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         ret = -EINVAL;
1756         cfqg = blkg_to_cfqg(ctx.blkg);
1757         cfqgd = blkcg_to_cfqgd(blkcg);
1758         if (!cfqg || !cfqgd)
1759                 goto err;
1760
1761         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1762                 if (!is_leaf_weight) {
1763                         cfqg->dev_weight = ctx.v;
1764                         cfqg->new_weight = ctx.v ?: cfqgd->weight;
1765                 } else {
1766                         cfqg->dev_leaf_weight = ctx.v;
1767                         cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1768                 }
1769                 ret = 0;
1770         }
1771
1772 err:
1773         blkg_conf_finish(&ctx);
1774         return ret ?: nbytes;
1775 }
1776
1777 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1778                                       char *buf, size_t nbytes, loff_t off)
1779 {
1780         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1781 }
1782
1783 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1784                                            char *buf, size_t nbytes, loff_t off)
1785 {
1786         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1787 }
1788
1789 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1790                             u64 val, bool is_leaf_weight)
1791 {
1792         struct blkcg *blkcg = css_to_blkcg(css);
1793         struct blkcg_gq *blkg;
1794         struct cfq_group_data *cfqgd;
1795         int ret = 0;
1796
1797         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1798                 return -EINVAL;
1799
1800         spin_lock_irq(&blkcg->lock);
1801         cfqgd = blkcg_to_cfqgd(blkcg);
1802         if (!cfqgd) {
1803                 ret = -EINVAL;
1804                 goto out;
1805         }
1806
1807         if (!is_leaf_weight)
1808                 cfqgd->weight = val;
1809         else
1810                 cfqgd->leaf_weight = val;
1811
1812         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1813                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1814
1815                 if (!cfqg)
1816                         continue;
1817
1818                 if (!is_leaf_weight) {
1819                         if (!cfqg->dev_weight)
1820                                 cfqg->new_weight = cfqgd->weight;
1821                 } else {
1822                         if (!cfqg->dev_leaf_weight)
1823                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1824                 }
1825         }
1826
1827 out:
1828         spin_unlock_irq(&blkcg->lock);
1829         return ret;
1830 }
1831
1832 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1833                           u64 val)
1834 {
1835         return __cfq_set_weight(css, cft, val, false);
1836 }
1837
1838 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1839                                struct cftype *cft, u64 val)
1840 {
1841         return __cfq_set_weight(css, cft, val, true);
1842 }
1843
1844 static int cfqg_print_stat(struct seq_file *sf, void *v)
1845 {
1846         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1847                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1848         return 0;
1849 }
1850
1851 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1852 {
1853         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1854                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1855         return 0;
1856 }
1857
1858 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1859                                       struct blkg_policy_data *pd, int off)
1860 {
1861         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1862                                           &blkcg_policy_cfq, off);
1863         return __blkg_prfill_u64(sf, pd, sum);
1864 }
1865
1866 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1867                                         struct blkg_policy_data *pd, int off)
1868 {
1869         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1870                                                         &blkcg_policy_cfq, off);
1871         return __blkg_prfill_rwstat(sf, pd, &sum);
1872 }
1873
1874 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1875 {
1876         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1877                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1878                           seq_cft(sf)->private, false);
1879         return 0;
1880 }
1881
1882 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1883 {
1884         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1885                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1886                           seq_cft(sf)->private, true);
1887         return 0;
1888 }
1889
1890 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1891                                int off)
1892 {
1893         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1894
1895         return __blkg_prfill_u64(sf, pd, sum >> 9);
1896 }
1897
1898 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1899 {
1900         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1901                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1902         return 0;
1903 }
1904
1905 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1906                                          struct blkg_policy_data *pd, int off)
1907 {
1908         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1909                                         offsetof(struct blkcg_gq, stat_bytes));
1910         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1911                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1912
1913         return __blkg_prfill_u64(sf, pd, sum >> 9);
1914 }
1915
1916 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1917 {
1918         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1919                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1920                           false);
1921         return 0;
1922 }
1923
1924 #ifdef CONFIG_DEBUG_BLK_CGROUP
1925 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1926                                       struct blkg_policy_data *pd, int off)
1927 {
1928         struct cfq_group *cfqg = pd_to_cfqg(pd);
1929         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1930         u64 v = 0;
1931
1932         if (samples) {
1933                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1934                 v = div64_u64(v, samples);
1935         }
1936         __blkg_prfill_u64(sf, pd, v);
1937         return 0;
1938 }
1939
1940 /* print avg_queue_size */
1941 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1942 {
1943         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1944                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1945                           0, false);
1946         return 0;
1947 }
1948 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1949
1950 static struct cftype cfq_blkcg_files[] = {
1951         /* on root, weight is mapped to leaf_weight */
1952         {
1953                 .name = "weight_device",
1954                 .flags = CFTYPE_ONLY_ON_ROOT,
1955                 .seq_show = cfqg_print_leaf_weight_device,
1956                 .write = cfqg_set_leaf_weight_device,
1957         },
1958         {
1959                 .name = "weight",
