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