block/cfq-iosched.c

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