1960                 .flags = CFTYPE_ONLY_ON_ROOT,
1961                 .seq_show = cfq_print_leaf_weight,
1962                 .write_u64 = cfq_set_leaf_weight,
1963         },
1964
1965         /* no such mapping necessary for !roots */
1966         {
1967                 .name = "weight_device",
1968                 .flags = CFTYPE_NOT_ON_ROOT,
1969                 .seq_show = cfqg_print_weight_device,
1970                 .write = cfqg_set_weight_device,
1971         },
1972         {
1973                 .name = "weight",
1974                 .flags = CFTYPE_NOT_ON_ROOT,
1975                 .seq_show = cfq_print_weight,
1976                 .write_u64 = cfq_set_weight,
1977         },
1978
1979         {
1980                 .name = "leaf_weight_device",
1981                 .seq_show = cfqg_print_leaf_weight_device,
1982                 .write = cfqg_set_leaf_weight_device,
1983         },
1984         {
1985                 .name = "leaf_weight",
1986                 .seq_show = cfq_print_leaf_weight,
1987                 .write_u64 = cfq_set_leaf_weight,
1988         },
1989
1990         /* statistics, covers only the tasks in the cfqg */
1991         {
1992                 .name = "time",
1993                 .private = offsetof(struct cfq_group, stats.time),
1994                 .seq_show = cfqg_print_stat,
1995         },
1996         {
1997                 .name = "sectors",
1998                 .seq_show = cfqg_print_stat_sectors,
1999         },
2000         {
2001                 .name = "io_service_bytes",
2002                 .private = (unsigned long)&blkcg_policy_cfq,
2003                 .seq_show = blkg_print_stat_bytes,
2004         },
2005         {
2006                 .name = "io_serviced",
2007                 .private = (unsigned long)&blkcg_policy_cfq,
2008                 .seq_show = blkg_print_stat_ios,
2009         },
2010         {
2011                 .name = "io_service_time",
2012                 .private = offsetof(struct cfq_group, stats.service_time),
2013                 .seq_show = cfqg_print_rwstat,
2014         },
2015         {
2016                 .name = "io_wait_time",
2017                 .private = offsetof(struct cfq_group, stats.wait_time),
2018                 .seq_show = cfqg_print_rwstat,
2019         },
2020         {
2021                 .name = "io_merged",
2022                 .private = offsetof(struct cfq_group, stats.merged),
2023                 .seq_show = cfqg_print_rwstat,
2024         },
2025         {
2026                 .name = "io_queued",
2027                 .private = offsetof(struct cfq_group, stats.queued),
2028                 .seq_show = cfqg_print_rwstat,
2029         },
2030
2031         /* the same statictics which cover the cfqg and its descendants */
2032         {
2033                 .name = "time_recursive",
2034                 .private = offsetof(struct cfq_group, stats.time),
2035                 .seq_show = cfqg_print_stat_recursive,
2036         },
2037         {
2038                 .name = "sectors_recursive",
2039                 .seq_show = cfqg_print_stat_sectors_recursive,
2040         },
2041         {
2042                 .name = "io_service_bytes_recursive",
2043                 .private = (unsigned long)&blkcg_policy_cfq,
2044                 .seq_show = blkg_print_stat_bytes_recursive,
2045         },
2046         {
2047                 .name = "io_serviced_recursive",
2048                 .private = (unsigned long)&blkcg_policy_cfq,
2049                 .seq_show = blkg_print_stat_ios_recursive,
2050         },
2051         {
2052                 .name = "io_service_time_recursive",
2053                 .private = offsetof(struct cfq_group, stats.service_time),
2054                 .seq_show = cfqg_print_rwstat_recursive,
2055         },
2056         {
2057                 .name = "io_wait_time_recursive",
2058                 .private = offsetof(struct cfq_group, stats.wait_time),
2059                 .seq_show = cfqg_print_rwstat_recursive,
2060         },
2061         {
2062                 .name = "io_merged_recursive",
2063                 .private = offsetof(struct cfq_group, stats.merged),
2064                 .seq_show = cfqg_print_rwstat_recursive,
2065         },
2066         {
2067                 .name = "io_queued_recursive",
2068                 .private = offsetof(struct cfq_group, stats.queued),
2069                 .seq_show = cfqg_print_rwstat_recursive,
2070         },
2071 #ifdef CONFIG_DEBUG_BLK_CGROUP
2072         {
2073                 .name = "avg_queue_size",
2074                 .seq_show = cfqg_print_avg_queue_size,
2075         },
2076         {
2077                 .name = "group_wait_time",
2078                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2079                 .seq_show = cfqg_print_stat,
2080         },
2081         {
2082                 .name = "idle_time",
2083                 .private = offsetof(struct cfq_group, stats.idle_time),
2084                 .seq_show = cfqg_print_stat,
2085         },
2086         {
2087                 .name = "empty_time",
2088                 .private = offsetof(struct cfq_group, stats.empty_time),
2089                 .seq_show = cfqg_print_stat,
2090         },
2091         {
2092                 .name = "dequeue",
2093                 .private = offsetof(struct cfq_group, stats.dequeue),
2094                 .seq_show = cfqg_print_stat,
2095         },
2096         {
2097                 .name = "unaccounted_time",
2098                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2099                 .seq_show = cfqg_print_stat,
2100         },
2101 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2102         { }     /* terminate */
2103 };
2104 #else /* GROUP_IOSCHED */
2105 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2106                                          struct blkcg *blkcg)
2107 {
2108         return cfqd->root_group;
2109 }
2110
2111 static inline void
2112 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2113         cfqq->cfqg = cfqg;
2114 }
2115
2116 #endif /* GROUP_IOSCHED */
2117
2118 /*
2119  * The cfqd->service_trees holds all pending cfq_queue's that have
2120  * requests waiting to be processed. It is sorted in the order that
2121  * we will service the queues.
2122  */
2123 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2124                                  bool add_front)
2125 {
2126         struct rb_node **p, *parent;
2127         struct cfq_queue *__cfqq;
2128         unsigned long rb_key;
2129         struct cfq_rb_root *st;
2130         int left;
2131         int new_cfqq = 1;
2132
2133         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2134         if (cfq_class_idle(cfqq)) {
2135                 rb_key = CFQ_IDLE_DELAY;
2136                 parent = rb_last(&st->rb);
2137                 if (parent && parent != &cfqq->rb_node) {
2138                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2139                         rb_key += __cfqq->rb_key;
2140                 } else
2141                         rb_key += jiffies;
2142         } else if (!add_front) {
2143                 /*
2144                  * Get our rb key offset. Subtract any residual slice
2145                  * value carried from last service. A negative resid
2146                  * count indicates slice overrun, and this should position
2147                  * the next service time further away in the tree.
