591          */
3592         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3593             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3594             cfq_class_idle(cfqq))) {
3595                 cfqq->slice_end = ktime_get_ns() + 1;
3596                 cfq_slice_expired(cfqd, 0);
3597         }
3598
3599         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3600         return 1;
3601 }
3602
3603 /*
3604  * task holds one reference to the queue, dropped when task exits. each rq
3605  * in-flight on this queue also holds a reference, dropped when rq is freed.
3606  *
3607  * Each cfq queue took a reference on the parent group. Drop it now.
3608  * queue lock must be held here.
3609  */
3610 static void cfq_put_queue(struct cfq_queue *cfqq)
3611 {
3612         struct cfq_data *cfqd = cfqq->cfqd;
3613         struct cfq_group *cfqg;
3614
3615         BUG_ON(cfqq->ref <= 0);
3616
3617         cfqq->ref--;
3618         if (cfqq->ref)
3619                 return;
3620
3621         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3622         BUG_ON(rb_first(&cfqq->sort_list));
3623         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3624         cfqg = cfqq->cfqg;
3625
3626         if (unlikely(cfqd->active_queue == cfqq)) {
3627                 __cfq_slice_expired(cfqd, cfqq, 0);
3628                 cfq_schedule_dispatch(cfqd);
3629         }
3630
3631         BUG_ON(cfq_cfqq_on_rr(cfqq));
3632         kmem_cache_free(cfq_pool, cfqq);
3633         cfqg_put(cfqg);
3634 }
3635
3636 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3637 {
3638         struct cfq_queue *__cfqq, *next;
3639
3640         /*
3641          * If this queue was scheduled to merge with another queue, be
3642          * sure to drop the reference taken on that queue (and others in
3643          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3644          */
3645         __cfqq = cfqq->new_cfqq;
3646         while (__cfqq) {
3647                 if (__cfqq == cfqq) {
3648                         WARN(1, "cfqq->new_cfqq loop detected\n");
3649                         break;
3650                 }
3651                 next = __cfqq->new_cfqq;
3652                 cfq_put_queue(__cfqq);
3653                 __cfqq = next;
3654         }
3655 }
3656
3657 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3658 {
3659         if (unlikely(cfqq == cfqd->active_queue)) {
3660                 __cfq_slice_expired(cfqd, cfqq, 0);
3661                 cfq_schedule_dispatch(cfqd);
3662         }
3663
3664         cfq_put_cooperator(cfqq);
3665
3666         cfq_put_queue(cfqq);
3667 }
3668
3669 static void cfq_init_icq(struct io_cq *icq)
3670 {
3671         struct cfq_io_cq *cic = icq_to_cic(icq);
3672
3673         cic->ttime.last_end_request = ktime_get_ns();
3674 }
3675
3676 static void cfq_exit_icq(struct io_cq *icq)
3677 {
3678         struct cfq_io_cq *cic = icq_to_cic(icq);
3679         struct cfq_data *cfqd = cic_to_cfqd(cic);
3680
3681         if (cic_to_cfqq(cic, false)) {
3682                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3683                 cic_set_cfqq(cic, NULL, false);
3684         }
3685
3686         if (cic_to_cfqq(cic, true)) {
3687                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3688                 cic_set_cfqq(cic, NULL, true);
3689         }
3690 }
3691
3692 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3693 {
3694         struct task_struct *tsk = current;
3695         int ioprio_class;
3696
3697         if (!cfq_cfqq_prio_changed(cfqq))
3698                 return;
3699
3700         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3701         switch (ioprio_class) {
3702         default:
3703                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3704         case IOPRIO_CLASS_NONE:
3705                 /*
3706                  * no prio set, inherit CPU scheduling settings
3707                  */
3708                 cfqq->ioprio = task_nice_ioprio(tsk);
3709                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3710                 break;
3711         case IOPRIO_CLASS_RT:
3712                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3713                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3714                 break;
3715         case IOPRIO_CLASS_BE:
3716                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3717                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3718                 break;
3719         case IOPRIO_CLASS_IDLE:
3720                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3721                 cfqq->ioprio = 7;
3722                 cfq_clear_cfqq_idle_window(cfqq);
3723                 break;
3724         }
3725
3726         /*
3727          * keep track of original prio settings in case we have to temporarily
3728          * elevate the priority of this queue
3729          */
3730         cfqq->org_ioprio = cfqq->ioprio;
3731         cfqq->org_ioprio_class = cfqq->ioprio_class;
3732         cfq_clear_cfqq_prio_changed(cfqq);
3733 }
3734
3735 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3736 {
3737         int ioprio = cic->icq.ioc->ioprio;
3738         struct cfq_data *cfqd = cic_to_cfqd(cic);
3739         struct cfq_queue *cfqq;
3740
3741         /*
3742          * Check whether ioprio has changed.  The condition may trigger
3743          * spuriously on a newly created cic but there's no harm.
