block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <linux/sched/clock.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/elevator.h>
  14 #include <linux/ktime.h>
  15 #include <linux/rbtree.h>
  16 #include <linux/ioprio.h>
  17 #include <linux/blktrace_api.h>
  18 #include <linux/blk-cgroup.h>
  19 #include "blk.h"
  20 #include "blk-wbt.h"
  21
  22 /*
  23  * tunables
  24  */
  25 /* max queue in one round of service */
  26 static const int cfq_quantum = 8;
  27 static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
  28 /* maximum backwards seek, in KiB */
  29 static const int cfq_back_max = 16 * 1024;
  30 /* penalty of a backwards seek */
  31 static const int cfq_back_penalty = 2;
  32 static const u64 cfq_slice_sync = NSEC_PER_SEC / 10;
  33 static u64 cfq_slice_async = NSEC_PER_SEC / 25;
  34 static const int cfq_slice_async_rq = 2;
  35 static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
  36 static u64 cfq_group_idle = NSEC_PER_SEC / 125;
  37 static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
  38 static const int cfq_hist_divisor = 4;
  39
  40 /*
  41  * offset from end of service tree
  42  */
  43 #define CFQ_IDLE_DELAY          (NSEC_PER_SEC / 5)
  44
  45 /*
  46  * below this threshold, we consider thinktime immediate
  47  */
  48 #define CFQ_MIN_TT              (2 * NSEC_PER_SEC / HZ)
  49
  50 #define CFQ_SLICE_SCALE         (5)
  51 #define CFQ_HW_QUEUE_MIN        (5)
  52 #define CFQ_SERVICE_SHIFT       12
  53
  54 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  55 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  56 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  57 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  58
  59 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  60 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  61 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  62
  63 static struct kmem_cache *cfq_pool;
  64
  65 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  66 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  67 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  68
  69 #define sample_valid(samples)   ((samples) > 80)
  70 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  71
  72 /* blkio-related constants */
  73 #define CFQ_WEIGHT_LEGACY_MIN   10
  74 #define CFQ_WEIGHT_LEGACY_DFL   500
  75 #define CFQ_WEIGHT_LEGACY_MAX   1000
  76
  77 struct cfq_ttime {
  78         u64 last_end_request;
  79
  80         u64 ttime_total;
  81         u64 ttime_mean;
  82         unsigned long ttime_samples;
  83 };
  84
  85 /*
  86  * Most of our rbtree usage is for sorting with min extraction, so
  87  * if we cache the leftmost node we don't have to walk down the tree
  88  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  89  * move this into the elevator for the rq sorting as well.
  90  */
  91 struct cfq_rb_root {
  92         struct rb_root rb;
  93         struct rb_node *left;
  94         unsigned count;
  95         u64 min_vdisktime;
  96         struct cfq_ttime ttime;
  97 };
  98 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  99                         .ttime = {.last_end_request = ktime_get_ns(),},}
 100
 101 /*
 102  * Per process-grouping structure
 103  */
 104 struct cfq_queue {
 105         /* reference count */
 106         int ref;
 107         /* various state flags, see below */
 108         unsigned int flags;
 109         /* parent cfq_data */
 110         struct cfq_data *cfqd;
 111         /* service_tree member */
 112         struct rb_node rb_node;
 113         /* service_tree key */
 114         u64 rb_key;
 115         /* prio tree member */
 116         struct rb_node p_node;
 117         /* prio tree root we belong to, if any */
 118         struct rb_root *p_root;
 119         /* sorted list of pending requests */
 120         struct rb_root sort_list;
 121         /* if fifo isn't expired, next request to serve */
 122         struct request *next_rq;
 123         /* requests queued in sort_list */
 124         int queued[2];
 125         /* currently allocated requests */
 126         int allocated[2];
 127         /* fifo list of requests in sort_list */
 128         struct list_head fifo;
 129
 130         /* time when queue got scheduled in to dispatch first request. */
 131         u64 dispatch_start;
 132         u64 allocated_slice;
 133         u64 slice_dispatch;
 134         /* time when first request from queue completed and slice started. */
 135         u64 slice_start;
 136         u64 slice_end;
 137         s64 slice_resid;
 138
 139         /* pending priority requests */
 140         int prio_pending;
 141         /* number of requests that are on the dispatch list or inside driver */
 142         int dispatched;
 143
 144         /* io prio of this group */
 145         unsigned short ioprio, org_ioprio;
 146         unsigned short ioprio_class, org_ioprio_class;
 147
 148         pid_t pid;
 149
 150         u32 seek_history;
 151         sector_t last_request_pos;
 152
 153         struct cfq_rb_root *service_tree;
 154         struct cfq_queue *new_cfqq;
 155         struct cfq_group *cfqg;
 156         /* Number of sectors dispatched from queue in single dispatch round */
 157         unsigned long nr_sectors;
 158 };
 159
 160 /*
 161  * First index in the service_trees.
 162  * IDLE is handled separately, so it has negative index
 163  */
 164 enum wl_class_t {
 165         BE_WORKLOAD = 0,
 166         RT_WORKLOAD = 1,
 167         IDLE_WORKLOAD = 2,
 168         CFQ_PRIO_NR,
 169 };
 170
 171 /*
 172  * Second index in the service_trees.
 173  */
 174 enum wl_type_t {
 175         ASYNC_WORKLOAD = 0,
 176         SYNC_NOIDLE_WORKLOAD = 1,
 177         SYNC_WORKLOAD = 2
 178 };
 179
 180 struct cfqg_stats {
 181 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 182         /* number of ios merged */
 183         struct blkg_rwstat              merged;
 184         /* total time spent on device in ns, may not be accurate w/ queueing */
 185         struct blkg_rwstat              service_time;
 186         /* total time spent waiting in scheduler queue in ns */
 187         struct blkg_rwstat              wait_time;
 188         /* number of IOs queued up */
 189         struct blkg_rwstat              queued;
 190         /* total disk time and nr sectors dispatched by this group */
 191         struct blkg_stat                time;
 192 #ifdef CONFIG_DEBUG_BLK_CGROUP
 193         /* time not charged to this cgroup */
 194         struct blkg_stat                unaccounted_time;
 195         /* sum of number of ios queued across all samples */
 196         struct blkg_stat                avg_queue_size_sum;
 197         /* count of samples taken for average */
 198         struct blkg_stat                avg_queue_size_samples;
 199         /* how many times this group has been removed from service tree */
 200         struct blkg_stat                dequeue;
 201         /* total time spent waiting for it to be assigned a timeslice. */
 202         struct blkg_stat                group_wait_time;
 203         /* time spent idling for this blkcg_gq */
 204         struct blkg_stat                idle_time;
 205         /* total time with empty current active q with other requests queued */
 206         struct blkg_stat                empty_time;
 207         /* fields after this shouldn't be cleared on stat reset */
 208         uint64_t                        start_group_wait_time;
 209         uint64_t                        start_idle_time;
 210         uint64_t                        start_empty_time;
 211         uint16_t                        flags;
 212 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 213 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 214 };
 215
 216 /* Per-cgroup data */
 217 struct cfq_group_data {
 218         /* must be the first member */
 219         struct blkcg_policy_data cpd;
 220
 221         unsigned int weight;
 222         unsigned int leaf_weight;
 223 };
 224
 225 /* This is per cgroup per device grouping structure */
 226 struct cfq_group {
 227         /* must be the first member */
 228         struct blkg_policy_data pd;
 229
 230         /* group service_tree member */
 231         struct rb_node rb_node;
 232
 233         /* group service_tree key */
 234         u64 vdisktime;
 235
 236         /*
 237          * The number of active cfqgs and sum of their weights under this
 238          * cfqg.  This covers this cfqg's leaf_weight and all children's
 239          * weights, but does not cover weights of further descendants.
 240          *
 241          * If a cfqg is on the service tree, it's active.  An active cfqg
 242          * also activates its parent and contributes to the children_weight
 243          * of the parent.
 244          */
 245         int nr_active;
 246         unsigned int children_weight;
 247
 248         /*
 249          * vfraction is the fraction of vdisktime that the tasks in this
 250          * cfqg are entitled to.  This is determined by compounding the
 251          * ratios walking up from this cfqg to the root.
 252          *
 253          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 254          * vfractions on a service tree is approximately 1.  The sum may
 255          * deviate a bit due to rounding errors and fluctuations caused by
 256          * cfqgs entering and leaving the service tree.
 257          */
 258         unsigned int vfraction;
 259
 260         /*
 261          * There are two weights - (internal) weight is the weight of this
 262          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 263          * this cfqg against the child cfqgs.  For the root cfqg, both
 264          * weights are kept in sync for backward compatibility.
 265          */
 266         unsigned int weight;
 267         unsigned int new_weight;
 268         unsigned int dev_weight;
 269
 270         unsigned int leaf_weight;
 271         unsigned int new_leaf_weight;
 272         unsigned int dev_leaf_weight;
 273
 274         /* number of cfqq currently on this group */
 275         int nr_cfqq;
 276
 277         /*
 278          * Per group busy queues average. Useful for workload slice calc. We
 279          * create the array for each prio class but at run time it is used
 280          * only for RT and BE class and slot for IDLE class remains unused.
 281          * This is primarily done to avoid confusion and a gcc warning.
 282          */
 283         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 284         /*
 285          * rr lists of queues with requests. We maintain service trees for
 286          * RT and BE classes. These trees are subdivided in subclasses
 287          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 288          * class there is no subclassification and all the cfq queues go on
 289          * a single tree service_tree_idle.
 290          * Counts are embedded in the cfq_rb_root
 291          */
 292         struct cfq_rb_root service_trees[2][3];
 293         struct cfq_rb_root service_tree_idle;
 294
 295         u64 saved_wl_slice;
 296         enum wl_type_t saved_wl_type;
 297         enum wl_class_t saved_wl_class;
 298
 299         /* number of requests that are on the dispatch list or inside driver */
 300         int dispatched;
 301         struct cfq_ttime ttime;
 302         struct cfqg_stats stats;        /* stats for this cfqg */
 303
 304         /* async queue for each priority case */
 305         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 306         struct cfq_queue *async_idle_cfqq;
 307
 308 };
 309
 310 struct cfq_io_cq {
 311         struct io_cq            icq;            /* must be the first member */
 312         struct cfq_queue        *cfqq[2];
 313         struct cfq_ttime        ttime;
 314         int                     ioprio;         /* the current ioprio */
 315 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 316         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 317 #endif
 318 };
 319
 320 /*
 321  * Per block device queue structure
 322  */
 323 struct cfq_data {
 324         struct request_queue *queue;
 325         /* Root service tree for cfq_groups */
 326         struct cfq_rb_root grp_service_tree;
 327         struct cfq_group *root_group;
 328
 329         /*
 330          * The priority currently being served
 331          */
 332         enum wl_class_t serving_wl_class;
 333         enum wl_type_t serving_wl_type;
 334         u64 workload_expires;
 335         struct cfq_group *serving_group;
 336
 337         /*
 338          * Each priority tree is sorted by next_request position.  These
 339          * trees are used when determining if two or more queues are
 340          * interleaving requests (see cfq_close_cooperator).