2148                  */
2149                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2150                 rb_key -= cfqq->slice_resid;
2151                 cfqq->slice_resid = 0;
2152         } else {
2153                 rb_key = -HZ;
2154                 __cfqq = cfq_rb_first(st);
2155                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2156         }
2157
2158         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2159                 new_cfqq = 0;
2160                 /*
2161                  * same position, nothing more to do
2162                  */
2163                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2164                         return;
2165
2166                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2167                 cfqq->service_tree = NULL;
2168         }
2169
2170         left = 1;
2171         parent = NULL;
2172         cfqq->service_tree = st;
2173         p = &st->rb.rb_node;
2174         while (*p) {
2175                 parent = *p;
2176                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2177
2178                 /*
2179                  * sort by key, that represents service time.
2180                  */
2181                 if (time_before(rb_key, __cfqq->rb_key))
2182                         p = &parent->rb_left;
2183                 else {
2184                         p = &parent->rb_right;
2185                         left = 0;
2186                 }
2187         }
2188
2189         if (left)
2190                 st->left = &cfqq->rb_node;
2191
2192         cfqq->rb_key = rb_key;
2193         rb_link_node(&cfqq->rb_node, parent, p);
2194         rb_insert_color(&cfqq->rb_node, &st->rb);
2195         st->count++;
2196         if (add_front || !new_cfqq)
2197                 return;
2198         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2199 }
2200
2201 static struct cfq_queue *
2202 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2203                      sector_t sector, struct rb_node **ret_parent,
2204                      struct rb_node ***rb_link)
2205 {
2206         struct rb_node **p, *parent;
2207         struct cfq_queue *cfqq = NULL;
2208
2209         parent = NULL;
2210         p = &root->rb_node;
2211         while (*p) {
2212                 struct rb_node **n;
2213
2214                 parent = *p;
2215                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2216
2217                 /*
2218                  * Sort strictly based on sector.  Smallest to the left,
2219                  * largest to the right.
2220                  */
2221                 if (sector > blk_rq_pos(cfqq->next_rq))
2222                         n = &(*p)->rb_right;
2223                 else if (sector < blk_rq_pos(cfqq->next_rq))
2224                         n = &(*p)->rb_left;
2225                 else
2226                         break;
2227                 p = n;
2228                 cfqq = NULL;
2229         }
2230
2231         *ret_parent = parent;
2232         if (rb_link)
2233                 *rb_link = p;
2234         return cfqq;
2235 }
2236
2237 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2238 {
2239         struct rb_node **p, *parent;
2240         struct cfq_queue *__cfqq;
2241
2242         if (cfqq->p_root) {
2243                 rb_erase(&cfqq->p_node, cfqq->p_root);
2244                 cfqq->p_root = NULL;
2245         }
2246
2247         if (cfq_class_idle(cfqq))
2248                 return;
2249         if (!cfqq->next_rq)
2250                 return;
2251
2252         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2253         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2254                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2255         if (!__cfqq) {
2256                 rb_link_node(&cfqq->p_node, parent, p);
2257                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2258         } else
2259                 cfqq->p_root = NULL;
2260 }
2261
2262 /*
2263  * Update cfqq's position in the service tree.
2264  */
2265 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2266 {
2267         /*
2268          * Resorting requires the cfqq to be on the RR list already.
2269          */
2270         if (cfq_cfqq_on_rr(cfqq)) {
2271                 cfq_service_tree_add(cfqd, cfqq, 0);
2272                 cfq_prio_tree_add(cfqd, cfqq);
2273         }
2274 }
2275
2276 /*
2277  * add to busy list of queues for service, trying to be fair in ordering
2278  * the pending list according to last request service
2279  */
2280 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2281 {
2282         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2283         BUG_ON(cfq_cfqq_on_rr(cfqq));
2284         cfq_mark_cfqq_on_rr(cfqq);
2285         cfqd->busy_queues++;
2286         if (cfq_cfqq_sync(cfqq))
2287                 cfqd->busy_sync_queues++;
2288
2289         cfq_resort_rr_list(cfqd, cfqq);
2290 }
2291
2292 /*
2293  * Called when the cfqq no longer has requests pending, remove it from
2294  * the service tree.
2295  */
2296 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2297 {
2298         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2299         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2300         cfq_clear_cfqq_on_rr(cfqq);
2301
2302         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2303                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2304                 cfqq->service_tree = NULL;
2305         }
2306         if (cfqq->p_root) {
2307                 rb_erase(&cfqq->p_node, cfqq->p_root);
2308                 cfqq->p_root = NULL;
2309         }
2310
2311         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2312         BUG_ON(!cfqd->busy_queues);
2313         cfqd->busy_queues--;
2314         if (cfq_cfqq_sync(cfqq))
2315                 cfqd->busy_sync_queues--;
2316 }
2317
2318 /*
2319  * rb tree support functions
2320  */
2321 static void cfq_del_rq_rb(struct request *rq)
2322 {
2323         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2324         const int sync = rq_is_sync(rq);
2325
2326         BUG_ON(!cfqq->queued[sync]);
2327         cfqq->queued[sync]--;
2328
2329         elv_rb_del(&cfqq->sort_list, rq);
2330
2331         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2332                 /*
2333                  * Queue will be deleted from service tree when we actually
2334                  * expire it later. Right now just remove it from prio tree
2335                  * as it is empty.