3744          */
3745         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3746                 return;
3747
3748         cfqq = cic_to_cfqq(cic, false);
3749         if (cfqq) {
3750                 cfq_put_queue(cfqq);
3751                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3752                 cic_set_cfqq(cic, cfqq, false);
3753         }
3754
3755         cfqq = cic_to_cfqq(cic, true);
3756         if (cfqq)
3757                 cfq_mark_cfqq_prio_changed(cfqq);
3758
3759         cic->ioprio = ioprio;
3760 }
3761
3762 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3763                           pid_t pid, bool is_sync)
3764 {
3765         RB_CLEAR_NODE(&cfqq->rb_node);
3766         RB_CLEAR_NODE(&cfqq->p_node);
3767         INIT_LIST_HEAD(&cfqq->fifo);
3768
3769         cfqq->ref = 0;
3770         cfqq->cfqd = cfqd;
3771
3772         cfq_mark_cfqq_prio_changed(cfqq);
3773
3774         if (is_sync) {
3775                 if (!cfq_class_idle(cfqq))
3776                         cfq_mark_cfqq_idle_window(cfqq);
3777                 cfq_mark_cfqq_sync(cfqq);
3778         }
3779         cfqq->pid = pid;
3780 }
3781
3782 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3783 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3784 {
3785         struct cfq_data *cfqd = cic_to_cfqd(cic);
3786         struct cfq_queue *cfqq;
3787         uint64_t serial_nr;
3788
3789         rcu_read_lock();
3790         serial_nr = bio_blkcg(bio)->css.serial_nr;
3791         rcu_read_unlock();
3792
3793         /*
3794          * Check whether blkcg has changed.  The condition may trigger
3795          * spuriously on a newly created cic but there's no harm.
3796          */
3797         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3798                 return;
3799
3800         /*
3801          * Drop reference to queues.  New queues will be assigned in new
3802          * group upon arrival of fresh requests.
3803          */
3804         cfqq = cic_to_cfqq(cic, false);
3805         if (cfqq) {
3806                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3807                 cic_set_cfqq(cic, NULL, false);
3808                 cfq_put_queue(cfqq);
3809         }
3810
3811         cfqq = cic_to_cfqq(cic, true);
3812         if (cfqq) {
3813                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3814                 cic_set_cfqq(cic, NULL, true);
3815                 cfq_put_queue(cfqq);
3816         }
3817
3818         cic->blkcg_serial_nr = serial_nr;
3819 }
3820 #else
3821 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3822 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3823
3824 static struct cfq_queue **
3825 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3826 {
3827         switch (ioprio_class) {
3828         case IOPRIO_CLASS_RT:
3829                 return &cfqg->async_cfqq[0][ioprio];
3830         case IOPRIO_CLASS_NONE:
3831                 ioprio = IOPRIO_NORM;
3832                 /* fall through */
3833         case IOPRIO_CLASS_BE:
3834                 return &cfqg->async_cfqq[1][ioprio];
3835         case IOPRIO_CLASS_IDLE:
3836                 return &cfqg->async_idle_cfqq;
3837         default:
3838                 BUG();
3839         }
3840 }
3841
3842 static struct cfq_queue *
3843 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3844               struct bio *bio)
3845 {
3846         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3847         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3848         struct cfq_queue **async_cfqq = NULL;
3849         struct cfq_queue *cfqq;
3850         struct cfq_group *cfqg;
3851
3852         rcu_read_lock();
3853         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3854         if (!cfqg) {
3855                 cfqq = &cfqd->oom_cfqq;
3856                 goto out;
3857         }
3858
3859         if (!is_sync) {
3860                 if (!ioprio_valid(cic->ioprio)) {
3861                         struct task_struct *tsk = current;
3862                         ioprio = task_nice_ioprio(tsk);
3863                         ioprio_class = task_nice_ioclass(tsk);
3864                 }
3865                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3866                 cfqq = *async_cfqq;
3867                 if (cfqq)
3868                         goto out;
3869         }
3870
3871         cfqq = kmem_cache_alloc_node(cfq_pool,
3872                                      GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
3873                                      cfqd->queue->node);
3874         if (!cfqq) {
3875                 cfqq = &cfqd->oom_cfqq;
3876                 goto out;
3877         }
3878
3879         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3880         cfq_init_prio_data(cfqq, cic);
3881         cfq_link_cfqq_cfqg(cfqq, cfqg);
3882         cfq_log_cfqq(cfqd, cfqq, "alloced");
3883
3884         if (async_cfqq) {
3885                 /* a new async queue is created, pin and remember */
3886                 cfqq->ref++;
3887                 *async_cfqq = cfqq;
3888         }
3889 out:
3890         cfqq->ref++;
3891         rcu_read_unlock();
3892         return cfqq;
3893 }
3894
3895 static void
3896 __cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle)
3897 {
3898         u64 elapsed = ktime_get_ns() - ttime->last_end_request;
3899         elapsed = min(elapsed, 2UL * slice_idle);
3900
3901         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3902         ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
3903         ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
3904                                      ttime->ttime_samples);
3905 }
3906
3907 static void
3908 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3909                         struct cfq_io_cq *cic)
3910 {
3911         if (cfq_cfqq_sync(cfqq)) {
3912                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3913                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3914                         cfqd->cfq_slice_idle);
3915         }
3916 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3917         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3918 #endif
3919 }
3920
3921 static void
3922 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3923                        struct request *rq)
3924 {
3925         sector_t sdist = 0;
3926         sector_t n_sec = blk_rq_sectors(rq);
3927         if (cfqq->last_request_pos) {
3928                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3929                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3930                 else
3931                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3932         }
3933
3934         cfqq->seek_history <<= 1;
3935         if (blk_queue_nonrot(cfqd->queue))
3936                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3937         else
3938                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3939 }
3940
3941 /*
3942  * Disable idle window if the process thinks too long or seeks so much that
3943  * it doesn't matter
3944  */
3945 static void
3946 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3947                        struct cfq_io_cq *cic)
3948 {
3949         int old_idle, enable_idle;
3950
3951         /*
3952          * Don't idle for async or idle io prio class
3953          */
3954         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3955                 return;
3956
3957         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3958
3959         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3960                 cfq_mark_cfqq_deep(cfqq);
3961
3962         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3963                 enable_idle = 0;
3964         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3965                  !cfqd->cfq_slice_idle ||
3966                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3967                 enable_idle = 0;
3968         else if (sample_valid(cic->ttime.ttime_samples)) {
3969                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3970                         enable_idle = 0;
3971                 else
3972                         enable_idle = 1;
3973         }
3974
3975         if (old_idle != enable_idle) {
3976                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3977                 if (enable_idle)
3978                         cfq_mark_cfqq_idle_window(cfqq);
3979                 else
3980                         cfq_clear_cfqq_idle_window(cfqq);
3981         }
3982 }
3983
3984 /*
3985  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3986  * no or if we aren't sure, a 1 will cause a preempt.