 341          */
 342         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 343
 344         unsigned int busy_queues;
 345         unsigned int busy_sync_queues;
 346
 347         int rq_in_driver;
 348         int rq_in_flight[2];
 349
 350         /*
 351          * queue-depth detection
 352          */
 353         int rq_queued;
 354         int hw_tag;
 355         /*
 356          * hw_tag can be
 357          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 358          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 359          *  0 => no NCQ
 360          */
 361         int hw_tag_est_depth;
 362         unsigned int hw_tag_samples;
 363
 364         /*
 365          * idle window management
 366          */
 367         struct hrtimer idle_slice_timer;
 368         struct work_struct unplug_work;
 369
 370         struct cfq_queue *active_queue;
 371         struct cfq_io_cq *active_cic;
 372
 373         sector_t last_position;
 374
 375         /*
 376          * tunables, see top of file
 377          */
 378         unsigned int cfq_quantum;
 379         unsigned int cfq_back_penalty;
 380         unsigned int cfq_back_max;
 381         unsigned int cfq_slice_async_rq;
 382         unsigned int cfq_latency;
 383         u64 cfq_fifo_expire[2];
 384         u64 cfq_slice[2];
 385         u64 cfq_slice_idle;
 386         u64 cfq_group_idle;
 387         u64 cfq_target_latency;
 388
 389         /*
 390          * Fallback dummy cfqq for extreme OOM conditions
 391          */
 392         struct cfq_queue oom_cfqq;
 393
 394         u64 last_delayed_sync;
 395 };
 396
 397 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 398 static void cfq_put_queue(struct cfq_queue *cfqq);
 399
 400 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 401                                             enum wl_class_t class,
 402                                             enum wl_type_t type)
 403 {
 404         if (!cfqg)
 405                 return NULL;
 406
 407         if (class == IDLE_WORKLOAD)
 408                 return &cfqg->service_tree_idle;
 409
 410         return &cfqg->service_trees[class][type];
 411 }
 412
 413 enum cfqq_state_flags {
 414         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 415         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 416         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 417         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 418         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 419         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 420         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 421         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 422         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 423         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 424         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 425         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 426         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 427 };
 428
 429 #define CFQ_CFQQ_FNS(name)                                              \
 430 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 431 {                                                                       \
 432         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 433 }                                                                       \
 434 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 435 {                                                                       \
 436         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 437 }                                                                       \
 438 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 439 {                                                                       \
 440         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 441 }
 442
 443 CFQ_CFQQ_FNS(on_rr);
 444 CFQ_CFQQ_FNS(wait_request);
 445 CFQ_CFQQ_FNS(must_dispatch);
 446 CFQ_CFQQ_FNS(must_alloc_slice);
 447 CFQ_CFQQ_FNS(fifo_expire);
 448 CFQ_CFQQ_FNS(idle_window);
 449 CFQ_CFQQ_FNS(prio_changed);
 450 CFQ_CFQQ_FNS(slice_new);
 451 CFQ_CFQQ_FNS(sync);
 452 CFQ_CFQQ_FNS(coop);
 453 CFQ_CFQQ_FNS(split_coop);
 454 CFQ_CFQQ_FNS(deep);
 455 CFQ_CFQQ_FNS(wait_busy);
 456 #undef CFQ_CFQQ_FNS
 457
 458 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 459
 460 /* cfqg stats flags */
 461 enum cfqg_stats_flags {
 462         CFQG_stats_waiting = 0,
 463         CFQG_stats_idling,
 464         CFQG_stats_empty,
 465 };
 466
 467 #define CFQG_FLAG_FNS(name)                                             \
 468 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 469 {                                                                       \
 470         stats->flags |= (1 << CFQG_stats_##name);                       \
 471 }                                                                       \
 472 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 473 {                                                                       \
 474         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 475 }                                                                       \
 476 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 477 {                                                                       \
 478         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 479 }                                                                       \
 480
 481 CFQG_FLAG_FNS(waiting)
 482 CFQG_FLAG_FNS(idling)
 483 CFQG_FLAG_FNS(empty)
 484 #undef CFQG_FLAG_FNS
 485
 486 /* This should be called with the queue_lock held. */
 487 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 488 {
 489         unsigned long long now;
 490
 491         if (!cfqg_stats_waiting(stats))
 492                 return;
 493
 494         now = sched_clock();
 495         if (time_after64(now, stats->start_group_wait_time))
 496                 blkg_stat_add(&stats->group_wait_time,
 497                               now - stats->start_group_wait_time);
 498         cfqg_stats_clear_waiting(stats);
 499 }
 500
 501 /* This should be called with the queue_lock held. */
 502 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 503                                                  struct cfq_group *curr_cfqg)
 504 {
 505         struct cfqg_stats *stats = &cfqg->stats;
 506
 507         if (cfqg_stats_waiting(stats))
 508                 return;
 509         if (cfqg == curr_cfqg)
 510                 return;
 511         stats->start_group_wait_time = sched_clock();
 512         cfqg_stats_mark_waiting(stats);
 513 }
 514
 515 /* This should be called with the queue_lock held. */
 516 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 517 {
 518         unsigned long long now;
 519
 520         if (!cfqg_stats_empty(stats))
 521                 return;
 522
 523         now = sched_clock();
 524         if (time_after64(now, stats->start_empty_time))
 525                 blkg_stat_add(&stats->empty_time,
 526                               now - stats->start_empty_time);
 527         cfqg_stats_clear_empty(stats);
 528 }
 529
 530 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 531 {
 532         blkg_stat_add(&cfqg->stats.dequeue, 1);
 533 }
 534
 535 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 536 {
 537         struct cfqg_stats *stats = &cfqg->stats;
 538
 539         if (blkg_rwstat_total(&stats->queued))
 540                 return;
 541
 542         /*
 543          * group is already marked empty. This can happen if cfqq got new
 544          * request in parent group and moved to this group while being added
 545          * to service tree. Just ignore the event and move on.
 546          */
 547         if (cfqg_stats_empty(stats))
 548                 return;
 549
 550         stats->start_empty_time = sched_clock();
 551         cfqg_stats_mark_empty(stats);
 552 }
 553
 554 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 555 {
 556         struct cfqg_stats *stats = &cfqg->stats;
 557
 558         if (cfqg_stats_idling(stats)) {
 559                 unsigned long long now = sched_clock();
 560
 561                 if (time_after64(now, stats->start_idle_time))
 562                         blkg_stat_add(&stats->idle_time,
 563                                       now - stats->start_idle_time);
 564                 cfqg_stats_clear_idling(stats);
 565         }
 566 }
 567
 568 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 569 {
 570         struct cfqg_stats *stats = &cfqg->stats;
 571
 572         BUG_ON(cfqg_stats_idling(stats));
 573
 574         stats->start_idle_time = sched_clock();
 575         cfqg_stats_mark_idling(stats);
 576 }
 577
 578 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 579 {
 580         struct cfqg_stats *stats = &cfqg->stats;
 581
 582         blkg_stat_add(&stats->avg_queue_size_sum,
 583                       blkg_rwstat_total(&stats->queued));
 584         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 585         cfqg_stats_update_group_wait_time(stats);
 586 }
 587
 588 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 589
 590 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 591 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 592 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 593 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 594 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 595 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 596 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 597
 598 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 599
 600 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 601
 602 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 603 {
 604         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 605 }
 606
 607 static struct cfq_group_data
 608 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 609 {
 610         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 611 }
 612
 613 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 614 {
 615         return pd_to_blkg(&cfqg->pd);
 616 }
 617
 618 static struct blkcg_policy blkcg_policy_cfq;
 619
 620 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 621 {
 622         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 623 }
 624
 625 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 626 {
 627         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 628 }
 629
 630 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 631 {
 632         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 633
 634         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 635 }
 636
 637 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 638                                       struct cfq_group *ancestor)
 639 {
 640         return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
 641                                     cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
 642 }
 643
 644 static inline void cfqg_get(struct cfq_group *cfqg)
 645 {
 646         return blkg_get(cfqg_to_blkg(cfqg));
 647 }
 648
 649 static inline void cfqg_put(struct cfq_group *cfqg)
 650 {
 651         return blkg_put(cfqg_to_blkg(cfqg));
 652 }
 653
 654 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 655         char __pbuf[128];                                               \
 656                                                                         \
 657         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 658         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 659                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 660                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 661                           __pbuf, ##args);                              \
 662 } while (0)
 663
 664 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 665         char __pbuf[128];                                               \
 666                                                                         \
 667         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 668         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 669 } while (0)
 670
 671 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 672                                             struct cfq_group *curr_cfqg,
 673                                             unsigned int op)
 674 {
 675         blkg_rwstat_add(&cfqg->stats.queued, op, 1);
 676         cfqg_stats_end_empty_time(&cfqg->stats);
 677         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 678 }
 679
 680 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 681                         uint64_t time, unsigned long unaccounted_time)
 682 {
 683         blkg_stat_add(&cfqg->stats.time, time);
 684 #ifdef CONFIG_DEBUG_BLK_CGROUP
 685         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 686 #endif
 687 }
 688
 689 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
 690                                                unsigned int op)
 691 {
 692         blkg_rwstat_add(&cfqg->stats.queued, op, -1);
 693 }
 694
 695 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
 696                                                unsigned int op)
 697 {
 698         blkg_rwstat_add(&cfqg->stats.merged, op, 1);
 699 }
 700
 701 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 702                         uint64_t start_time, uint64_t io_start_time,
 703                         unsigned int op)
 704 {
 705         struct cfqg_stats *stats = &cfqg->stats;
 706         unsigned long long now = sched_clock();
 707
 708         if (time_after64(now, io_start_time))
 709                 blkg_rwstat_add(&stats->service_time, op, now - io_start_time);
 710         if (time_after64(io_start_time, start_time))
 711                 blkg_rwstat_add(&stats->wait_time, op,
 712                                 io_start_time - start_time);
 713 }
 714
 715 /* @stats = 0 */
 716 static void cfqg_stats_reset(struct cfqg_stats *stats)
 717 {
 718         /* queued stats shouldn't be cleared */
 719         blkg_rwstat_reset(&stats->merged);
 720         blkg_rwstat_reset(&stats->service_time);
 721         blkg_rwstat_reset(&stats->wait_time);
 722         blkg_stat_reset(&stats->time);
 723 #ifdef CONFIG_DEBUG_BLK_CGROUP
 724         blkg_stat_reset(&stats->unaccounted_time);
 725         blkg_stat_reset(&stats->avg_queue_size_sum);
 726         blkg_stat_reset(&stats->avg_queue_size_samples);
 727         blkg_stat_reset(&stats->dequeue);
 728         blkg_stat_reset(&stats->group_wait_time);
 729         blkg_stat_reset(&stats->idle_time);
 730         blkg_stat_reset(&stats->empty_time);
 731 #endif
 732 }
 733
 734 /* @to += @from */
 735 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 736 {
 737         /* queued stats shouldn't be cleared */
 738         blkg_rwstat_add_aux(&to->merged, &from->merged);
 739         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 740         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 741         blkg_stat_add_aux(&from->time, &from->time);
 742 #ifdef CONFIG_DEBUG_BLK_CGROUP
 743         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 744         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 745         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 746         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 747         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 748         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 749         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 750 #endif
 751 }
 752
 753 /*
 754  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 755  * recursive stats can still account for the amount used by this cfqg after
 756  * it's gone.
 757  */
 758 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 759 {
 760         struct cfq_group *parent = cfqg_parent(cfqg);
 761
 762         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 763
 764         if (unlikely(!parent))
 765                 return;
 766
 767         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 768         cfqg_stats_reset(&cfqg->stats);
 769 }
 770
 771 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 772
 773 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 774 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
 775                                       struct cfq_group *ancestor)
 776 {
 777         return true;
 778 }
 779 static inline void cfqg_get(struct cfq_group *cfqg) { }
 780 static inline void cfqg_put(struct cfq_group *cfqg) { }
 781
 782 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 783         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 784                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 785                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 786                                 ##args)
 787 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 788
 789 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 790                         struct cfq_group *curr_cfqg, unsigned int op) { }
 791 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 792                         uint64_t time, unsigned long unaccounted_time) { }
 793 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
 794                         unsigned int op) { }
 795 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
 796                         unsigned int op) { }
 797 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 798                         uint64_t start_time, uint64_t io_start_time,
 799                         unsigned int op) { }
 800
 801 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 802
 803 #define cfq_log(cfqd, fmt, args...)     \
 804         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 805
 806 /* Traverses through cfq group service trees */
 807 #define for_each_cfqg_st(cfqg, i, j, st) \
 808         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 809                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 810                         : &cfqg->service_tree_idle; \
 811                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 812                         (i == IDLE_WORKLOAD && j == 0); \
 813                         j++, st = i < IDLE_WORKLOAD ? \
 814                         &cfqg->service_trees[i][j]: NULL) \
 815
 816 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 817         struct cfq_ttime *ttime, bool group_idle)
 818 {
 819         u64 slice;
 820         if (!sample_valid(ttime->ttime_samples))
 821                 return false;
 822         if (group_idle)
 823                 slice = cfqd->cfq_group_idle;
 824         else
 825                 slice = cfqd->cfq_slice_idle;
 826         return ttime->ttime_mean > slice;
 827 }
 828
 829 static inline bool iops_mode(struct cfq_data *cfqd)
 830 {
 831         /*
 832          * If we are not idling on queues and it is a NCQ drive, parallel
 833          * execution of requests is on and measuring time is not possible
 834          * in most of the cases until and unless we drive shallower queue
 835          * depths and that becomes a performance bottleneck. In such cases
 836          * switch to start providing fairness in terms of number of IOs.