2336                  */
2337                 if (cfqq->p_root) {
2338                         rb_erase(&cfqq->p_node, cfqq->p_root);
2339                         cfqq->p_root = NULL;
2340                 }
2341         }
2342 }
2343
2344 static void cfq_add_rq_rb(struct request *rq)
2345 {
2346         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2347         struct cfq_data *cfqd = cfqq->cfqd;
2348         struct request *prev;
2349
2350         cfqq->queued[rq_is_sync(rq)]++;
2351
2352         elv_rb_add(&cfqq->sort_list, rq);
2353
2354         if (!cfq_cfqq_on_rr(cfqq))
2355                 cfq_add_cfqq_rr(cfqd, cfqq);
2356
2357         /*
2358          * check if this request is a better next-serve candidate
2359          */
2360         prev = cfqq->next_rq;
2361         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2362
2363         /*
2364          * adjust priority tree position, if ->next_rq changes
2365          */
2366         if (prev != cfqq->next_rq)
2367                 cfq_prio_tree_add(cfqd, cfqq);
2368
2369         BUG_ON(!cfqq->next_rq);
2370 }
2371
2372 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2373 {
2374         elv_rb_del(&cfqq->sort_list, rq);
2375         cfqq->queued[rq_is_sync(rq)]--;
2376         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2377         cfq_add_rq_rb(rq);
2378         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2379                                  rq->cmd_flags);
2380 }
2381
2382 static struct request *
2383 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2384 {
2385         struct task_struct *tsk = current;
2386         struct cfq_io_cq *cic;
2387         struct cfq_queue *cfqq;
2388
2389         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2390         if (!cic)
2391                 return NULL;
2392
2393         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2394         if (cfqq)
2395                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2396
2397         return NULL;
2398 }
2399
2400 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2401 {
2402         struct cfq_data *cfqd = q->elevator->elevator_data;
2403
2404         cfqd->rq_in_driver++;
2405         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2406                                                 cfqd->rq_in_driver);
2407
2408         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2409 }
2410
2411 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2412 {
2413         struct cfq_data *cfqd = q->elevator->elevator_data;
2414
2415         WARN_ON(!cfqd->rq_in_driver);
2416         cfqd->rq_in_driver--;
2417         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2418                                                 cfqd->rq_in_driver);
2419 }
2420
2421 static void cfq_remove_request(struct request *rq)
2422 {
2423         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2424
2425         if (cfqq->next_rq == rq)
2426                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2427
2428         list_del_init(&rq->queuelist);
2429         cfq_del_rq_rb(rq);
2430
2431         cfqq->cfqd->rq_queued--;
2432         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2433         if (rq->cmd_flags & REQ_PRIO) {
2434                 WARN_ON(!cfqq->prio_pending);
2435                 cfqq->prio_pending--;
2436         }
2437 }
2438
2439 static int cfq_merge(struct request_queue *q, struct request **req,
2440                      struct bio *bio)
2441 {
2442         struct cfq_data *cfqd = q->elevator->elevator_data;
2443         struct request *__rq;
2444
2445         __rq = cfq_find_rq_fmerge(cfqd, bio);
2446         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2447                 *req = __rq;
2448                 return ELEVATOR_FRONT_MERGE;
2449         }
2450
2451         return ELEVATOR_NO_MERGE;
2452 }
2453
2454 static void cfq_merged_request(struct request_queue *q, struct request *req,
2455                                int type)
2456 {
2457         if (type == ELEVATOR_FRONT_MERGE) {
2458                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2459
2460                 cfq_reposition_rq_rb(cfqq, req);
2461         }
2462 }
2463
2464 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2465                                 struct bio *bio)
2466 {
2467         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2468 }
2469
2470 static void
2471 cfq_merged_requests(struct request_queue *q, struct request *rq,
2472                     struct request *next)
2473 {
2474         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2475         struct cfq_data *cfqd = q->elevator->elevator_data;
2476
2477         /*
2478          * reposition in fifo if next is older than rq
2479          */
2480         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2481             time_before(next->fifo_time, rq->fifo_time) &&
2482             cfqq == RQ_CFQQ(next)) {
2483                 list_move(&rq->queuelist, &next->queuelist);
2484                 rq->fifo_time = next->fifo_time;
2485         }
2486
2487         if (cfqq->next_rq == next)
2488                 cfqq->next_rq = rq;
2489         cfq_remove_request(next);
2490         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2491
2492         cfqq = RQ_CFQQ(next);
2493         /*
2494          * all requests of this queue are merged to other queues, delete it
2495          * from the service tree. If it's the active_queue,
2496          * cfq_dispatch_requests() will choose to expire it or do idle
2497          */
2498         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2499             cfqq != cfqd->active_queue)
2500                 cfq_del_cfqq_rr(cfqd, cfqq);
2501 }
2502
2503 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2504                            struct bio *bio)
2505 {
2506         struct cfq_data *cfqd = q->elevator->elevator_data;
2507         struct cfq_io_cq *cic;
2508         struct cfq_queue *cfqq;
2509
2510         /*
2511          * Disallow merge of a sync bio into an async request.
2512          */
2513         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2514                 return false;
2515
2516         /*
2517          * Lookup the cfqq that this bio will be queued with and allow
2518          * merge only if rq is queued there.