3987  */
3988 static bool
3989 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3990                    struct request *rq)
3991 {
3992         struct cfq_queue *cfqq;
3993
3994         cfqq = cfqd->active_queue;
3995         if (!cfqq)
3996                 return false;
3997
3998         if (cfq_class_idle(new_cfqq))
3999                 return false;
4000
4001         if (cfq_class_idle(cfqq))
4002                 return true;
4003
4004         /*
4005          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
4006          */
4007         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
4008                 return false;
4009
4010         /*
4011          * if the new request is sync, but the currently running queue is
4012          * not, let the sync request have priority.
4013          */
4014         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
4015                 return true;
4016
4017         /*
4018          * Treat ancestors of current cgroup the same way as current cgroup.
4019          * For anybody else we disallow preemption to guarantee service
4020          * fairness among cgroups.
4021          */
4022         if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg))
4023                 return false;
4024
4025         if (cfq_slice_used(cfqq))
4026                 return true;
4027
4028         /*
4029          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
4030          */
4031         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
4032                 return true;
4033
4034         WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class);
4035         /* Allow preemption only if we are idling on sync-noidle tree */
4036         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
4037             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
4038             RB_EMPTY_ROOT(&cfqq->sort_list))
4039                 return true;
4040
4041         /*
4042          * So both queues are sync. Let the new request get disk time if
4043          * it's a metadata request and the current queue is doing regular IO.
4044          */
4045         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
4046                 return true;
4047
4048         /* An idle queue should not be idle now for some reason */
4049         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
4050                 return true;
4051
4052         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
4053                 return false;
4054
4055         /*
4056          * if this request is as-good as one we would expect from the
4057          * current cfqq, let it preempt
4058          */
4059         if (cfq_rq_close(cfqd, cfqq, rq))
4060                 return true;
4061
4062         return false;
4063 }
4064
4065 /*
4066  * cfqq preempts the active queue. if we allowed preempt with no slice left,
4067  * let it have half of its nominal slice.
4068  */
4069 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4070 {
4071         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4072
4073         cfq_log_cfqq(cfqd, cfqq, "preempt");
4074         cfq_slice_expired(cfqd, 1);
4075
4076         /*
4077          * workload type is changed, don't save slice, otherwise preempt
4078          * doesn't happen
4079          */
4080         if (old_type != cfqq_type(cfqq))
4081                 cfqq->cfqg->saved_wl_slice = 0;
4082
4083         /*
4084          * Put the new queue at the front of the of the current list,
4085          * so we know that it will be selected next.
4086          */
4087         BUG_ON(!cfq_cfqq_on_rr(cfqq));
4088
4089         cfq_service_tree_add(cfqd, cfqq, 1);
4090
4091         cfqq->slice_end = 0;
4092         cfq_mark_cfqq_slice_new(cfqq);
4093 }
4094
4095 /*
4096  * Called when a new fs request (rq) is added (to cfqq). Check if there's
4097  * something we should do about it
4098  */
4099 static void
4100 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4101                 struct request *rq)
4102 {
4103         struct cfq_io_cq *cic = RQ_CIC(rq);
4104
4105         cfqd->rq_queued++;
4106         if (rq->cmd_flags & REQ_PRIO)
4107                 cfqq->prio_pending++;
4108
4109         cfq_update_io_thinktime(cfqd, cfqq, cic);
4110         cfq_update_io_seektime(cfqd, cfqq, rq);
4111         cfq_update_idle_window(cfqd, cfqq, cic);
4112
4113         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4114
4115         if (cfqq == cfqd->active_queue) {
4116                 /*
4117                  * Remember that we saw a request from this process, but
4118                  * don't start queuing just yet. Otherwise we risk seeing lots
4119                  * of tiny requests, because we disrupt the normal plugging
4120                  * and merging. If the request is already larger than a single
4121                  * page, let it rip immediately. For that case we assume that
4122                  * merging is already done. Ditto for a busy system that
4123                  * has other work pending, don't risk delaying until the
4124                  * idle timer unplug to continue working.