 837          */
 838         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 839                 return true;
 840         else
 841                 return false;
 842 }
 843
 844 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 845 {
 846         if (cfq_class_idle(cfqq))
 847                 return IDLE_WORKLOAD;
 848         if (cfq_class_rt(cfqq))
 849                 return RT_WORKLOAD;
 850         return BE_WORKLOAD;
 851 }
 852
 853
 854 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 855 {
 856         if (!cfq_cfqq_sync(cfqq))
 857                 return ASYNC_WORKLOAD;
 858         if (!cfq_cfqq_idle_window(cfqq))
 859                 return SYNC_NOIDLE_WORKLOAD;
 860         return SYNC_WORKLOAD;
 861 }
 862
 863 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 864                                         struct cfq_data *cfqd,
 865                                         struct cfq_group *cfqg)
 866 {
 867         if (wl_class == IDLE_WORKLOAD)
 868                 return cfqg->service_tree_idle.count;
 869
 870         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 871                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 872                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 873 }
 874
 875 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 876                                         struct cfq_group *cfqg)
 877 {
 878         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 879                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 880 }
 881
 882 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 883 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 884                                        struct cfq_io_cq *cic, struct bio *bio);
 885
 886 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 887 {
 888         /* cic->icq is the first member, %NULL will convert to %NULL */
 889         return container_of(icq, struct cfq_io_cq, icq);
 890 }
 891
 892 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 893                                                struct io_context *ioc)
 894 {
 895         if (ioc)
 896                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 897         return NULL;
 898 }
 899
 900 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 901 {
 902         return cic->cfqq[is_sync];
 903 }
 904
 905 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 906                                 bool is_sync)
 907 {
 908         cic->cfqq[is_sync] = cfqq;
 909 }
 910
 911 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 912 {
 913         return cic->icq.q->elevator->elevator_data;
 914 }
 915
 916 /*
 917  * scheduler run of queue, if there are requests pending and no one in the
 918  * driver that will restart queueing
 919  */
 920 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 921 {
 922         if (cfqd->busy_queues) {
 923                 cfq_log(cfqd, "schedule dispatch");
 924                 kblockd_schedule_work(&cfqd->unplug_work);
 925         }
 926 }
 927
 928 /*
 929  * Scale schedule slice based on io priority. Use the sync time slice only
 930  * if a queue is marked sync and has sync io queued. A sync queue with async
 931  * io only, should not get full sync slice length.
 932  */
 933 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 934                                  unsigned short prio)
 935 {
 936         u64 base_slice = cfqd->cfq_slice[sync];
 937         u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
 938
 939         WARN_ON(prio >= IOPRIO_BE_NR);
 940
 941         return base_slice + (slice * (4 - prio));
 942 }
 943
 944 static inline u64
 945 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 946 {
 947         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 948 }
 949
 950 /**
 951  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 952  * @charge: disk time being charged
 953  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 954  *
 955  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 956  * scaling is inversely proportional.
 957  *
 958  * scaled = charge / vfraction
 959  *
 960  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 961  */
 962 static inline u64 cfqg_scale_charge(u64 charge,
 963                                     unsigned int vfraction)
 964 {
 965         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 966
 967         /* charge / vfraction */
 968         c <<= CFQ_SERVICE_SHIFT;
 969         return div_u64(c, vfraction);
 970 }
 971
 972 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 973 {
 974         s64 delta = (s64)(vdisktime - min_vdisktime);
 975         if (delta > 0)
 976                 min_vdisktime = vdisktime;
 977
 978         return min_vdisktime;
 979 }
 980
 981 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 982 {
 983         s64 delta = (s64)(vdisktime - min_vdisktime);
 984         if (delta < 0)
 985                 min_vdisktime = vdisktime;
 986
 987         return min_vdisktime;
 988 }
 989
 990 static void update_min_vdisktime(struct cfq_rb_root *st)
 991 {
 992         struct cfq_group *cfqg;
 993
 994         if (st->left) {
 995                 cfqg = rb_entry_cfqg(st->left);
 996                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 997                                                   cfqg->vdisktime);
 998         }
 999 }
1000
1001 /*
1002  * get averaged number of queues of RT/BE priority.
1003  * average is updated, with a formula that gives more weight to higher numbers,
1004  * to quickly follows sudden increases and decrease slowly
1005  */
1006
1007 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1008                                         struct cfq_group *cfqg, bool rt)
1009 {
1010         unsigned min_q, max_q;
1011         unsigned mult  = cfq_hist_divisor - 1;
1012         unsigned round = cfq_hist_divisor / 2;
1013         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1014
1015         min_q = min(cfqg->busy_queues_avg[rt], busy);
1016         max_q = max(cfqg->busy_queues_avg[rt], busy);
1017         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1018                 cfq_hist_divisor;
1019         return cfqg->busy_queues_avg[rt];
1020 }
1021
1022 static inline u64
1023 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1024 {
1025         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1026 }
1027
1028 static inline u64
1029 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1030 {
1031         u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1032         if (cfqd->cfq_latency) {
1033                 /*
1034                  * interested queues (we consider only the ones with the same
1035                  * priority class in the cfq group)
1036                  */
1037                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1038                                                 cfq_class_rt(cfqq));
1039                 u64 sync_slice = cfqd->cfq_slice[1];
1040                 u64 expect_latency = sync_slice * iq;
1041                 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1042
1043                 if (expect_latency > group_slice) {
1044                         u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1045                         u64 low_slice;
1046
1047                         /* scale low_slice according to IO priority
1048                          * and sync vs async */
1049                         low_slice = div64_u64(base_low_slice*slice, sync_slice);
1050                         low_slice = min(slice, low_slice);
1051                         /* the adapted slice value is scaled to fit all iqs
1052                          * into the target latency */
1053                         slice = div64_u64(slice*group_slice, expect_latency);
1054                         slice = max(slice, low_slice);
1055                 }
1056         }
1057         return slice;
1058 }
1059
1060 static inline void
1061 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1062 {
1063         u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1064         u64 now = ktime_get_ns();
1065
1066         cfqq->slice_start = now;
1067         cfqq->slice_end = now + slice;
1068         cfqq->allocated_slice = slice;
1069         cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1070 }
1071
1072 /*
1073  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1074  * isn't valid until the first request from the dispatch is activated
1075  * and the slice time set.
1076  */
1077 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1078 {
1079         if (cfq_cfqq_slice_new(cfqq))
1080                 return false;
1081         if (ktime_get_ns() < cfqq->slice_end)
1082                 return false;
1083
1084         return true;
1085 }
1086
1087 /*
1088  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1089  * We choose the request that is closest to the head right now. Distance
1090  * behind the head is penalized and only allowed to a certain extent.
1091  */
1092 static struct request *
1093 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1094 {
1095         sector_t s1, s2, d1 = 0, d2 = 0;
1096         unsigned long back_max;
1097 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1098 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1099         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1100
1101         if (rq1 == NULL || rq1 == rq2)
1102                 return rq2;
1103         if (rq2 == NULL)
1104                 return rq1;
1105
1106         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1107                 return rq_is_sync(rq1) ? rq1 : rq2;
1108
1109         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1110                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1111
1112         s1 = blk_rq_pos(rq1);
1113         s2 = blk_rq_pos(rq2);
1114
1115         /*
1116          * by definition, 1KiB is 2 sectors
1117          */
1118         back_max = cfqd->cfq_back_max * 2;
1119
1120         /*
1121          * Strict one way elevator _except_ in the case where we allow
1122          * short backward seeks which are biased as twice the cost of a
1123          * similar forward seek.
1124          */
1125         if (s1 >= last)
1126                 d1 = s1 - last;
1127         else if (s1 + back_max >= last)
1128                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1129         else
1130                 wrap |= CFQ_RQ1_WRAP;
1131
1132         if (s2 >= last)
1133                 d2 = s2 - last;
1134         else if (s2 + back_max >= last)
1135                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1136         else
1137                 wrap |= CFQ_RQ2_WRAP;
1138
1139         /* Found required data */
1140
1141         /*
1142          * By doing switch() on the bit mask "wrap" we avoid having to
1143          * check two variables for all permutations: --> faster!
1144          */
1145         switch (wrap) {
1146         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1147                 if (d1 < d2)
1148                         return rq1;
1149                 else if (d2 < d1)
1150                         return rq2;
1151                 else {
1152                         if (s1 >= s2)
1153                                 return rq1;
1154                         else
1155                                 return rq2;
1156                 }
1157
1158         case CFQ_RQ2_WRAP:
1159                 return rq1;
1160         case CFQ_RQ1_WRAP:
1161                 return rq2;
1162         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1163         default:
1164                 /*
1165                  * Since both rqs are wrapped,
1166                  * start with the one that's further behind head
1167                  * (--> only *one* back seek required),
1168                  * since back seek takes more time than forward.
1169                  */
1170                 if (s1 <= s2)
1171                         return rq1;
1172                 else
1173                         return rq2;
1174         }
1175 }
1176
1177 /*
1178  * The below is leftmost cache rbtree addon
1179  */
1180 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1181 {
1182         /* Service tree is empty */
1183         if (!root->count)
1184                 return NULL;
1185
1186         if (!root->left)
1187                 root->left = rb_first(&root->rb);
1188
1189         if (root->left)
1190                 return rb_entry(root->left, struct cfq_queue, rb_node);
1191
1192         return NULL;
1193 }
1194
1195 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1196 {
1197         if (!root->left)
1198                 root->left = rb_first(&root->rb);
1199
1200         if (root->left)
1201                 return rb_entry_cfqg(root->left);
1202
1203         return NULL;
1204 }
1205
1206 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1207 {
1208         rb_erase(n, root);
1209         RB_CLEAR_NODE(n);
1210 }
1211
1212 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1213 {
1214         if (root->left == n)
1215                 root->left = NULL;
1216         rb_erase_init(n, &root->rb);
1217         --root->count;
1218 }
1219
1220 /*
1221  * would be nice to take fifo expire time into account as well
1222  */
1223 static struct request *
1224 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1225                   struct request *last)
1226 {
1227         struct rb_node *rbnext = rb_next(&last->rb_node);
1228         struct rb_node *rbprev = rb_prev(&last->rb_node);
1229         struct request *next = NULL, *prev = NULL;
1230
1231         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1232
1233         if (rbprev)
1234                 prev = rb_entry_rq(rbprev);
1235
1236         if (rbnext)
1237                 next = rb_entry_rq(rbnext);
1238         else {
1239                 rbnext = rb_first(&cfqq->sort_list);
1240                 if (rbnext && rbnext != &last->rb_node)
1241                         next = rb_entry_rq(rbnext);
1242         }
1243
1244         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1245 }
1246
1247 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1248                             struct cfq_queue *cfqq)
1249 {
1250         /*
1251          * just an approximation, should be ok.