2519          */
2520         cic = cfq_cic_lookup(cfqd, current->io_context);
2521         if (!cic)
2522                 return false;
2523
2524         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2525         return cfqq == RQ_CFQQ(rq);
2526 }
2527
2528 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2529 {
2530         del_timer(&cfqd->idle_slice_timer);
2531         cfqg_stats_update_idle_time(cfqq->cfqg);
2532 }
2533
2534 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2535                                    struct cfq_queue *cfqq)
2536 {
2537         if (cfqq) {
2538                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2539                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2540                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2541                 cfqq->slice_start = 0;
2542                 cfqq->dispatch_start = jiffies;
2543                 cfqq->allocated_slice = 0;
2544                 cfqq->slice_end = 0;
2545                 cfqq->slice_dispatch = 0;
2546                 cfqq->nr_sectors = 0;
2547
2548                 cfq_clear_cfqq_wait_request(cfqq);
2549                 cfq_clear_cfqq_must_dispatch(cfqq);
2550                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2551                 cfq_clear_cfqq_fifo_expire(cfqq);
2552                 cfq_mark_cfqq_slice_new(cfqq);
2553
2554                 cfq_del_timer(cfqd, cfqq);
2555         }
2556
2557         cfqd->active_queue = cfqq;
2558 }
2559
2560 /*
2561  * current cfqq expired its slice (or was too idle), select new one
2562  */
2563 static void
2564 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2565                     bool timed_out)
2566 {
2567         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2568
2569         if (cfq_cfqq_wait_request(cfqq))
2570                 cfq_del_timer(cfqd, cfqq);
2571
2572         cfq_clear_cfqq_wait_request(cfqq);
2573         cfq_clear_cfqq_wait_busy(cfqq);
2574
2575         /*
2576          * If this cfqq is shared between multiple processes, check to
2577          * make sure that those processes are still issuing I/Os within
2578          * the mean seek distance.  If not, it may be time to break the
2579          * queues apart again.
2580          */
2581         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2582                 cfq_mark_cfqq_split_coop(cfqq);
2583
2584         /*
2585          * store what was left of this slice, if the queue idled/timed out
2586          */
2587         if (timed_out) {
2588                 if (cfq_cfqq_slice_new(cfqq))
2589                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2590                 else
2591                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2592                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2593         }
2594
2595         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2596
2597         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2598                 cfq_del_cfqq_rr(cfqd, cfqq);
2599
2600         cfq_resort_rr_list(cfqd, cfqq);
2601
2602         if (cfqq == cfqd->active_queue)
2603                 cfqd->active_queue = NULL;
2604
2605         if (cfqd->active_cic) {
2606                 put_io_context(cfqd->active_cic->icq.ioc);
2607                 cfqd->active_cic = NULL;
2608         }
2609 }
2610
2611 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2612 {
2613         struct cfq_queue *cfqq = cfqd->active_queue;
2614
2615         if (cfqq)
2616                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2617 }
2618
2619 /*
2620  * Get next queue for service. Unless we have a queue preemption,
2621  * we'll simply select the first cfqq in the service tree.
2622  */
2623 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2624 {
2625         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2626                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2627
2628         if (!cfqd->rq_queued)
2629                 return NULL;
2630
2631         /* There is nothing to dispatch */
2632         if (!st)
2633                 return NULL;
2634         if (RB_EMPTY_ROOT(&st->rb))
2635                 return NULL;
2636         return cfq_rb_first(st);
2637 }
2638
2639 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2640 {
2641         struct cfq_group *cfqg;
2642         struct cfq_queue *cfqq;
2643         int i, j;
2644         struct cfq_rb_root *st;
2645
2646         if (!cfqd->rq_queued)
2647                 return NULL;
2648
2649         cfqg = cfq_get_next_cfqg(cfqd);
2650         if (!cfqg)
2651                 return NULL;
2652
2653         for_each_cfqg_st(cfqg, i, j, st)
2654                 if ((cfqq = cfq_rb_first(st)) != NULL)
2655                         return cfqq;
2656         return NULL;
2657 }
2658
2659 /*
2660  * Get and set a new active queue for service.
2661  */
2662 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2663                                               struct cfq_queue *cfqq)
2664 {
2665         if (!cfqq)
2666                 cfqq = cfq_get_next_queue(cfqd);
2667
2668         __cfq_set_active_queue(cfqd, cfqq);
2669         return cfqq;
2670 }
2671
2672 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2673                                           struct request *rq)
2674 {
2675         if (blk_rq_pos(rq) >= cfqd->last_position)
2676                 return blk_rq_pos(rq) - cfqd->last_position;
2677         else
2678                 return cfqd->last_position - blk_rq_pos(rq);
2679 }
2680
2681 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2682                                struct request *rq)
2683 {
2684         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2685 }
2686
2687 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2688                                     struct cfq_queue *cur_cfqq)
2689 {
2690         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2691         struct rb_node *parent, *node;
2692         struct cfq_queue *__cfqq;
2693         sector_t sector = cfqd->last_position;
2694
2695         if (RB_EMPTY_ROOT(root))
2696                 return NULL;
2697
2698         /*
2699          * First, if we find a request starting at the end of the last
2700          * request, choose it.
2701          */
2702         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2703         if (__cfqq)
2704                 return __cfqq;
2705
2706         /*
2707          * If the exact sector wasn't found, the parent of the NULL leaf
2708          * will contain the closest sector.
2709          */
2710         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2711         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2712                 return __cfqq;
2713
2714         if (blk_rq_pos(__cfqq->next_rq) < sector)
2715                 node = rb_next(&__cfqq->p_node);
2716         else
2717                 node = rb_prev(&__cfqq->p_node);
2718         if (!node)
2719                 return NULL;
2720
2721         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2722         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2723                 return __cfqq;
2724
2725         return NULL;
2726 }
2727
2728 /*
2729  * cfqd - obvious
2730  * cur_cfqq - passed in so that we don't decide that the current queue is
2731  *            closely cooperating with itself.
2732  *
2733  * So, basically we're assuming that that cur_cfqq has dispatched at least
2734  * one request, and that cfqd->last_position reflects a position on the disk
2735  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2736  * assumption.
2737  */
2738 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2739                                               struct cfq_queue *cur_cfqq)
2740 {
2741         struct cfq_queue *cfqq;
2742
2743         if (cfq_class_idle(cur_cfqq))
2744                 return NULL;
2745         if (!cfq_cfqq_sync(cur_cfqq))
2746                 return NULL;
2747         if (CFQQ_SEEKY(cur_cfqq))
2748                 return NULL;
2749
2750         /*
2751          * Don't search priority tree if it's the only queue in the group.