4125                  */
4126                 if (cfq_cfqq_wait_request(cfqq)) {
4127                         if (blk_rq_bytes(rq) > PAGE_SIZE ||
4128                             cfqd->busy_queues > 1) {
4129                                 cfq_del_timer(cfqd, cfqq);
4130                                 cfq_clear_cfqq_wait_request(cfqq);
4131                                 __blk_run_queue(cfqd->queue);
4132                         } else {
4133                                 cfqg_stats_update_idle_time(cfqq->cfqg);
4134                                 cfq_mark_cfqq_must_dispatch(cfqq);
4135                         }
4136                 }
4137         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
4138                 /*
4139                  * not the active queue - expire current slice if it is
4140                  * idle and has expired it's mean thinktime or this new queue
4141                  * has some old slice time left and is of higher priority or
4142                  * this new queue is RT and the current one is BE
4143                  */
4144                 cfq_preempt_queue(cfqd, cfqq);
4145                 __blk_run_queue(cfqd->queue);
4146         }
4147 }
4148
4149 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4150 {
4151         struct cfq_data *cfqd = q->elevator->elevator_data;
4152         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4153
4154         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4155         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4156
4157         rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4158         list_add_tail(&rq->queuelist, &cfqq->fifo);
4159         cfq_add_rq_rb(rq);
4160         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group, req_op(rq),
4161                                  rq->cmd_flags);
4162         cfq_rq_enqueued(cfqd, cfqq, rq);
4163 }
4164
4165 /*
4166  * Update hw_tag based on peak queue depth over 50 samples under
4167  * sufficient load.
4168  */
4169 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4170 {
4171         struct cfq_queue *cfqq = cfqd->active_queue;
4172
4173         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4174                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4175
4176         if (cfqd->hw_tag == 1)
4177                 return;
4178
4179         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4180             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4181                 return;
4182
4183         /*
4184          * If active queue hasn't enough requests and can idle, cfq might not
4185          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4186          * case
4187          */
4188         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4189             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4190             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4191                 return;
4192
4193         if (cfqd->hw_tag_samples++ < 50)
4194                 return;
4195
4196         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4197                 cfqd->hw_tag = 1;
4198         else
4199                 cfqd->hw_tag = 0;
4200 }
4201
4202 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4203 {
4204         struct cfq_io_cq *cic = cfqd->active_cic;
4205         u64 now = ktime_get_ns();
4206
4207         /* If the queue already has requests, don't wait */
4208         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4209                 return false;
4210
4211         /* If there are other queues in the group, don't wait */
4212         if (cfqq->cfqg->nr_cfqq > 1)
4213                 return false;
4214
4215         /* the only queue in the group, but think time is big */
4216         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4217                 return false;
4218
4219         if (cfq_slice_used(cfqq))
4220                 return true;
4221
4222         /* if slice left is less than think time, wait busy */
4223         if (cic && sample_valid(cic->ttime.ttime_samples)
4224             && (cfqq->slice_end - now < cic->ttime.ttime_mean))
4225                 return true;
4226
4227         /*
4228          * If think times is less than a jiffy than ttime_mean=0 and above
4229          * will not be true. It might happen that slice has not expired yet
4230          * but will expire soon (4-5 ns) during select_queue(). To cover the
4231          * case where think time is less than a jiffy, mark the queue wait
4232          * busy if only 1 jiffy is left in the slice.
4233          */
4234         if (cfqq->slice_end - now <= jiffies_to_nsecs(1))
4235                 return true;
4236
4237         return false;
4238 }
4239
4240 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4241 {
4242         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4243         struct cfq_data *cfqd = cfqq->cfqd;
4244         const int sync = rq_is_sync(rq);
4245         u64 now = ktime_get_ns();
4246
4247         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4248                      !!(rq->cmd_flags & REQ_NOIDLE));
4249
4250         cfq_update_hw_tag(cfqd);
4251
4252         WARN_ON(!cfqd->rq_in_driver);
4253         WARN_ON(!cfqq->dispatched);
4254         cfqd->rq_in_driver--;
4255         cfqq->dispatched--;
4256         (RQ_CFQG(rq))->dispatched--;
4257         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4258                                      rq_io_start_time_ns(rq), req_op(rq),
4259                                      rq->cmd_flags);
4260
4261         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4262
4263         if (sync) {
4264                 struct cfq_rb_root *st;
4265
4266                 RQ_CIC(rq)->ttime.last_end_request = now;
4267
4268                 if (cfq_cfqq_on_rr(cfqq))
4269                         st = cfqq->service_tree;
4270                 else
4271                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4272                                         cfqq_type(cfqq));
4273
4274                 st->ttime.last_end_request = now;
4275                 /*
4276                  * We have to do this check in jiffies since start_time is in
4277                  * jiffies and it is not trivial to convert to ns. If
4278                  * cfq_fifo_expire[1] ever comes close to 1 jiffie, this test
4279                  * will become problematic but so far we are fine (the default
4280                  * is 128 ms).
4281                  */
4282                 if (!time_after(rq->start_time +
4283                                   nsecs_to_jiffies(cfqd->cfq_fifo_expire[1]),
4284                                 jiffies))
4285                         cfqd->last_delayed_sync = now;
4286         }
4287
4288 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4289         cfqq->cfqg->ttime.last_end_request = now;
4290 #endif
4291
4292         /*
4293          * If this is the active queue, check if it needs to be expired,
4294          * or if we want to idle in case it has no pending requests.
4295          */
4296         if (cfqd->active_queue == cfqq) {
4297                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4298
4299                 if (cfq_cfqq_slice_new(cfqq)) {
4300                         cfq_set_prio_slice(cfqd, cfqq);
4301                         cfq_clear_cfqq_slice_new(cfqq);
4302                 }
4303
4304                 /*
4305                  * Should we wait for next request to come in before we expire
4306                  * the queue.