1252          */
1253         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1254                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1255 }
1256
1257 static inline s64
1258 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1259 {
1260         return cfqg->vdisktime - st->min_vdisktime;
1261 }
1262
1263 static void
1264 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1265 {
1266         struct rb_node **node = &st->rb.rb_node;
1267         struct rb_node *parent = NULL;
1268         struct cfq_group *__cfqg;
1269         s64 key = cfqg_key(st, cfqg);
1270         int left = 1;
1271
1272         while (*node != NULL) {
1273                 parent = *node;
1274                 __cfqg = rb_entry_cfqg(parent);
1275
1276                 if (key < cfqg_key(st, __cfqg))
1277                         node = &parent->rb_left;
1278                 else {
1279                         node = &parent->rb_right;
1280                         left = 0;
1281                 }
1282         }
1283
1284         if (left)
1285                 st->left = &cfqg->rb_node;
1286
1287         rb_link_node(&cfqg->rb_node, parent, node);
1288         rb_insert_color(&cfqg->rb_node, &st->rb);
1289 }
1290
1291 /*
1292  * This has to be called only on activation of cfqg
1293  */
1294 static void
1295 cfq_update_group_weight(struct cfq_group *cfqg)
1296 {
1297         if (cfqg->new_weight) {
1298                 cfqg->weight = cfqg->new_weight;
1299                 cfqg->new_weight = 0;
1300         }
1301 }
1302
1303 static void
1304 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1305 {
1306         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1307
1308         if (cfqg->new_leaf_weight) {
1309                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1310                 cfqg->new_leaf_weight = 0;
1311         }
1312 }
1313
1314 static void
1315 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1316 {
1317         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1318         struct cfq_group *pos = cfqg;
1319         struct cfq_group *parent;
1320         bool propagate;
1321
1322         /* add to the service tree */
1323         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1324
1325         /*
1326          * Update leaf_weight.  We cannot update weight at this point
1327          * because cfqg might already have been activated and is
1328          * contributing its current weight to the parent's child_weight.
1329          */
1330         cfq_update_group_leaf_weight(cfqg);
1331         __cfq_group_service_tree_add(st, cfqg);
1332
1333         /*
1334          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1335          * entitled to.  vfraction is calculated by walking the tree
1336          * towards the root calculating the fraction it has at each level.
1337          * The compounded ratio is how much vfraction @cfqg owns.
1338          *
1339          * Start with the proportion tasks in this cfqg has against active
1340          * children cfqgs - its leaf_weight against children_weight.
1341          */
1342         propagate = !pos->nr_active++;
1343         pos->children_weight += pos->leaf_weight;
1344         vfr = vfr * pos->leaf_weight / pos->children_weight;
1345
1346         /*
1347          * Compound ->weight walking up the tree.  Both activation and
1348          * vfraction calculation are done in the same loop.  Propagation
1349          * stops once an already activated node is met.  vfraction
1350          * calculation should always continue to the root.
1351          */
1352         while ((parent = cfqg_parent(pos))) {
1353                 if (propagate) {
1354                         cfq_update_group_weight(pos);
1355                         propagate = !parent->nr_active++;
1356                         parent->children_weight += pos->weight;
1357                 }
1358                 vfr = vfr * pos->weight / parent->children_weight;
1359                 pos = parent;
1360         }
1361
1362         cfqg->vfraction = max_t(unsigned, vfr, 1);
1363 }
1364
1365 static void
1366 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1367 {
1368         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1369         struct cfq_group *__cfqg;
1370         struct rb_node *n;
1371
1372         cfqg->nr_cfqq++;
1373         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1374                 return;
1375
1376         /*
1377          * Currently put the group at the end. Later implement something
1378          * so that groups get lesser vtime based on their weights, so that
1379          * if group does not loose all if it was not continuously backlogged.
1380          */
1381         n = rb_last(&st->rb);
1382         if (n) {
1383                 __cfqg = rb_entry_cfqg(n);
1384                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1385         } else
1386                 cfqg->vdisktime = st->min_vdisktime;
1387         cfq_group_service_tree_add(st, cfqg);
1388 }
1389
1390 static void
1391 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1392 {
1393         struct cfq_group *pos = cfqg;
1394         bool propagate;
1395
1396         /*
1397          * Undo activation from cfq_group_service_tree_add().  Deactivate
1398          * @cfqg and propagate deactivation upwards.
1399          */
1400         propagate = !--pos->nr_active;
1401         pos->children_weight -= pos->leaf_weight;
1402
1403         while (propagate) {
1404                 struct cfq_group *parent = cfqg_parent(pos);
1405
1406                 /* @pos has 0 nr_active at this point */
1407                 WARN_ON_ONCE(pos->children_weight);
1408                 pos->vfraction = 0;
1409
1410                 if (!parent)
1411                         break;
1412
1413                 propagate = !--parent->nr_active;
1414                 parent->children_weight -= pos->weight;
1415                 pos = parent;
1416         }
1417
1418         /* remove from the service tree */
1419         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1420                 cfq_rb_erase(&cfqg->rb_node, st);
1421 }
1422
1423 static void
1424 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1425 {
1426         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1427
1428         BUG_ON(cfqg->nr_cfqq < 1);
1429         cfqg->nr_cfqq--;
1430
1431         /* If there are other cfq queues under this group, don't delete it */
1432         if (cfqg->nr_cfqq)
1433                 return;
1434
1435         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1436         cfq_group_service_tree_del(st, cfqg);
1437         cfqg->saved_wl_slice = 0;
1438         cfqg_stats_update_dequeue(cfqg);
1439 }
1440
1441 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1442                                        u64 *unaccounted_time)
1443 {
1444         u64 slice_used;
1445         u64 now = ktime_get_ns();
1446
1447         /*
1448          * Queue got expired before even a single request completed or
1449          * got expired immediately after first request completion.
1450          */
1451         if (!cfqq->slice_start || cfqq->slice_start == now) {
1452                 /*
1453                  * Also charge the seek time incurred to the group, otherwise
1454                  * if there are mutiple queues in the group, each can dispatch
1455                  * a single request on seeky media and cause lots of seek time
1456                  * and group will never know it.
1457                  */
1458                 slice_used = max_t(u64, (now - cfqq->dispatch_start),
1459                                         jiffies_to_nsecs(1));
1460         } else {
1461                 slice_used = now - cfqq->slice_start;
1462                 if (slice_used > cfqq->allocated_slice) {
1463                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1464                         slice_used = cfqq->allocated_slice;
1465                 }
1466                 if (cfqq->slice_start > cfqq->dispatch_start)
1467                         *unaccounted_time += cfqq->slice_start -
1468                                         cfqq->dispatch_start;
1469         }
1470
1471         return slice_used;
1472 }
1473
1474 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1475                                 struct cfq_queue *cfqq)
1476 {
1477         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1478         u64 used_sl, charge, unaccounted_sl = 0;
1479         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1480                         - cfqg->service_tree_idle.count;
1481         unsigned int vfr;
1482         u64 now = ktime_get_ns();
1483
1484         BUG_ON(nr_sync < 0);
1485         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1486
1487         if (iops_mode(cfqd))
1488                 charge = cfqq->slice_dispatch;
1489         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1490                 charge = cfqq->allocated_slice;
1491
1492         /*
1493          * Can't update vdisktime while on service tree and cfqg->vfraction
1494          * is valid only while on it.  Cache vfr, leave the service tree,
1495          * update vdisktime and go back on.  The re-addition to the tree
1496          * will also update the weights as necessary.
1497          */
1498         vfr = cfqg->vfraction;
1499         cfq_group_service_tree_del(st, cfqg);
1500         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1501         cfq_group_service_tree_add(st, cfqg);
1502
1503         /* This group is being expired. Save the context */
1504         if (cfqd->workload_expires > now) {
1505                 cfqg->saved_wl_slice = cfqd->workload_expires - now;
1506                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1507                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1508         } else
1509                 cfqg->saved_wl_slice = 0;
1510
1511         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1512                                         st->min_vdisktime);
1513         cfq_log_cfqq(cfqq->cfqd, cfqq,
1514                      "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
1515                      used_sl, cfqq->slice_dispatch, charge,
1516                      iops_mode(cfqd), cfqq->nr_sectors);
1517         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1518         cfqg_stats_set_start_empty_time(cfqg);
1519 }
1520
1521 /**
1522  * cfq_init_cfqg_base - initialize base part of a cfq_group
1523  * @cfqg: cfq_group to initialize
1524  *
1525  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1526  * is enabled or not.
1527  */
1528 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1529 {
1530         struct cfq_rb_root *st;
1531         int i, j;
1532
1533         for_each_cfqg_st(cfqg, i, j, st)
1534                 *st = CFQ_RB_ROOT;
1535         RB_CLEAR_NODE(&cfqg->rb_node);
1536
1537         cfqg->ttime.last_end_request = ktime_get_ns();
1538 }
1539
1540 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1541 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1542                             bool on_dfl, bool reset_dev, bool is_leaf_weight);
1543
1544 static void cfqg_stats_exit(struct cfqg_stats *stats)
1545 {
1546         blkg_rwstat_exit(&stats->merged);
1547         blkg_rwstat_exit(&stats->service_time);
1548         blkg_rwstat_exit(&stats->wait_time);
1549         blkg_rwstat_exit(&stats->queued);
1550         blkg_stat_exit(&stats->time);
1551 #ifdef CONFIG_DEBUG_BLK_CGROUP
1552         blkg_stat_exit(&stats->unaccounted_time);
1553         blkg_stat_exit(&stats->avg_queue_size_sum);
1554         blkg_stat_exit(&stats->avg_queue_size_samples);
1555         blkg_stat_exit(&stats->dequeue);
1556         blkg_stat_exit(&stats->group_wait_time);
1557         blkg_stat_exit(&stats->idle_time);
1558         blkg_stat_exit(&stats->empty_time);
1559 #endif
1560 }
1561
1562 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1563 {
1564         if (blkg_rwstat_init(&stats->merged, gfp) ||
1565             blkg_rwstat_init(&stats->service_time, gfp) ||
1566             blkg_rwstat_init(&stats->wait_time, gfp) ||
1567             blkg_rwstat_init(&stats->queued, gfp) ||
1568             blkg_stat_init(&stats->time, gfp))
1569                 goto err;
1570
1571 #ifdef CONFIG_DEBUG_BLK_CGROUP
1572         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1573             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1574             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1575             blkg_stat_init(&stats->dequeue, gfp) ||
1576             blkg_stat_init(&stats->group_wait_time, gfp) ||
1577             blkg_stat_init(&stats->idle_time, gfp) ||
1578             blkg_stat_init(&stats->empty_time, gfp))
1579                 goto err;
1580 #endif
1581         return 0;
1582 err:
1583         cfqg_stats_exit(stats);
1584         return -ENOMEM;
1585 }
1586
1587 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1588 {
1589         struct cfq_group_data *cgd;
1590
1591         cgd = kzalloc(sizeof(*cgd), gfp);
1592         if (!cgd)
1593                 return NULL;
1594         return &cgd->cpd;
1595 }
1596
1597 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1598 {
1599         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1600         unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1601                               CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1602
1603         if (cpd_to_blkcg(cpd) == &blkcg_root)
1604                 weight *= 2;
1605
1606         cgd->weight = weight;
1607         cgd->leaf_weight = weight;
1608 }
1609
1610 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1611 {
1612         kfree(cpd_to_cfqgd(cpd));
1613 }
1614
1615 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1616 {
1617         struct blkcg *blkcg = cpd_to_blkcg(cpd);
1618         bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1619         unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1620
1621         if (blkcg == &blkcg_root)
1622                 weight *= 2;
1623
1624         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1625         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1626 }
1627
1628 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1629 {
1630         struct cfq_group *cfqg;
1631
1632         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1633         if (!cfqg)
1634                 return NULL;
1635
1636         cfq_init_cfqg_base(cfqg);
1637         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1638                 kfree(cfqg);
1639                 return NULL;
1640         }
1641
1642         return &cfqg->pd;
1643 }
1644
1645 static void cfq_pd_init(struct blkg_policy_data *pd)
1646 {
1647         struct cfq_group *cfqg = pd_to_cfqg(pd);
1648         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1649
1650         cfqg->weight = cgd->weight;
1651         cfqg->leaf_weight = cgd->leaf_weight;
1652 }
1653
1654 static void cfq_pd_offline(struct blkg_policy_data *pd)
1655 {
1656         struct cfq_group *cfqg = pd_to_cfqg(pd);
1657         int i;
1658
1659         for (i = 0; i < IOPRIO_BE_NR; i++) {
1660                 if (cfqg->async_cfqq[0][i])
1661                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1662                 if (cfqg->async_cfqq[1][i])
1663                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1664         }
1665
1666         if (cfqg->async_idle_cfqq)
1667                 cfq_put_queue(cfqg->async_idle_cfqq);
1668
1669         /*
1670          * @blkg is going offline and will be ignored by
1671          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1672          * that they don't get lost.  If IOs complete after this point, the
1673          * stats for them will be lost.  Oh well...