2752          */
2753         if (cur_cfqq->cfqg->nr_cfqq == 1)
2754                 return NULL;
2755
2756         /*
2757          * We should notice if some of the queues are cooperating, eg
2758          * working closely on the same area of the disk. In that case,
2759          * we can group them together and don't waste time idling.
2760          */
2761         cfqq = cfqq_close(cfqd, cur_cfqq);
2762         if (!cfqq)
2763                 return NULL;
2764
2765         /* If new queue belongs to different cfq_group, don't choose it */
2766         if (cur_cfqq->cfqg != cfqq->cfqg)
2767                 return NULL;
2768
2769         /*
2770          * It only makes sense to merge sync queues.
2771          */
2772         if (!cfq_cfqq_sync(cfqq))
2773                 return NULL;
2774         if (CFQQ_SEEKY(cfqq))
2775                 return NULL;
2776
2777         /*
2778          * Do not merge queues of different priority classes
2779          */
2780         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2781                 return NULL;
2782
2783         return cfqq;
2784 }
2785
2786 /*
2787  * Determine whether we should enforce idle window for this queue.
2788  */
2789
2790 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2791 {
2792         enum wl_class_t wl_class = cfqq_class(cfqq);
2793         struct cfq_rb_root *st = cfqq->service_tree;
2794
2795         BUG_ON(!st);
2796         BUG_ON(!st->count);
2797
2798         if (!cfqd->cfq_slice_idle)
2799                 return false;
2800
2801         /* We never do for idle class queues. */
2802         if (wl_class == IDLE_WORKLOAD)
2803                 return false;
2804
2805         /* We do for queues that were marked with idle window flag. */
2806         if (cfq_cfqq_idle_window(cfqq) &&
2807            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2808                 return true;
2809
2810         /*
2811          * Otherwise, we do only if they are the last ones
2812          * in their service tree.
2813          */
2814         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2815            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2816                 return true;
2817         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2818         return false;
2819 }
2820
2821 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2822 {
2823         struct cfq_queue *cfqq = cfqd->active_queue;
2824         struct cfq_io_cq *cic;
2825         unsigned long sl, group_idle = 0;
2826
2827         /*
2828          * SSD device without seek penalty, disable idling. But only do so
2829          * for devices that support queuing, otherwise we still have a problem
2830          * with sync vs async workloads.
2831          */
2832         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2833                 return;
2834
2835         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2836         WARN_ON(cfq_cfqq_slice_new(cfqq));
2837
2838         /*
2839          * idle is disabled, either manually or by past process history
2840          */
2841         if (!cfq_should_idle(cfqd, cfqq)) {
2842                 /* no queue idling. Check for group idling */
2843                 if (cfqd->cfq_group_idle)
2844                         group_idle = cfqd->cfq_group_idle;
2845                 else
2846                         return;
2847         }
2848
2849         /*
2850          * still active requests from this queue, don't idle
2851          */
2852         if (cfqq->dispatched)
2853                 return;
2854
2855         /*
2856          * task has exited, don't wait
2857          */
2858         cic = cfqd->active_cic;
2859         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2860                 return;
2861
2862         /*
2863          * If our average think time is larger than the remaining time
2864          * slice, then don't idle. This avoids overrunning the allotted
2865          * time slice.
2866          */
2867         if (sample_valid(cic->ttime.ttime_samples) &&
2868             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2869                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2870                              cic->ttime.ttime_mean);
2871                 return;
2872         }
2873
2874         /* There are other queues in the group, don't do group idle */
2875         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2876                 return;
2877
2878         cfq_mark_cfqq_wait_request(cfqq);
2879
2880         if (group_idle)
2881                 sl = cfqd->cfq_group_idle;
2882         else
2883                 sl = cfqd->cfq_slice_idle;
2884
2885         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2886         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2887         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2888                         group_idle ? 1 : 0);
2889 }
2890
2891 /*
2892  * Move request from internal lists to the request queue dispatch list.
2893  */
2894 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2895 {
2896         struct cfq_data *cfqd = q->elevator->elevator_data;
2897         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2898
2899         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2900
2901         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2902         cfq_remove_request(rq);
2903         cfqq->dispatched++;
2904         (RQ_CFQG(rq))->dispatched++;
2905         elv_dispatch_sort(q, rq);
2906
2907         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2908         cfqq->nr_sectors += blk_rq_sectors(rq);
2909 }
2910
2911 /*
2912  * return expired entry, or NULL to just start from scratch in rbtree
2913  */
2914 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2915 {
2916         struct request *rq = NULL;
2917
2918         if (cfq_cfqq_fifo_expire(cfqq))
2919                 return NULL;
2920
2921         cfq_mark_cfqq_fifo_expire(cfqq);
2922
2923         if (list_empty(&cfqq->fifo))
2924                 return NULL;
2925
2926         rq = rq_entry_fifo(cfqq->fifo.next);
2927         if (time_before(jiffies, rq->fifo_time))
2928                 rq = NULL;
2929
2930         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2931         return rq;
2932 }
2933
2934 static inline int
2935 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2936 {
2937         const int base_rq = cfqd->cfq_slice_async_rq;
2938
2939         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2940
2941         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2942 }
2943
2944 /*
2945  * Must be called with the queue_lock held.
2946  */
2947 static int cfqq_process_refs(struct cfq_queue *cfqq)
2948 {
2949         int process_refs, io_refs;
2950
2951         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2952         process_refs = cfqq->ref - io_refs;
2953         BUG_ON(process_refs < 0);
2954         return process_refs;
2955 }
2956
2957 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2958 {
2959         int process_refs, new_process_refs;
2960         struct cfq_queue *__cfqq;
2961
2962         /*
2963          * If there are no process references on the new_cfqq, then it is
2964          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2965          * chain may have dropped their last reference (not just their
2966          * last process reference).