4307                  */
4308                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4309                         u64 extend_sl = cfqd->cfq_slice_idle;
4310                         if (!cfqd->cfq_slice_idle)
4311                                 extend_sl = cfqd->cfq_group_idle;
4312                         cfqq->slice_end = now + extend_sl;
4313                         cfq_mark_cfqq_wait_busy(cfqq);
4314                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4315                 }
4316
4317                 /*
4318                  * Idling is not enabled on:
4319                  * - expired queues
4320                  * - idle-priority queues
4321                  * - async queues
4322                  * - queues with still some requests queued
4323                  * - when there is a close cooperator
4324                  */
4325                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4326                         cfq_slice_expired(cfqd, 1);
4327                 else if (sync && cfqq_empty &&
4328                          !cfq_close_cooperator(cfqd, cfqq)) {
4329                         cfq_arm_slice_timer(cfqd);
4330                 }
4331         }
4332
4333         if (!cfqd->rq_in_driver)
4334                 cfq_schedule_dispatch(cfqd);
4335 }
4336
4337 static void cfqq_boost_on_prio(struct cfq_queue *cfqq, int op_flags)
4338 {
4339         /*
4340          * If REQ_PRIO is set, boost class and prio level, if it's below
4341          * BE/NORM. If prio is not set, restore the potentially boosted
4342          * class/prio level.
4343          */
4344         if (!(op_flags & REQ_PRIO)) {
4345                 cfqq->ioprio_class = cfqq->org_ioprio_class;
4346                 cfqq->ioprio = cfqq->org_ioprio;
4347         } else {
4348                 if (cfq_class_idle(cfqq))
4349                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
4350                 if (cfqq->ioprio > IOPRIO_NORM)
4351                         cfqq->ioprio = IOPRIO_NORM;
4352         }
4353 }
4354
4355 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4356 {
4357         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4358                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4359                 return ELV_MQUEUE_MUST;
4360         }
4361
4362         return ELV_MQUEUE_MAY;
4363 }
4364
4365 static int cfq_may_queue(struct request_queue *q, int op, int op_flags)
4366 {
4367         struct cfq_data *cfqd = q->elevator->elevator_data;
4368         struct task_struct *tsk = current;
4369         struct cfq_io_cq *cic;
4370         struct cfq_queue *cfqq;
4371
4372         /*
4373          * don't force setup of a queue from here, as a call to may_queue
4374          * does not necessarily imply that a request actually will be queued.
4375          * so just lookup a possibly existing queue, or return 'may queue'
4376          * if that fails
4377          */
4378         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4379         if (!cic)
4380                 return ELV_MQUEUE_MAY;
4381
4382         cfqq = cic_to_cfqq(cic, rw_is_sync(op, op_flags));
4383         if (cfqq) {
4384                 cfq_init_prio_data(cfqq, cic);
4385                 cfqq_boost_on_prio(cfqq, op_flags);
4386
4387                 return __cfq_may_queue(cfqq);
4388         }
4389
4390         return ELV_MQUEUE_MAY;
4391 }
4392
4393 /*
4394  * queue lock held here
4395  */
4396 static void cfq_put_request(struct request *rq)
4397 {
4398         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4399
4400         if (cfqq) {
4401                 const int rw = rq_data_dir(rq);
4402
4403                 BUG_ON(!cfqq->allocated[rw]);
4404                 cfqq->allocated[rw]--;
4405
4406                 /* Put down rq reference on cfqg */
4407                 cfqg_put(RQ_CFQG(rq));
4408                 rq->elv.priv[0] = NULL;
4409                 rq->elv.priv[1] = NULL;
4410
4411                 cfq_put_queue(cfqq);
4412         }
4413 }
4414
4415 static struct cfq_queue *
4416 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4417                 struct cfq_queue *cfqq)
4418 {
4419         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4420         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4421         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4422         cfq_put_queue(cfqq);
4423         return cic_to_cfqq(cic, 1);
4424 }
4425
4426 /*
4427  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4428  * was the last process referring to said cfqq.
4429  */
4430 static struct cfq_queue *
4431 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4432 {
4433         if (cfqq_process_refs(cfqq) == 1) {
4434                 cfqq->pid = current->pid;
4435                 cfq_clear_cfqq_coop(cfqq);
4436                 cfq_clear_cfqq_split_coop(cfqq);
4437                 return cfqq;
4438         }
4439
4440         cic_set_cfqq(cic, NULL, 1);
4441
4442         cfq_put_cooperator(cfqq);
4443
4444         cfq_put_queue(cfqq);
4445         return NULL;
4446 }
4447 /*
4448  * Allocate cfq data structures associated with this request.
4449  */
4450 static int
4451 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4452                 gfp_t gfp_mask)
4453 {
4454         struct cfq_data *cfqd = q->elevator->elevator_data;
4455         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4456         const int rw = rq_data_dir(rq);
4457         const bool is_sync = rq_is_sync(rq);
4458         struct cfq_queue *cfqq;
4459
4460         spin_lock_irq(q->queue_lock);
4461
4462         check_ioprio_changed(cic, bio);
4463         check_blkcg_changed(cic, bio);
4464 new_queue:
4465         cfqq = cic_to_cfqq(cic, is_sync);
4466         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4467                 if (cfqq)
4468                         cfq_put_queue(cfqq);
4469                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4470                 cic_set_cfqq(cic, cfqq, is_sync);
4471         } else {
4472                 /*
4473                  * If the queue was seeky for too long, break it apart.
4474                  */
4475                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4476                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4477                         cfqq = split_cfqq(cic, cfqq);
4478                         if (!cfqq)
4479                                 goto new_queue;
4480                 }
4481
4482                 /*
4483                  * Check to see if this queue is scheduled to merge with
4484                  * another, closely cooperating queue.  The merging of
4485                  * queues happens here as it must be done in process context.