1674          */
1675         cfqg_stats_xfer_dead(cfqg);
1676 }
1677
1678 static void cfq_pd_free(struct blkg_policy_data *pd)
1679 {
1680         struct cfq_group *cfqg = pd_to_cfqg(pd);
1681
1682         cfqg_stats_exit(&cfqg->stats);
1683         return kfree(cfqg);
1684 }
1685
1686 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1687 {
1688         struct cfq_group *cfqg = pd_to_cfqg(pd);
1689
1690         cfqg_stats_reset(&cfqg->stats);
1691 }
1692
1693 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1694                                          struct blkcg *blkcg)
1695 {
1696         struct blkcg_gq *blkg;
1697
1698         blkg = blkg_lookup(blkcg, cfqd->queue);
1699         if (likely(blkg))
1700                 return blkg_to_cfqg(blkg);
1701         return NULL;
1702 }
1703
1704 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1705 {
1706         cfqq->cfqg = cfqg;
1707         /* cfqq reference on cfqg */
1708         cfqg_get(cfqg);
1709 }
1710
1711 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1712                                      struct blkg_policy_data *pd, int off)
1713 {
1714         struct cfq_group *cfqg = pd_to_cfqg(pd);
1715
1716         if (!cfqg->dev_weight)
1717                 return 0;
1718         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1719 }
1720
1721 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1722 {
1723         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1724                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1725                           0, false);
1726         return 0;
1727 }
1728
1729 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1730                                           struct blkg_policy_data *pd, int off)
1731 {
1732         struct cfq_group *cfqg = pd_to_cfqg(pd);
1733
1734         if (!cfqg->dev_leaf_weight)
1735                 return 0;
1736         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1737 }
1738
1739 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1740 {
1741         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1742                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1743                           0, false);
1744         return 0;
1745 }
1746
1747 static int cfq_print_weight(struct seq_file *sf, void *v)
1748 {
1749         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1750         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1751         unsigned int val = 0;
1752
1753         if (cgd)
1754                 val = cgd->weight;
1755
1756         seq_printf(sf, "%u\n", val);
1757         return 0;
1758 }
1759
1760 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1761 {
1762         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1763         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1764         unsigned int val = 0;
1765
1766         if (cgd)
1767                 val = cgd->leaf_weight;
1768
1769         seq_printf(sf, "%u\n", val);
1770         return 0;
1771 }
1772
1773 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1774                                         char *buf, size_t nbytes, loff_t off,
1775                                         bool on_dfl, bool is_leaf_weight)
1776 {
1777         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1778         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1779         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1780         struct blkg_conf_ctx ctx;
1781         struct cfq_group *cfqg;
1782         struct cfq_group_data *cfqgd;
1783         int ret;
1784         u64 v;
1785
1786         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1787         if (ret)
1788                 return ret;
1789
1790         if (sscanf(ctx.body, "%llu", &v) == 1) {
1791                 /* require "default" on dfl */
1792                 ret = -ERANGE;
1793                 if (!v && on_dfl)
1794                         goto out_finish;
1795         } else if (!strcmp(strim(ctx.body), "default")) {
1796                 v = 0;
1797         } else {
1798                 ret = -EINVAL;
1799                 goto out_finish;
1800         }
1801
1802         cfqg = blkg_to_cfqg(ctx.blkg);
1803         cfqgd = blkcg_to_cfqgd(blkcg);
1804
1805         ret = -ERANGE;
1806         if (!v || (v >= min && v <= max)) {
1807                 if (!is_leaf_weight) {
1808                         cfqg->dev_weight = v;
1809                         cfqg->new_weight = v ?: cfqgd->weight;
1810                 } else {
1811                         cfqg->dev_leaf_weight = v;
1812                         cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1813                 }
1814                 ret = 0;
1815         }
1816 out_finish:
1817         blkg_conf_finish(&ctx);
1818         return ret ?: nbytes;
1819 }
1820
1821 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1822                                       char *buf, size_t nbytes, loff_t off)
1823 {
1824         return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1825 }
1826
1827 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1828                                            char *buf, size_t nbytes, loff_t off)
1829 {
1830         return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1831 }
1832
1833 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1834                             bool on_dfl, bool reset_dev, bool is_leaf_weight)
1835 {
1836         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1837         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1838         struct blkcg *blkcg = css_to_blkcg(css);
1839         struct blkcg_gq *blkg;
1840         struct cfq_group_data *cfqgd;
1841         int ret = 0;
1842
1843         if (val < min || val > max)
1844                 return -ERANGE;
1845
1846         spin_lock_irq(&blkcg->lock);
1847         cfqgd = blkcg_to_cfqgd(blkcg);
1848         if (!cfqgd) {
1849                 ret = -EINVAL;
1850                 goto out;
1851         }
1852
1853         if (!is_leaf_weight)
1854                 cfqgd->weight = val;
1855         else
1856                 cfqgd->leaf_weight = val;
1857
1858         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1859                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1860
1861                 if (!cfqg)
1862                         continue;
1863
1864                 if (!is_leaf_weight) {
1865                         if (reset_dev)
1866                                 cfqg->dev_weight = 0;
1867                         if (!cfqg->dev_weight)
1868                                 cfqg->new_weight = cfqgd->weight;
1869                 } else {
1870                         if (reset_dev)
1871                                 cfqg->dev_leaf_weight = 0;
1872                         if (!cfqg->dev_leaf_weight)
1873                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1874                 }
1875         }
1876
1877 out:
1878         spin_unlock_irq(&blkcg->lock);
1879         return ret;
1880 }
1881
1882 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1883                           u64 val)
1884 {
1885         return __cfq_set_weight(css, val, false, false, false);
1886 }
1887
1888 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1889                                struct cftype *cft, u64 val)
1890 {
1891         return __cfq_set_weight(css, val, false, false, true);
1892 }
1893
1894 static int cfqg_print_stat(struct seq_file *sf, void *v)
1895 {
1896         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1897                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1898         return 0;
1899 }
1900
1901 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1902 {
1903         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1904                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1905         return 0;
1906 }
1907
1908 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1909                                       struct blkg_policy_data *pd, int off)
1910 {
1911         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1912                                           &blkcg_policy_cfq, off);
1913         return __blkg_prfill_u64(sf, pd, sum);
1914 }
1915
1916 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1917                                         struct blkg_policy_data *pd, int off)
1918 {
1919         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1920                                                         &blkcg_policy_cfq, off);
1921         return __blkg_prfill_rwstat(sf, pd, &sum);
1922 }
1923
1924 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1925 {
1926         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1927                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1928                           seq_cft(sf)->private, false);
1929         return 0;
1930 }
1931
1932 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1933 {
1934         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1935                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1936                           seq_cft(sf)->private, true);
1937         return 0;
1938 }
1939
1940 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1941                                int off)
1942 {
1943         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1944
1945         return __blkg_prfill_u64(sf, pd, sum >> 9);
1946 }
1947
1948 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1949 {
1950         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1951                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1952         return 0;
1953 }
1954
1955 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1956                                          struct blkg_policy_data *pd, int off)
1957 {
1958         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1959                                         offsetof(struct blkcg_gq, stat_bytes));
1960         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1961                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1962
1963         return __blkg_prfill_u64(sf, pd, sum >> 9);
1964 }
1965
1966 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1967 {
1968         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1969                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1970                           false);
1971         return 0;
1972 }
1973
1974 #ifdef CONFIG_DEBUG_BLK_CGROUP
1975 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1976                                       struct blkg_policy_data *pd, int off)
1977 {
1978         struct cfq_group *cfqg = pd_to_cfqg(pd);
1979         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1980         u64 v = 0;
1981
1982         if (samples) {
1983                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1984                 v = div64_u64(v, samples);
1985         }
1986         __blkg_prfill_u64(sf, pd, v);
1987         return 0;
1988 }
1989
1990 /* print avg_queue_size */
1991 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1992 {
1993         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1994                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1995                           0, false);
1996         return 0;
1997 }
1998 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1999
2000 static struct cftype cfq_blkcg_legacy_files[] = {
2001         /* on root, weight is mapped to leaf_weight */
2002         {
2003                 .name = "weight_device",
2004                 .flags = CFTYPE_ONLY_ON_ROOT,
2005                 .seq_show = cfqg_print_leaf_weight_device,
2006                 .write = cfqg_set_leaf_weight_device,
2007         },
2008         {
2009                 .name = "weight",
2010                 .flags = CFTYPE_ONLY_ON_ROOT,
2011                 .seq_show = cfq_print_leaf_weight,
2012                 .write_u64 = cfq_set_leaf_weight,
2013         },
2014
2015         /* no such mapping necessary for !roots */
2016         {
2017                 .name = "weight_device",
2018                 .flags = CFTYPE_NOT_ON_ROOT,
2019                 .seq_show = cfqg_print_weight_device,
2020                 .write = cfqg_set_weight_device,
2021         },
2022         {
2023                 .name = "weight",
2024                 .flags = CFTYPE_NOT_ON_ROOT,
2025                 .seq_show = cfq_print_weight,
2026                 .write_u64 = cfq_set_weight,
2027         },
2028
2029         {
2030                 .name = "leaf_weight_device",
2031                 .seq_show = cfqg_print_leaf_weight_device,
2032                 .write = cfqg_set_leaf_weight_device,
2033         },
2034         {
2035                 .name = "leaf_weight",
2036                 .seq_show = cfq_print_leaf_weight,
2037                 .write_u64 = cfq_set_leaf_weight,
2038         },
2039
2040         /* statistics, covers only the tasks in the cfqg */
2041         {
2042                 .name = "time",
2043                 .private = offsetof(struct cfq_group, stats.time),
2044                 .seq_show = cfqg_print_stat,
2045         },
2046         {
2047                 .name = "sectors",
2048                 .seq_show = cfqg_print_stat_sectors,
2049         },
2050         {
2051                 .name = "io_service_bytes",
2052                 .private = (unsigned long)&blkcg_policy_cfq,
2053                 .seq_show = blkg_print_stat_bytes,
2054         },
2055         {
2056                 .name = "io_serviced",
2057                 .private = (unsigned long)&blkcg_policy_cfq,
2058                 .seq_show = blkg_print_stat_ios,
2059         },
2060         {
2061                 .name = "io_service_time",
2062                 .private = offsetof(struct cfq_group, stats.service_time),
2063                 .seq_show = cfqg_print_rwstat,
2064         },
2065         {
2066                 .name = "io_wait_time",
2067                 .private = offsetof(struct cfq_group, stats.wait_time),
2068                 .seq_show = cfqg_print_rwstat,
2069         },
2070         {
2071                 .name = "io_merged",
2072                 .private = offsetof(struct cfq_group, stats.merged),
2073                 .seq_show = cfqg_print_rwstat,
2074         },
2075         {
2076                 .name = "io_queued",
2077                 .private = offsetof(struct cfq_group, stats.queued),
2078                 .seq_show = cfqg_print_rwstat,
2079         },
2080
2081         /* the same statictics which cover the cfqg and its descendants */
2082         {
2083                 .name = "time_recursive",
2084                 .private = offsetof(struct cfq_group, stats.time),
2085                 .seq_show = cfqg_print_stat_recursive,
2086         },
2087         {
2088                 .name = "sectors_recursive",
2089                 .seq_show = cfqg_print_stat_sectors_recursive,
2090         },
2091         {
2092                 .name = "io_service_bytes_recursive",
2093                 .private = (unsigned long)&blkcg_policy_cfq,
2094                 .seq_show = blkg_print_stat_bytes_recursive,
2095         },
2096         {
2097                 .name = "io_serviced_recursive",
2098                 .private = (unsigned long)&blkcg_policy_cfq,
2099                 .seq_show = blkg_print_stat_ios_recursive,
2100         },
2101         {
2102                 .name = "io_service_time_recursive",
2103                 .private = offsetof(struct cfq_group, stats.service_time),
2104                 .seq_show = cfqg_print_rwstat_recursive,
2105         },
2106         {
2107                 .name = "io_wait_time_recursive",
2108                 .private = offsetof(struct cfq_group, stats.wait_time),
2109                 .seq_show = cfqg_print_rwstat_recursive,
2110         },
2111         {
2112                 .name = "io_merged_recursive",
2113                 .private = offsetof(struct cfq_group, stats.merged),
2114                 .seq_show = cfqg_print_rwstat_recursive,
2115         },
2116         {
2117                 .name = "io_queued_recursive",
2118                 .private = offsetof(struct cfq_group, stats.queued),
2119                 .seq_show = cfqg_print_rwstat_recursive,
2120         },
2121 #ifdef CONFIG_DEBUG_BLK_CGROUP
2122         {
2123                 .name = "avg_queue_size",
2124                 .seq_show = cfqg_print_avg_queue_size,
2125         },
2126         {
2127                 .name = "group_wait_time",
2128                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2129                 .seq_show = cfqg_print_stat,
2130         },
2131         {
2132                 .name = "idle_time",
2133                 .private = offsetof(struct cfq_group, stats.idle_time),
2134                 .seq_show = cfqg_print_stat,
2135         },
2136         {
2137                 .name = "empty_time",
2138                 .private = offsetof(struct cfq_group, stats.empty_time),
2139                 .seq_show = cfqg_print_stat,
2140         },
2141         {
2142                 .name = "dequeue",
2143                 .private = offsetof(struct cfq_group, stats.dequeue),
2144                 .seq_show = cfqg_print_stat,
2145         },
2146         {
2147                 .name = "unaccounted_time",
2148                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2149                 .seq_show = cfqg_print_stat,
2150         },
2151 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2152         { }     /* terminate */
2153 };
2154
2155 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2156 {
2157         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2158         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2159
2160         seq_printf(sf, "default %u\n", cgd->weight);
2161         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2162                           &blkcg_policy_cfq, 0, false);
2163         return 0;
2164 }
2165
2166 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2167                                      char *buf, size_t nbytes, loff_t off)
2168 {
2169         char *endp;
2170         int ret;
2171         u64 v;
2172
2173         buf = strim(buf);
2174
2175         /* "WEIGHT" or "default WEIGHT" sets the default weight */
2176         v = simple_strtoull(buf, &endp, 0);
2177         if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2178                 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2179                 return ret ?: nbytes;
2180         }
2181
2182         /* "MAJ:MIN WEIGHT" */
2183         return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2184 }
2185
2186 static struct cftype cfq_blkcg_files[] = {
2187         {
2188                 .name = "weight",
2189                 .flags = CFTYPE_NOT_ON_ROOT,
2190                 .seq_show = cfq_print_weight_on_dfl,
2191                 .write = cfq_set_weight_on_dfl,
2192         },
2193         { }     /* terminate */
2194 };
2195
2196 #else /* GROUP_IOSCHED */
2197 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2198                                          struct blkcg *blkcg)
2199 {
2200         return cfqd->root_group;
2201 }
2202
2203 static inline void
2204 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2205         cfqq->cfqg = cfqg;
2206 }
2207
2208 #endif /* GROUP_IOSCHED */
2209
2210 /*
2211  * The cfqd->service_trees holds all pending cfq_queue's that have
2212  * requests waiting to be processed. It is sorted in the order that
2213  * we will service the queues.