2967          */
2968         if (!cfqq_process_refs(new_cfqq))
2969                 return;
2970
2971         /* Avoid a circular list and skip interim queue merges */
2972         while ((__cfqq = new_cfqq->new_cfqq)) {
2973                 if (__cfqq == cfqq)
2974                         return;
2975                 new_cfqq = __cfqq;
2976         }
2977
2978         process_refs = cfqq_process_refs(cfqq);
2979         new_process_refs = cfqq_process_refs(new_cfqq);
2980         /*
2981          * If the process for the cfqq has gone away, there is no
2982          * sense in merging the queues.
2983          */
2984         if (process_refs == 0 || new_process_refs == 0)
2985                 return;
2986
2987         /*
2988          * Merge in the direction of the lesser amount of work.
2989          */
2990         if (new_process_refs >= process_refs) {
2991                 cfqq->new_cfqq = new_cfqq;
2992                 new_cfqq->ref += process_refs;
2993         } else {
2994                 new_cfqq->new_cfqq = cfqq;
2995                 cfqq->ref += new_process_refs;
2996         }
2997 }
2998
2999 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3000                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3001 {
3002         struct cfq_queue *queue;
3003         int i;
3004         bool key_valid = false;
3005         unsigned long lowest_key = 0;
3006         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3007
3008         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3009                 /* select the one with lowest rb_key */
3010                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3011                 if (queue &&
3012                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
3013                         lowest_key = queue->rb_key;
3014                         cur_best = i;
3015                         key_valid = true;
3016                 }
3017         }
3018
3019         return cur_best;
3020 }
3021
3022 static void
3023 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3024 {
3025         unsigned slice;
3026         unsigned count;
3027         struct cfq_rb_root *st;
3028         unsigned group_slice;
3029         enum wl_class_t original_class = cfqd->serving_wl_class;
3030
3031         /* Choose next priority. RT > BE > IDLE */
3032         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3033                 cfqd->serving_wl_class = RT_WORKLOAD;
3034         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3035                 cfqd->serving_wl_class = BE_WORKLOAD;
3036         else {
3037                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3038                 cfqd->workload_expires = jiffies + 1;
3039                 return;
3040         }
3041
3042         if (original_class != cfqd->serving_wl_class)
3043                 goto new_workload;
3044
3045         /*
3046          * For RT and BE, we have to choose also the type
3047          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3048          * expiration time
3049          */
3050         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3051         count = st->count;
3052
3053         /*
3054          * check workload expiration, and that we still have other queues ready
3055          */
3056         if (count && !time_after(jiffies, cfqd->workload_expires))
3057                 return;
3058
3059 new_workload:
3060         /* otherwise select new workload type */
3061         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3062                                         cfqd->serving_wl_class);
3063         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3064         count = st->count;
3065
3066         /*
3067          * the workload slice is computed as a fraction of target latency
3068          * proportional to the number of queues in that workload, over
3069          * all the queues in the same priority class
3070          */
3071         group_slice = cfq_group_slice(cfqd, cfqg);
3072
3073         slice = group_slice * count /
3074                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3075                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3076                                         cfqg));
3077
3078         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3079                 unsigned int tmp;
3080
3081                 /*
3082                  * Async queues are currently system wide. Just taking
3083                  * proportion of queues with-in same group will lead to higher
3084                  * async ratio system wide as generally root group is going
3085                  * to have higher weight. A more accurate thing would be to
3086                  * calculate system wide asnc/sync ratio.
3087                  */
3088                 tmp = cfqd->cfq_target_latency *
3089                         cfqg_busy_async_queues(cfqd, cfqg);
3090                 tmp = tmp/cfqd->busy_queues;
3091                 slice = min_t(unsigned, slice, tmp);
3092
3093                 /* async workload slice is scaled down according to
3094                  * the sync/async slice ratio. */
3095                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3096         } else
3097                 /* sync workload slice is at least 2 * cfq_slice_idle */
3098                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3099
3100         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3101         cfq_log(cfqd, "workload slice:%d", slice);
3102         cfqd->workload_expires = jiffies + slice;
3103 }
3104
3105 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3106 {
3107         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3108         struct cfq_group *cfqg;
3109
3110         if (RB_EMPTY_ROOT(&st->rb))
3111                 return NULL;
3112         cfqg = cfq_rb_first_group(st);
3113         update_min_vdisktime(st);
3114         return cfqg;
3115 }
3116
3117 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3118 {
3119         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3120
3121         cfqd->serving_group = cfqg;
3122
3123         /* Restore the workload type data */
3124         if (cfqg->saved_wl_slice) {
3125                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3126                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3127                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3128         } else
3129                 cfqd->workload_expires = jiffies - 1;
3130
3131         choose_wl_class_and_type(cfqd, cfqg);
3132 }
3133
3134 /*
3135  * Select a queue for service. If we have a current active queue,
3136  * check whether to continue servicing it, or retrieve and set a new one.
3137  */
3138 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3139 {
3140         struct cfq_queue *cfqq, *new_cfqq = NULL;
3141
3142         cfqq = cfqd->active_queue;
3143         if (!cfqq)
3144                 goto new_queue;
3145
3146         if (!cfqd->rq_queued)
3147                 return NULL;
3148
3149         /*
3150          * We were waiting for group to get backlogged. Expire the queue
3151          */
3152         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3153                 goto expire;
3154
3155         /*
3156          * The active queue has run out of time, expire it and select new.
3157          */
3158         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3159                 /*
3160                  * If slice had not expired at the completion of last request
3161                  * we might not have turned on wait_busy flag. Don't expire
3162                  * the queue yet. Allow the group to get backlogged.
3163                  *
3164                  * The very fact that we have used the slice, that means we
3165                  * have been idling all along on this queue and it should be
3166                  * ok to wait for this request to complete.