4486                  * The reference on new_cfqq was taken in merge_cfqqs.
4487                  */
4488                 if (cfqq->new_cfqq)
4489                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4490         }
4491
4492         cfqq->allocated[rw]++;
4493
4494         cfqq->ref++;
4495         cfqg_get(cfqq->cfqg);
4496         rq->elv.priv[0] = cfqq;
4497         rq->elv.priv[1] = cfqq->cfqg;
4498         spin_unlock_irq(q->queue_lock);
4499         return 0;
4500 }
4501
4502 static void cfq_kick_queue(struct work_struct *work)
4503 {
4504         struct cfq_data *cfqd =
4505                 container_of(work, struct cfq_data, unplug_work);
4506         struct request_queue *q = cfqd->queue;
4507
4508         spin_lock_irq(q->queue_lock);
4509         __blk_run_queue(cfqd->queue);
4510         spin_unlock_irq(q->queue_lock);
4511 }
4512
4513 /*
4514  * Timer running if the active_queue is currently idling inside its time slice
4515  */
4516 static enum hrtimer_restart cfq_idle_slice_timer(struct hrtimer *timer)
4517 {
4518         struct cfq_data *cfqd = container_of(timer, struct cfq_data,
4519                                              idle_slice_timer);
4520         struct cfq_queue *cfqq;
4521         unsigned long flags;
4522         int timed_out = 1;
4523
4524         cfq_log(cfqd, "idle timer fired");
4525
4526         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4527
4528         cfqq = cfqd->active_queue;
4529         if (cfqq) {
4530                 timed_out = 0;
4531
4532                 /*
4533                  * We saw a request before the queue expired, let it through
4534                  */
4535                 if (cfq_cfqq_must_dispatch(cfqq))
4536                         goto out_kick;
4537
4538                 /*
4539                  * expired
4540                  */
4541                 if (cfq_slice_used(cfqq))
4542                         goto expire;
4543
4544                 /*
4545                  * only expire and reinvoke request handler, if there are
4546                  * other queues with pending requests
4547                  */
4548                 if (!cfqd->busy_queues)
4549                         goto out_cont;
4550
4551                 /*
4552                  * not expired and it has a request pending, let it dispatch
4553                  */
4554                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4555                         goto out_kick;
4556
4557                 /*
4558                  * Queue depth flag is reset only when the idle didn't succeed
4559                  */
4560                 cfq_clear_cfqq_deep(cfqq);
4561         }
4562 expire:
4563         cfq_slice_expired(cfqd, timed_out);
4564 out_kick:
4565         cfq_schedule_dispatch(cfqd);
4566 out_cont:
4567         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4568         return HRTIMER_NORESTART;
4569 }
4570
4571 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4572 {
4573         hrtimer_cancel(&cfqd->idle_slice_timer);
4574         cancel_work_sync(&cfqd->unplug_work);
4575 }
4576
4577 static void cfq_exit_queue(struct elevator_queue *e)
4578 {
4579         struct cfq_data *cfqd = e->elevator_data;
4580         struct request_queue *q = cfqd->queue;
4581
4582         cfq_shutdown_timer_wq(cfqd);
4583
4584         spin_lock_irq(q->queue_lock);
4585
4586         if (cfqd->active_queue)
4587                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4588
4589         spin_unlock_irq(q->queue_lock);
4590
4591         cfq_shutdown_timer_wq(cfqd);
4592
4593 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4594         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4595 #else
4596         kfree(cfqd->root_group);
4597 #endif
4598         kfree(cfqd);
4599 }
4600
4601 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4602 {
4603         struct cfq_data *cfqd;
4604         struct blkcg_gq *blkg __maybe_unused;
4605         int i, ret;
4606         struct elevator_queue *eq;
4607
4608         eq = elevator_alloc(q, e);
4609         if (!eq)
4610                 return -ENOMEM;
4611
4612         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4613         if (!cfqd) {
4614                 kobject_put(&eq->kobj);
4615                 return -ENOMEM;
4616         }
4617         eq->elevator_data = cfqd;
4618
4619         cfqd->queue = q;
4620         spin_lock_irq(q->queue_lock);
4621         q->elevator = eq;
4622         spin_unlock_irq(q->queue_lock);
4623
4624         /* Init root service tree */
4625         cfqd->grp_service_tree = CFQ_RB_ROOT;
4626
4627         /* Init root group and prefer root group over other groups by default */
4628 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4629         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4630         if (ret)
4631                 goto out_free;
4632
4633         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4634 #else
4635         ret = -ENOMEM;
4636         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4637                                         GFP_KERNEL, cfqd->queue->node);
4638         if (!cfqd->root_group)
4639                 goto out_free;
4640
4641         cfq_init_cfqg_base(cfqd->root_group);
4642         cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4643         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4644 #endif
4645
4646         /*
4647          * Not strictly needed (since RB_ROOT just clears the node and we
4648          * zeroed cfqd on alloc), but better be safe in case someone decides
4649          * to add magic to the rb code
4650          */
4651         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4652                 cfqd->prio_trees[i] = RB_ROOT;
4653
4654         /*
4655          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4656          * Grab a permanent reference to it, so that the normal code flow
4657          * will not attempt to free it.  oom_cfqq is linked to root_group
4658          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4659          * the reference from linking right away.