2214  */
2215 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2216                                  bool add_front)
2217 {
2218         struct rb_node **p, *parent;
2219         struct cfq_queue *__cfqq;
2220         u64 rb_key;
2221         struct cfq_rb_root *st;
2222         int left;
2223         int new_cfqq = 1;
2224         u64 now = ktime_get_ns();
2225
2226         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2227         if (cfq_class_idle(cfqq)) {
2228                 rb_key = CFQ_IDLE_DELAY;
2229                 parent = rb_last(&st->rb);
2230                 if (parent && parent != &cfqq->rb_node) {
2231                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2232                         rb_key += __cfqq->rb_key;
2233                 } else
2234                         rb_key += now;
2235         } else if (!add_front) {
2236                 /*
2237                  * Get our rb key offset. Subtract any residual slice
2238                  * value carried from last service. A negative resid
2239                  * count indicates slice overrun, and this should position
2240                  * the next service time further away in the tree.
2241                  */
2242                 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2243                 rb_key -= cfqq->slice_resid;
2244                 cfqq->slice_resid = 0;
2245         } else {
2246                 rb_key = -NSEC_PER_SEC;
2247                 __cfqq = cfq_rb_first(st);
2248                 rb_key += __cfqq ? __cfqq->rb_key : now;
2249         }
2250
2251         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2252                 new_cfqq = 0;
2253                 /*
2254                  * same position, nothing more to do
2255                  */
2256                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2257                         return;
2258
2259                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2260                 cfqq->service_tree = NULL;
2261         }
2262
2263         left = 1;
2264         parent = NULL;
2265         cfqq->service_tree = st;
2266         p = &st->rb.rb_node;
2267         while (*p) {
2268                 parent = *p;
2269                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2270
2271                 /*
2272                  * sort by key, that represents service time.
2273                  */
2274                 if (rb_key < __cfqq->rb_key)
2275                         p = &parent->rb_left;
2276                 else {
2277                         p = &parent->rb_right;
2278                         left = 0;
2279                 }
2280         }
2281
2282         if (left)
2283                 st->left = &cfqq->rb_node;
2284
2285         cfqq->rb_key = rb_key;
2286         rb_link_node(&cfqq->rb_node, parent, p);
2287         rb_insert_color(&cfqq->rb_node, &st->rb);
2288         st->count++;
2289         if (add_front || !new_cfqq)
2290                 return;
2291         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2292 }
2293
2294 static struct cfq_queue *
2295 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2296                      sector_t sector, struct rb_node **ret_parent,
2297                      struct rb_node ***rb_link)
2298 {
2299         struct rb_node **p, *parent;
2300         struct cfq_queue *cfqq = NULL;
2301
2302         parent = NULL;
2303         p = &root->rb_node;
2304         while (*p) {
2305                 struct rb_node **n;
2306
2307                 parent = *p;
2308                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2309
2310                 /*
2311                  * Sort strictly based on sector.  Smallest to the left,
2312                  * largest to the right.
2313                  */
2314                 if (sector > blk_rq_pos(cfqq->next_rq))
2315                         n = &(*p)->rb_right;
2316                 else if (sector < blk_rq_pos(cfqq->next_rq))
2317                         n = &(*p)->rb_left;
2318                 else
2319                         break;
2320                 p = n;
2321                 cfqq = NULL;
2322         }
2323
2324         *ret_parent = parent;
2325         if (rb_link)
2326                 *rb_link = p;
2327         return cfqq;
2328 }
2329
2330 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2331 {
2332         struct rb_node **p, *parent;
2333         struct cfq_queue *__cfqq;
2334
2335         if (cfqq->p_root) {
2336                 rb_erase(&cfqq->p_node, cfqq->p_root);
2337                 cfqq->p_root = NULL;
2338         }
2339
2340         if (cfq_class_idle(cfqq))
2341                 return;
2342         if (!cfqq->next_rq)
2343                 return;
2344
2345         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2346         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2347                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2348         if (!__cfqq) {
2349                 rb_link_node(&cfqq->p_node, parent, p);
2350                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2351         } else
2352                 cfqq->p_root = NULL;
2353 }
2354
2355 /*
2356  * Update cfqq's position in the service tree.
2357  */
2358 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2359 {
2360         /*
2361          * Resorting requires the cfqq to be on the RR list already.
2362          */
2363         if (cfq_cfqq_on_rr(cfqq)) {
2364                 cfq_service_tree_add(cfqd, cfqq, 0);
2365                 cfq_prio_tree_add(cfqd, cfqq);
2366         }
2367 }
2368
2369 /*
2370  * add to busy list of queues for service, trying to be fair in ordering
2371  * the pending list according to last request service
2372  */
2373 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2374 {
2375         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2376         BUG_ON(cfq_cfqq_on_rr(cfqq));
2377         cfq_mark_cfqq_on_rr(cfqq);
2378         cfqd->busy_queues++;
2379         if (cfq_cfqq_sync(cfqq))
2380                 cfqd->busy_sync_queues++;
2381
2382         cfq_resort_rr_list(cfqd, cfqq);
2383 }
2384
2385 /*
2386  * Called when the cfqq no longer has requests pending, remove it from
2387  * the service tree.
2388  */
2389 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2390 {
2391         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2392         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2393         cfq_clear_cfqq_on_rr(cfqq);
2394
2395         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2396                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2397                 cfqq->service_tree = NULL;
2398         }
2399         if (cfqq->p_root) {
2400                 rb_erase(&cfqq->p_node, cfqq->p_root);
2401                 cfqq->p_root = NULL;
2402         }
2403
2404         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2405         BUG_ON(!cfqd->busy_queues);
2406         cfqd->busy_queues--;
2407         if (cfq_cfqq_sync(cfqq))
2408                 cfqd->busy_sync_queues--;
2409 }
2410
2411 /*
2412  * rb tree support functions
2413  */
2414 static void cfq_del_rq_rb(struct request *rq)
2415 {
2416         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2417         const int sync = rq_is_sync(rq);
2418
2419         BUG_ON(!cfqq->queued[sync]);
2420         cfqq->queued[sync]--;
2421
2422         elv_rb_del(&cfqq->sort_list, rq);
2423
2424         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2425                 /*
2426                  * Queue will be deleted from service tree when we actually
2427                  * expire it later. Right now just remove it from prio tree
2428                  * as it is empty.