4660          */
4661         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4662         cfqd->oom_cfqq.ref++;
4663
4664         spin_lock_irq(q->queue_lock);
4665         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4666         cfqg_put(cfqd->root_group);
4667         spin_unlock_irq(q->queue_lock);
4668
4669         hrtimer_init(&cfqd->idle_slice_timer, CLOCK_MONOTONIC,
4670                      HRTIMER_MODE_REL);
4671         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4672
4673         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4674
4675         cfqd->cfq_quantum = cfq_quantum;
4676         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4677         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4678         cfqd->cfq_back_max = cfq_back_max;
4679         cfqd->cfq_back_penalty = cfq_back_penalty;
4680         cfqd->cfq_slice[0] = cfq_slice_async;
4681         cfqd->cfq_slice[1] = cfq_slice_sync;
4682         cfqd->cfq_target_latency = cfq_target_latency;
4683         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4684         cfqd->cfq_slice_idle = cfq_slice_idle;
4685         cfqd->cfq_group_idle = cfq_group_idle;
4686         cfqd->cfq_latency = 1;
4687         cfqd->hw_tag = -1;
4688         /*
4689          * we optimistically start assuming sync ops weren't delayed in last
4690          * second, in order to have larger depth for async operations.
4691          */
4692         cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC;
4693         return 0;
4694
4695 out_free:
4696         kfree(cfqd);
4697         kobject_put(&eq->kobj);
4698         return ret;
4699 }
4700
4701 static void cfq_registered_queue(struct request_queue *q)
4702 {
4703         struct elevator_queue *e = q->elevator;
4704         struct cfq_data *cfqd = e->elevator_data;
4705
4706         /*
4707          * Default to IOPS mode with no idling for SSDs
4708          */
4709         if (blk_queue_nonrot(q))
4710                 cfqd->cfq_slice_idle = 0;
4711 }
4712
4713 /*
4714  * sysfs parts below -->
4715  */
4716 static ssize_t
4717 cfq_var_show(unsigned int var, char *page)
4718 {
4719         return sprintf(page, "%u\n", var);
4720 }
4721
4722 static ssize_t
4723 cfq_var_store(unsigned int *var, const char *page, size_t count)
4724 {
4725         char *p = (char *) page;
4726
4727         *var = simple_strtoul(p, &p, 10);
4728         return count;
4729 }
4730
4731 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4732 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4733 {                                                                       \
4734         struct cfq_data *cfqd = e->elevator_data;                       \
4735         u64 __data = __VAR;                                             \
4736         if (__CONV)                                                     \
4737                 __data = div_u64(__data, NSEC_PER_MSEC);                        \
4738         return cfq_var_show(__data, (page));                            \
4739 }
4740 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4741 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4742 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4743 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4744 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4745 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4746 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4747 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4748 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4749 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4750 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4751 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4752 #undef SHOW_FUNCTION
4753
4754 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)                               \
4755 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4756 {                                                                       \
4757         struct cfq_data *cfqd = e->elevator_data;                       \
4758         u64 __data = __VAR;                                             \
4759         __data = div_u64(__data, NSEC_PER_USEC);                        \
4760         return cfq_var_show(__data, (page));                            \
4761 }
4762 USEC_SHOW_FUNCTION(cfq_slice_idle_us_show, cfqd->cfq_slice_idle);
4763 USEC_SHOW_FUNCTION(cfq_group_idle_us_show, cfqd->cfq_group_idle);
4764 USEC_SHOW_FUNCTION(cfq_slice_sync_us_show, cfqd->cfq_slice[1]);
4765 USEC_SHOW_FUNCTION(cfq_slice_async_us_show, cfqd->cfq_slice[0]);
4766 USEC_SHOW_FUNCTION(cfq_target_latency_us_show, cfqd->cfq_target_latency);
4767 #undef USEC_SHOW_FUNCTION
4768
4769 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4770 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4771 {                                                                       \
4772         struct cfq_data *cfqd = e->elevator_data;                       \
4773         unsigned int __data;                                            \
4774         int ret = cfq_var_store(&__data, (page), count);                \
4775         if (__data < (MIN))                                             \
4776                 __data = (MIN);                                         \
4777         else if (__data > (MAX))                                        \
4778                 __data = (MAX);                                         \
4779         if (__CONV)                                                     \
4780                 *(__PTR) = (u64)__data * NSEC_PER_MSEC;                 \
4781         else                                                            \
4782                 *(__PTR) = __data;                                      \
4783         return ret;                                                     \
4784 }
4785 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4786 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4787                 UINT_MAX, 1);
4788 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4789                 UINT_MAX, 1);
4790 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4791 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4792                 UINT_MAX, 0);
4793 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4794 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4795 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4796 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4797 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4798                 UINT_MAX, 0);