2429                  */
2430                 if (cfqq->p_root) {
2431                         rb_erase(&cfqq->p_node, cfqq->p_root);
2432                         cfqq->p_root = NULL;
2433                 }
2434         }
2435 }
2436
2437 static void cfq_add_rq_rb(struct request *rq)
2438 {
2439         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2440         struct cfq_data *cfqd = cfqq->cfqd;
2441         struct request *prev;
2442
2443         cfqq->queued[rq_is_sync(rq)]++;
2444
2445         elv_rb_add(&cfqq->sort_list, rq);
2446
2447         if (!cfq_cfqq_on_rr(cfqq))
2448                 cfq_add_cfqq_rr(cfqd, cfqq);
2449
2450         /*
2451          * check if this request is a better next-serve candidate
2452          */
2453         prev = cfqq->next_rq;
2454         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2455
2456         /*
2457          * adjust priority tree position, if ->next_rq changes
2458          */
2459         if (prev != cfqq->next_rq)
2460                 cfq_prio_tree_add(cfqd, cfqq);
2461
2462         BUG_ON(!cfqq->next_rq);
2463 }
2464
2465 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2466 {
2467         elv_rb_del(&cfqq->sort_list, rq);
2468         cfqq->queued[rq_is_sync(rq)]--;
2469         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2470         cfq_add_rq_rb(rq);
2471         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2472                                  rq->cmd_flags);
2473 }
2474
2475 static struct request *
2476 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2477 {
2478         struct task_struct *tsk = current;
2479         struct cfq_io_cq *cic;
2480         struct cfq_queue *cfqq;
2481
2482         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2483         if (!cic)
2484                 return NULL;
2485
2486         cfqq = cic_to_cfqq(cic, op_is_sync(bio->bi_opf));
2487         if (cfqq)
2488                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2489
2490         return NULL;
2491 }
2492
2493 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2494 {
2495         struct cfq_data *cfqd = q->elevator->elevator_data;
2496
2497         cfqd->rq_in_driver++;
2498         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2499                                                 cfqd->rq_in_driver);
2500
2501         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2502 }
2503
2504 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2505 {
2506         struct cfq_data *cfqd = q->elevator->elevator_data;
2507
2508         WARN_ON(!cfqd->rq_in_driver);
2509         cfqd->rq_in_driver--;
2510         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2511                                                 cfqd->rq_in_driver);
2512 }
2513
2514 static void cfq_remove_request(struct request *rq)
2515 {
2516         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2517
2518         if (cfqq->next_rq == rq)
2519                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2520
2521         list_del_init(&rq->queuelist);
2522         cfq_del_rq_rb(rq);
2523
2524         cfqq->cfqd->rq_queued--;
2525         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2526         if (rq->cmd_flags & REQ_PRIO) {
2527                 WARN_ON(!cfqq->prio_pending);
2528                 cfqq->prio_pending--;
2529         }
2530 }
2531
2532 static enum elv_merge cfq_merge(struct request_queue *q, struct request **req,
2533                      struct bio *bio)
2534 {
2535         struct cfq_data *cfqd = q->elevator->elevator_data;
2536         struct request *__rq;
2537
2538         __rq = cfq_find_rq_fmerge(cfqd, bio);
2539         if (__rq && elv_bio_merge_ok(__rq, bio)) {
2540                 *req = __rq;
2541                 return ELEVATOR_FRONT_MERGE;
2542         }
2543
2544         return ELEVATOR_NO_MERGE;
2545 }
2546
2547 static void cfq_merged_request(struct request_queue *q, struct request *req,
2548                                enum elv_merge type)
2549 {
2550         if (type == ELEVATOR_FRONT_MERGE) {
2551                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2552
2553                 cfq_reposition_rq_rb(cfqq, req);
2554         }
2555 }
2556
2557 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2558                                 struct bio *bio)
2559 {
2560         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_opf);
2561 }
2562
2563 static void
2564 cfq_merged_requests(struct request_queue *q, struct request *rq,
2565                     struct request *next)
2566 {
2567         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2568         struct cfq_data *cfqd = q->elevator->elevator_data;
2569
2570         /*
2571          * reposition in fifo if next is older than rq
2572          */
2573         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2574             next->fifo_time < rq->fifo_time &&
2575             cfqq == RQ_CFQQ(next)) {
2576                 list_move(&rq->queuelist, &next->queuelist);
2577                 rq->fifo_time = next->fifo_time;
2578         }
2579
2580         if (cfqq->next_rq == next)
2581                 cfqq->next_rq = rq;
2582         cfq_remove_request(next);
2583         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2584
2585         cfqq = RQ_CFQQ(next);
2586         /*
2587          * all requests of this queue are merged to other queues, delete it
2588          * from the service tree. If it's the active_queue,
2589          * cfq_dispatch_requests() will choose to expire it or do idle
2590          */
2591         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2592             cfqq != cfqd->active_queue)
2593                 cfq_del_cfqq_rr(cfqd, cfqq);
2594 }
2595
2596 static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2597                                struct bio *bio)
2598 {
2599         struct cfq_data *cfqd = q->elevator->elevator_data;
2600         bool is_sync = op_is_sync(bio->bi_opf);
2601         struct cfq_io_cq *cic;
2602         struct cfq_queue *cfqq;
2603
2604         /*
2605          * Disallow merge of a sync bio into an async request.
2606          */
2607         if (is_sync && !rq_is_sync(rq))
2608                 return false;
2609
2610         /*
2611          * Lookup the cfqq that this bio will be queued with and allow
2612          * merge only if rq is queued there.
2613          */
2614         cic = cfq_cic_lookup(cfqd, current->io_context);
2615         if (!cic)
2616                 return false;
2617
2618         cfqq = cic_to_cfqq(cic, is_sync);
2619         return cfqq == RQ_CFQQ(rq);
2620 }
2621
2622 static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2623                               struct request *next)
2624 {
2625         return RQ_CFQQ(rq) == RQ_CFQQ(next);
2626 }
2627
2628 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2629 {
2630         hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2631         cfqg_stats_update_idle_time(cfqq->cfqg);
2632 }
2633
2634 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2635                                    struct cfq_queue *cfqq)
2636 {
2637         if (cfqq) {
2638                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2639                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2640                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2641                 cfqq->slice_start = 0;
2642                 cfqq->dispatch_start = ktime_get_ns();
2643                 cfqq->allocated_slice = 0;
2644                 cfqq->slice_end = 0;
2645                 cfqq->slice_dispatch = 0;
2646                 cfqq->nr_sectors = 0;
2647
2648                 cfq_clear_cfqq_wait_request(cfqq);
2649                 cfq_clear_cfqq_must_dispatch(cfqq);
2650                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2651                 cfq_clear_cfqq_fifo_expire(cfqq);
2652                 cfq_mark_cfqq_slice_new(cfqq);
2653
2654                 cfq_del_timer(cfqd, cfqq);
2655         }
2656
2657         cfqd->active_queue = cfqq;
2658 }
2659
2660 /*
2661  * current cfqq expired its slice (or was too idle), select new one
2662  */
2663 static void
2664 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2665                     bool timed_out)
2666 {
2667         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2668
2669         if (cfq_cfqq_wait_request(cfqq))
2670                 cfq_del_timer(cfqd, cfqq);
2671
2672         cfq_clear_cfqq_wait_request(cfqq);
2673         cfq_clear_cfqq_wait_busy(cfqq);
2674
2675         /*
2676          * If this cfqq is shared between multiple processes, check to
2677          * make sure that those processes are still issuing I/Os within
2678          * the mean seek distance.  If not, it may be time to break the
2679          * queues apart again.
2680          */
2681         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2682                 cfq_mark_cfqq_split_coop(cfqq);
2683
2684         /*
2685          * store what was left of this slice, if the queue idled/timed out
2686          */
2687         if (timed_out) {
2688                 if (cfq_cfqq_slice_new(cfqq))
2689                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2690                 else
2691                         cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2692                 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2693         }
2694
2695         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2696
2697         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2698                 cfq_del_cfqq_rr(cfqd, cfqq);
2699
2700         cfq_resort_rr_list(cfqd, cfqq);
2701
2702         if (cfqq == cfqd->active_queue)
2703                 cfqd->active_queue = NULL;
2704
2705         if (cfqd->active_cic) {
2706                 put_io_context(cfqd->active_cic->icq.ioc);
2707                 cfqd->active_cic = NULL;
2708         }
2709 }
2710
2711 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2712 {
2713         struct cfq_queue *cfqq = cfqd->active_queue;
2714
2715         if (cfqq)
2716                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2717 }
2718
2719 /*
2720  * Get next queue for service. Unless we have a queue preemption,
2721  * we'll simply select the first cfqq in the service tree.
2722  */
2723 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2724 {
2725         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2726                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2727
2728         if (!cfqd->rq_queued)
2729                 return NULL;
2730
2731         /* There is nothing to dispatch */
2732         if (!st)
2733                 return NULL;
2734         if (RB_EMPTY_ROOT(&st->rb))
2735                 return NULL;
2736         return cfq_rb_first(st);
2737 }
2738
2739 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2740 {
2741         struct cfq_group *cfqg;
2742         struct cfq_queue *cfqq;
2743         int i, j;
2744         struct cfq_rb_root *st;
2745
2746         if (!cfqd->rq_queued)
2747                 return NULL;
2748
2749         cfqg = cfq_get_next_cfqg(cfqd);
2750         if (!cfqg)
2751                 return NULL;
2752
2753         for_each_cfqg_st(cfqg, i, j, st) {
2754                 cfqq = cfq_rb_first(st);
2755                 if (cfqq)
2756                         return cfqq;
2757         }
2758         return NULL;
2759 }
2760
2761 /*
2762  * Get and set a new active queue for service.
2763  */
2764 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2765                                               struct cfq_queue *cfqq)
2766 {
2767         if (!cfqq)
2768                 cfqq = cfq_get_next_queue(cfqd);
2769
2770         __cfq_set_active_queue(cfqd, cfqq);
2771         return cfqq;
2772 }
2773
2774 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2775                                           struct request *rq)
2776 {
2777         if (blk_rq_pos(rq) >= cfqd->last_position)
2778                 return blk_rq_pos(rq) - cfqd->last_position;
2779         else
2780                 return cfqd->last_position - blk_rq_pos(rq);
2781 }
2782
2783 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2784                                struct request *rq)
2785 {
2786         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2787 }
2788
2789 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2790                                     struct cfq_queue *cur_cfqq)
2791 {
2792         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2793         struct rb_node *parent, *node;
2794         struct cfq_queue *__cfqq;
2795         sector_t sector = cfqd->last_position;
2796
2797         if (RB_EMPTY_ROOT(root))
2798                 return NULL;
2799
2800         /*
2801          * First, if we find a request starting at the end of the last
2802          * request, choose it.
2803          */
2804         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2805         if (__cfqq)
2806                 return __cfqq;
2807
2808         /*
2809          * If the exact sector wasn't found, the parent of the NULL leaf
2810          * will contain the closest sector.
2811          */
2812         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2813         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2814                 return __cfqq;
2815
2816         if (blk_rq_pos(__cfqq->next_rq) < sector)
2817                 node = rb_next(&__cfqq->p_node);
2818         else
2819                 node = rb_prev(&__cfqq->p_node);
2820         if (!node)
2821                 return NULL;
2822
2823         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2824         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2825                 return __cfqq;
2826
2827         return NULL;
2828 }
2829
2830 /*
2831  * cfqd - obvious
2832  * cur_cfqq - passed in so that we don't decide that the current queue is
2833  *            closely cooperating with itself.
2834  *
2835  * So, basically we're assuming that that cur_cfqq has dispatched at least
2836  * one request, and that cfqd->last_position reflects a position on the disk
2837  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2838  * assumption.
2839  */
2840 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2841                                               struct cfq_queue *cur_cfqq)
2842 {
2843         struct cfq_queue *cfqq;
2844
2845         if (cfq_class_idle(cur_cfqq))
2846                 return NULL;
2847         if (!cfq_cfqq_sync(cur_cfqq))
2848                 return NULL;
2849         if (CFQQ_SEEKY(cur_cfqq))
2850                 return NULL;
2851
2852         /*
2853          * Don't search priority tree if it's the only queue in the group.
2854          */
2855         if (cur_cfqq->cfqg->nr_cfqq == 1)
2856                 return NULL;
2857
2858         /*
2859          * We should notice if some of the queues are cooperating, eg
2860          * working closely on the same area of the disk. In that case,
2861          * we can group them together and don't waste time idling.
2862          */
2863         cfqq = cfqq_close(cfqd, cur_cfqq);
2864         if (!cfqq)
2865                 return NULL;
2866
2867         /* If new queue belongs to different cfq_group, don't choose it */
2868         if (cur_cfqq->cfqg != cfqq->cfqg)
2869                 return NULL;
2870
2871         /*
2872          * It only makes sense to merge sync queues.
2873          */
2874         if (!cfq_cfqq_sync(cfqq))
2875                 return NULL;
2876         if (CFQQ_SEEKY(cfqq))
2877                 return NULL;
2878
2879         /*
2880          * Do not merge queues of different priority classes
2881          */
2882         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2883                 return NULL;
2884
2885         return cfqq;
2886 }
2887
2888 /*
2889  * Determine whether we should enforce idle window for this queue.
2890  */
2891
2892 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2893 {
2894         enum wl_class_t wl_class = cfqq_class(cfqq);
2895         struct cfq_rb_root *st = cfqq->service_tree;
2896
2897         BUG_ON(!st);
2898         BUG_ON(!st->count);
2899
2900         if (!cfqd->cfq_slice_idle)
2901                 return false;
2902
2903         /* We never do for idle class queues. */
2904         if (wl_class == IDLE_WORKLOAD)
2905                 return false;
2906
2907         /* We do for queues that were marked with idle window flag. */
2908         if (cfq_cfqq_idle_window(cfqq) &&
2909            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2910                 return true;
2911
2912         /*
2913          * Otherwise, we do only if they are the last ones
2914          * in their service tree.
2915          */
2916         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2917            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2918                 return true;
2919         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2920         return false;
2921 }
2922
2923 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2924 {
2925         struct cfq_queue *cfqq = cfqd->active_queue;
2926         struct cfq_rb_root *st = cfqq->service_tree;
2927         struct cfq_io_cq *cic;
2928         u64 sl, group_idle = 0;
2929         u64 now = ktime_get_ns();
2930
2931         /*
2932          * SSD device without seek penalty, disable idling. But only do so
2933          * for devices that support queuing, otherwise we still have a problem
2934          * with sync vs async workloads.
2935          */
2936         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2937                 return;
2938
2939         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2940         WARN_ON(cfq_cfqq_slice_new(cfqq));
2941
2942         /*
2943          * idle is disabled, either manually or by past process history
2944          */
2945         if (!cfq_should_idle(cfqd, cfqq)) {
2946                 /* no queue idling. Check for group idling */
2947                 if (cfqd->cfq_group_idle)
2948                         group_idle = cfqd->cfq_group_idle;
2949                 else
2950                         return;
2951         }
2952
2953         /*
2954          * still active requests from this queue, don't idle
2955          */
2956         if (cfqq->dispatched)
2957                 return;
2958
2959         /*
2960          * task has exited, don't wait
2961          */
2962         cic = cfqd->active_cic;
2963         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2964                 return;
2965
2966         /*
2967          * If our average think time is larger than the remaining time
2968          * slice, then don't idle. This avoids overrunning the allotted
2969          * time slice.
2970          */
2971         if (sample_valid(cic->ttime.ttime_samples) &&
2972             (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2973                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2974                              cic->ttime.ttime_mean);
2975                 return;
2976         }
2977
2978         /*
2979          * There are other queues in the group or this is the only group and
2980          * it has too big thinktime, don't do group idle.
2981          */
2982         if (group_idle &&
2983             (cfqq->cfqg->nr_cfqq > 1 ||
2984              cfq_io_thinktime_big(cfqd, &st->ttime, true)))
2985                 return;
2986
2987         cfq_mark_cfqq_wait_request(cfqq);
2988
2989         if (group_idle)
2990                 sl = cfqd->cfq_group_idle;
2991         else
2992                 sl = cfqd->cfq_slice_idle;
2993
2994         hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
2995                       HRTIMER_MODE_REL);
2996         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2997         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
2998                         group_idle ? 1 : 0);
2999 }
3000
3001 /*
3002  * Move request from internal lists to the request queue dispatch list.
3003  */
3004 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
3005 {
3006         struct cfq_data *cfqd = q->elevator->elevator_data;
3007         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3008
3009         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
3010
3011         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
3012         cfq_remove_request(rq);
3013         cfqq->dispatched++;
3014         (RQ_CFQG(rq))->dispatched++;
3015         elv_dispatch_sort(q, rq);
3016
3017         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
3018         cfqq->nr_sectors += blk_rq_sectors(rq);
3019 }
3020
3021 /*
3022  * return expired entry, or NULL to just start from scratch in rbtree
3023  */
3024 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3025 {
3026         struct request *rq = NULL;
3027
3028         if (cfq_cfqq_fifo_expire(cfqq))
3029                 return NULL;
3030
3031         cfq_mark_cfqq_fifo_expire(cfqq);
3032
3033         if (list_empty(&cfqq->fifo))
3034                 return NULL;
3035
3036         rq = rq_entry_fifo(cfqq->fifo.next);
3037         if (ktime_get_ns() < rq->fifo_time)
3038                 rq = NULL;
3039
3040         return rq;
3041 }
3042
3043 static inline int
3044 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3045 {
3046         const int base_rq = cfqd->cfq_slice_async_rq;
3047
3048         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3049
3050         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3051 }
3052
3053 /*
3054  * Must be called with the queue_lock held.
3055  */
3056 static int cfqq_process_refs(struct cfq_queue *cfqq)
3057 {
3058         int process_refs, io_refs;
3059
3060         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3061         process_refs = cfqq->ref - io_refs;
3062         BUG_ON(process_refs < 0);
3063         return process_refs;
3064 }
3065
3066 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3067 {
3068         int process_refs, new_process_refs;
3069         struct cfq_queue *__cfqq;
3070
3071         /*
3072          * If there are no process references on the new_cfqq, then it is
3073          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3074          * chain may have dropped their last reference (not just their
3075          * last process reference).
3076          */
3077         if (!cfqq_process_refs(new_cfqq))
3078                 return;
3079
3080         /* Avoid a circular list and skip interim queue merges */
3081         while ((__cfqq = new_cfqq->new_cfqq)) {
3082                 if (__cfqq == cfqq)
3083                         return;
3084                 new_cfqq = __cfqq;
3085         }
3086
3087         process_refs = cfqq_process_refs(cfqq);
3088         new_process_refs = cfqq_process_refs(new_cfqq);
3089         /*
3090          * If the process for the cfqq has gone away, there is no
3091          * sense in merging the queues.
3092          */
3093         if (process_refs == 0 || new_process_refs == 0)
3094                 return;
3095
3096         /*
3097          * Merge in the direction of the lesser amount of work.
3098          */
3099         if (new_process_refs >= process_refs) {
3100                 cfqq->new_cfqq = new_cfqq;
3101                 new_cfqq->ref += process_refs;
3102         } else {
3103                 new_cfqq->new_cfqq = cfqq;
3104                 cfqq->ref += new_process_refs;
3105         }
3106 }
3107
3108 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3109                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3110 {
3111         struct cfq_queue *queue;
3112         int i;
3113         bool key_valid = false;
3114         u64 lowest_key = 0;
3115         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3116
3117         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3118                 /* select the one with lowest rb_key */
3119                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3120                 if (queue &&
3121                     (!key_valid || queue->rb_key < lowest_key)) {
3122                         lowest_key = queue->rb_key;
3123                         cur_best = i;
3124                         key_valid = true;
3125                 }
3126         }
3127
3128         return cur_best;
3129 }
3130
3131 static void
3132 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3133 {
3134         u64 slice;
3135         unsigned count;
3136         struct cfq_rb_root *st;
3137         u64 group_slice;
3138         enum wl_class_t original_class = cfqd->serving_wl_class;
3139         u64 now = ktime_get_ns();
3140
3141         /* Choose next priority. RT > BE > IDLE */
3142         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3143                 cfqd->serving_wl_class = RT_WORKLOAD;
3144         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3145                 cfqd->serving_wl_class = BE_WORKLOAD;
3146         else {
3147                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3148                 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3149                 return;
3150         }
3151
3152         if (original_class != cfqd->serving_wl_class)
3153                 goto new_workload;
3154
3155         /*
3156          * For RT and BE, we have to choose also the type
3157          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3158          * expiration time
3159          */
3160         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3161         count = st->count;
3162
3163         /*
3164          * check workload expiration, and that we still have other queues ready
3165          */
3166         if (count && !(now > cfqd->workload_expires))
3167                 return;
3168
3169 new_workload:
3170         /* otherwise select new workload type */
3171         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3172                                         cfqd->serving_wl_class);
3173         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3174         count = st->count;
3175
3176         /*
3177          * the workload slice is computed as a fraction of target latency
3178          * proportional to the number of queues in that workload, over
3179          * all the queues in the same priority class
3180          */
3181         group_slice = cfq_group_slice(cfqd, cfqg);
3182
3183         slice = div_u64(group_slice * count,
3184                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3185                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3186                                         cfqg)));
3187
3188         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3189                 u64 tmp;
3190
3191                 /*
3192                  * Async queues are currently system wide. Just taking
3193                  * proportion of queues with-in same group will lead to higher
3194                  * async ratio system wide as generally root group is going
3195                  * to have higher weight. A more accurate thing would be to
3196                  * calculate system wide asnc/sync ratio.
3197                  */
3198                 tmp = cfqd->cfq_target_latency *
3199                         cfqg_busy_async_queues(cfqd, cfqg);
3200                 tmp = div_u64(tmp, cfqd->busy_queues);
3201                 slice = min_t(u64, slice, tmp);
3202
3203                 /* async workload slice is scaled down according to
3204                  * the sync/async slice ratio. */
3205                 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3206         } else
3207                 /* sync workload slice is at least 2 * cfq_slice_idle */
3208                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3209
3210         slice = max_t(u64, slice, CFQ_MIN_TT);
3211         cfq_log(cfqd, "workload slice:%llu", slice);
3212         cfqd->workload_expires = now + slice;
3213 }
3214
3215 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3216 {
3217         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3218         struct cfq_group *cfqg;
3219
3220         if (RB_EMPTY_ROOT(&st->rb))
3221                 return NULL;
3222         cfqg = cfq_rb_first_group(st);
3223         update_min_vdisktime(st);
3224         return cfqg;
3225 }
3226
3227 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3228 {
3229         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3230         u64 now = ktime_get_ns();
3231
3232         cfqd->serving_group = cfqg;
3233
3234         /* Restore the workload type data */
3235         if (cfqg->saved_wl_slice) {
3236                 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3237                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3238                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3239         } else
3240                 cfqd->workload_expires = now - 1;
3241
3242         choose_wl_class_and_type(cfqd, cfqg);
3243 }
3244
3245 /*
3246  * Select a queue for service. If we have a current active queue,
3247  * check whether to continue servicing it, or retrieve and set a new one.
3248  */
3249 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3250 {
3251         struct cfq_queue *cfqq, *new_cfqq = NULL;
3252         u64 now = ktime_get_ns();
3253
3254         cfqq = cfqd->active_queue;
3255         if (!cfqq)
3256                 goto new_queue;
3257
3258         if (!cfqd->rq_queued)
3259                 return NULL;
3260
3261         /*
3262          * We were waiting for group to get backlogged. Expire the queue
3263          */
3264         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3265                 goto expire;
3266
3267         /*
3268          * The active queue has run out of time, expire it and select new.
3269          */
3270         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3271                 /*
3272                  * If slice had not expired at the completion of last request
3273                  * we might not have turned on wait_busy flag. Don't expire
3274                  * the queue yet. Allow the group to get backlogged.
3275                  *
3276                  * The very fact that we have used the slice, that means we
3277                  * have been idling all along on this queue and it should be
3278                  * ok to wait for this request to complete.
3279                  */
3280                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3281                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3282                         cfqq = NULL;
3283                         goto keep_queue;
3284                 } else
3285                         goto check_group_idle;
3286         }
3287
3288         /*
3289          * The active queue has requests and isn't expired, allow it to
3290          * dispatch.
3291          */
3292         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3293                 goto keep_queue;
3294
3295         /*
3296          * If another queue has a request waiting within our mean seek
3297          * distance, let it run.  The expire code will check for close
3298          * cooperators and put the close queue at the front of the service
3299          * tree.  If possible, merge the expiring queue with the new cfqq.
3300          */
3301         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3302         if (new_cfqq) {
3303                 if (!cfqq->new_cfqq)
3304                         cfq_setup_merge(cfqq, new_cfqq);
3305                 goto expire;
3306         }
3307
3308         /*
3309          * No requests pending. If the active queue still has requests in
3310          * flight or is idling for a new request, allow either of these
3311          * conditions to happen (or time out) before selecting a new queue.
3312          */
3313         if (hrtimer_active(&cfqd->idle_slice_timer)) {
3314                 cfqq = NULL;
3315                 goto keep_queue;
3316         }
3317
3318         /*
3319          * This is a deep seek queue, but the device is much faster than
3320          * the queue can deliver, don't idle
3321          **/
3322         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3323             (cfq_cfqq_slice_new(cfqq) ||
3324             (cfqq->slice_end - now > now - cfqq->slice_start))) {
3325                 cfq_clear_cfqq_deep(cfqq);
3326                 cfq_clear_cfqq_idle_window(cfqq);
3327         }
3328
3329         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3330                 cfqq = NULL;
3331                 goto keep_queue;
3332         }
3333
3334         /*
3335          * If group idle is enabled and there are requests dispatched from
3336          * this group, wait for requests to complete.
3337          */
3338 check_group_idle:
3339         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3340             cfqq->cfqg->dispatched &&
3341             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3342                 cfqq = NULL;
3343                 goto keep_queue;
3344         }
3345
3346 expire:
3347         cfq_slice_expired(cfqd, 0);
3348 new_queue:
3349         /*
3350          * Current queue expired. Check if we have to switch to a new
3351          * service tree
3352          */
3353         if (!new_cfqq)
3354                 cfq_choose_cfqg(cfqd);
3355
3356         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3357 keep_queue:
3358         return cfqq;
3359 }
3360
3361 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3362 {
3363         int dispatched = 0;
3364