4799 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4800 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4801 #undef STORE_FUNCTION
4802
4803 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)                    \
4804 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4805 {                                                                       \
4806         struct cfq_data *cfqd = e->elevator_data;                       \
4807         unsigned int __data;                                            \
4808         int ret = cfq_var_store(&__data, (page), count);                \
4809         if (__data < (MIN))                                             \
4810                 __data = (MIN);                                         \
4811         else if (__data > (MAX))                                        \
4812                 __data = (MAX);                                         \
4813         *(__PTR) = (u64)__data * NSEC_PER_USEC;                         \
4814         return ret;                                                     \
4815 }
4816 USEC_STORE_FUNCTION(cfq_slice_idle_us_store, &cfqd->cfq_slice_idle, 0, UINT_MAX);
4817 USEC_STORE_FUNCTION(cfq_group_idle_us_store, &cfqd->cfq_group_idle, 0, UINT_MAX);
4818 USEC_STORE_FUNCTION(cfq_slice_sync_us_store, &cfqd->cfq_slice[1], 1, UINT_MAX);
4819 USEC_STORE_FUNCTION(cfq_slice_async_us_store, &cfqd->cfq_slice[0], 1, UINT_MAX);
4820 USEC_STORE_FUNCTION(cfq_target_latency_us_store, &cfqd->cfq_target_latency, 1, UINT_MAX);
4821 #undef USEC_STORE_FUNCTION
4822
4823 #define CFQ_ATTR(name) \
4824         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4825
4826 static struct elv_fs_entry cfq_attrs[] = {
4827         CFQ_ATTR(quantum),
4828         CFQ_ATTR(fifo_expire_sync),
4829         CFQ_ATTR(fifo_expire_async),
4830         CFQ_ATTR(back_seek_max),
4831         CFQ_ATTR(back_seek_penalty),
4832         CFQ_ATTR(slice_sync),
4833         CFQ_ATTR(slice_sync_us),
4834         CFQ_ATTR(slice_async),
4835         CFQ_ATTR(slice_async_us),
4836         CFQ_ATTR(slice_async_rq),
4837         CFQ_ATTR(slice_idle),
4838         CFQ_ATTR(slice_idle_us),
4839         CFQ_ATTR(group_idle),
4840         CFQ_ATTR(group_idle_us),
4841         CFQ_ATTR(low_latency),
4842         CFQ_ATTR(target_latency),
4843         CFQ_ATTR(target_latency_us),
4844         __ATTR_NULL
4845 };
4846
4847 static struct elevator_type iosched_cfq = {
4848         .ops = {
4849                 .elevator_merge_fn =            cfq_merge,
4850                 .elevator_merged_fn =           cfq_merged_request,
4851                 .elevator_merge_req_fn =        cfq_merged_requests,
4852                 .elevator_allow_bio_merge_fn =  cfq_allow_bio_merge,
4853                 .elevator_allow_rq_merge_fn =   cfq_allow_rq_merge,
4854                 .elevator_bio_merged_fn =       cfq_bio_merged,
4855                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4856                 .elevator_add_req_fn =          cfq_insert_request,
4857                 .elevator_activate_req_fn =     cfq_activate_request,
4858                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4859                 .elevator_completed_req_fn =    cfq_completed_request,
4860                 .elevator_former_req_fn =       elv_rb_former_request,
4861                 .elevator_latter_req_fn =       elv_rb_latter_request,
4862                 .elevator_init_icq_fn =         cfq_init_icq,
4863                 .elevator_exit_icq_fn =         cfq_exit_icq,
4864                 .elevator_set_req_fn =          cfq_set_request,
4865                 .elevator_put_req_fn =          cfq_put_request,
4866                 .elevator_may_queue_fn =        cfq_may_queue,
4867                 .elevator_init_fn =             cfq_init_queue,
4868                 .elevator_exit_fn =             cfq_exit_queue,
4869                 .elevator_registered_fn =       cfq_registered_queue,
4870         },
4871         .icq_size       =       sizeof(struct cfq_io_cq),
4872         .icq_align      =       __alignof__(struct cfq_io_cq),
4873         .elevator_attrs =       cfq_attrs,
4874         .elevator_name  =       "cfq",
4875         .elevator_owner =       THIS_MODULE,
4876 };
4877
4878 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4879 static struct blkcg_policy blkcg_policy_cfq = {
4880         .dfl_cftypes            = cfq_blkcg_files,
4881         .legacy_cftypes         = cfq_blkcg_legacy_files,
4882
4883         .cpd_alloc_fn           = cfq_cpd_alloc,
4884         .cpd_init_fn            = cfq_cpd_init,
4885         .cpd_free_fn            = cfq_cpd_free,
4886         .cpd_bind_fn            = cfq_cpd_bind,
4887
4888         .pd_alloc_fn            = cfq_pd_alloc,
4889         .pd_init_fn             = cfq_pd_init,
4890         .pd_offline_fn          = cfq_pd_offline,
4891         .pd_free_fn             = cfq_pd_free,
4892         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4893 };
4894 #endif
4895
4896 static int __init cfq_init(void)
4897 {
4898         int ret;
4899
4900 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4901         ret = blkcg_policy_register(&blkcg_policy_cfq);
4902         if (ret)
4903                 return ret;
4904 #else
4905         cfq_group_idle = 0;
4906 #endif
4907
4908         ret = -ENOMEM;
4909         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4910         if (!cfq_pool)
4911                 goto err_pol_unreg;
4912
4913         ret = elv_register(&iosched_cfq);
4914         if (ret)
4915                 goto err_free_pool;
4916
4917         return 0;
4918
4919 err_free_pool:
4920         kmem_cache_destroy(cfq_pool);
4921 err_pol_unreg:
4922 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4923         blkcg_policy_unregister(&blkcg_policy_cfq);
4924 #endif
4925         return ret;
4926 }
4927
4928 static void __exit cfq_exit(void)
4929 {
4930 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4931         blkcg_policy_unregister(&blkcg_policy_cfq);
4932 #endif
4933         elv_unregister(&iosched_cfq);
4934         kmem_cache_destroy(cfq_pool);
4935 }
4936
4937 module_init(cfq_init);
4938 module_exit(cfq_exit);
4939
4940 MODULE_AUTHOR("Jens Axboe");
4941 MODULE_LICENSE("GPL");
4942 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");