3167                  */
3168                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3169                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3170                         cfqq = NULL;
3171                         goto keep_queue;
3172                 } else
3173                         goto check_group_idle;
3174         }
3175
3176         /*
3177          * The active queue has requests and isn't expired, allow it to
3178          * dispatch.
3179          */
3180         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3181                 goto keep_queue;
3182
3183         /*
3184          * If another queue has a request waiting within our mean seek
3185          * distance, let it run.  The expire code will check for close
3186          * cooperators and put the close queue at the front of the service
3187          * tree.  If possible, merge the expiring queue with the new cfqq.
3188          */
3189         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3190         if (new_cfqq) {
3191                 if (!cfqq->new_cfqq)
3192                         cfq_setup_merge(cfqq, new_cfqq);
3193                 goto expire;
3194         }
3195
3196         /*
3197          * No requests pending. If the active queue still has requests in
3198          * flight or is idling for a new request, allow either of these
3199          * conditions to happen (or time out) before selecting a new queue.
3200          */
3201         if (timer_pending(&cfqd->idle_slice_timer)) {
3202                 cfqq = NULL;
3203                 goto keep_queue;
3204         }
3205
3206         /*
3207          * This is a deep seek queue, but the device is much faster than
3208          * the queue can deliver, don't idle
3209          **/
3210         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3211             (cfq_cfqq_slice_new(cfqq) ||
3212             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3213                 cfq_clear_cfqq_deep(cfqq);
3214                 cfq_clear_cfqq_idle_window(cfqq);
3215         }
3216
3217         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3218                 cfqq = NULL;
3219                 goto keep_queue;
3220         }
3221
3222         /*
3223          * If group idle is enabled and there are requests dispatched from
3224          * this group, wait for requests to complete.
3225          */
3226 check_group_idle:
3227         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3228             cfqq->cfqg->dispatched &&
3229             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3230                 cfqq = NULL;
3231                 goto keep_queue;
3232         }
3233
3234 expire:
3235         cfq_slice_expired(cfqd, 0);
3236 new_queue:
3237         /*
3238          * Current queue expired. Check if we have to switch to a new
3239          * service tree
3240          */
3241         if (!new_cfqq)
3242                 cfq_choose_cfqg(cfqd);
3243
3244         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3245 keep_queue:
3246         return cfqq;
3247 }
3248
3249 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3250 {
3251         int dispatched = 0;
3252
3253         while (cfqq->next_rq) {
3254                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3255                 dispatched++;
3256         }
3257
3258         BUG_ON(!list_empty(&cfqq->fifo));
3259
3260         /* By default cfqq is not expired if it is empty. Do it explicitly */
3261         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3262         return dispatched;
3263 }
3264
3265 /*
3266  * Drain our current requests. Used for barriers and when switching
3267  * io schedulers on-the-fly.
3268  */
3269 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3270 {
3271         struct cfq_queue *cfqq;
3272         int dispatched = 0;
3273
3274         /* Expire the timeslice of the current active queue first */
3275         cfq_slice_expired(cfqd, 0);
3276         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3277                 __cfq_set_active_queue(cfqd, cfqq);
3278                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3279         }
3280
3281         BUG_ON(cfqd->busy_queues);
3282
3283         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3284         return dispatched;
3285 }
3286
3287 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3288         struct cfq_queue *cfqq)
3289 {
3290         /* the queue hasn't finished any request, can't estimate */
3291         if (cfq_cfqq_slice_new(cfqq))
3292                 return true;
3293         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3294                 cfqq->slice_end))
3295                 return true;
3296
3297         return false;
3298 }
3299
3300 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3301 {
3302         unsigned int max_dispatch;
3303
3304         /*
3305          * Drain async requests before we start sync IO
3306          */
3307         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3308                 return false;
3309
3310         /*
3311          * If this is an async queue and we have sync IO in flight, let it wait
3312          */
3313         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3314                 return false;
3315
3316         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3317         if (cfq_class_idle(cfqq))
3318                 max_dispatch = 1;
3319
3320         /*
3321          * Does this cfqq already have too much IO in flight?
3322          */
3323         if (cfqq->dispatched >= max_dispatch) {
3324                 bool promote_sync = false;
3325                 /*
3326                  * idle queue must always only have a single IO in flight
3327                  */
3328                 if (cfq_class_idle(cfqq))
3329                         return false;
3330
3331                 /*
3332                  * If there is only one sync queue
3333                  * we can ignore async queue here and give the sync
3334                  * queue no dispatch limit. The reason is a sync queue can
3335                  * preempt async queue, limiting the sync queue doesn't make
3336                  * sense. This is useful for aiostress test.
3337                  */
3338                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3339                         promote_sync = true;
3340
3341                 /*
3342                  * We have other queues, don't allow more IO from this one
3343                  */
3344                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3345                                 !promote_sync)
3346                         return false;
3347
3348                 /*
3349                  * Sole queue user, no limit
3350                  */
3351                 if (cfqd->busy_queues == 1 || promote_sync)
3352                         max_dispatch = -1;
3353                 else
3354                         /*
3355                          * Normally we start throttling cfqq when cfq_quantum/2
3356                          * requests have been dispatched. But we can drive
3357                          * deeper queue depths at the beginning of slice
3358                          * subjected to upper limit of cfq_quantum.
3359                          * */
3360                         max_dispatch = cfqd->cfq_quantum;
3361         }
3362
3363         /*
3364          * Async queues must wait a bit before being allowed dispatch.
3365          * We also ramp up the dispatch depth gradually for async IO,
3366          * based on the last sync IO we serviced
3367          */
3368         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3369                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3370                 unsigned int depth;
3371
3372                 depth = last_sync / cfqd->cfq_slice[1];
3373                 if (!depth && !cfqq->dispatched)
3374                         depth = 1;
3375                 if (depth < max_dispatch)
3376                         max_dispatch = depth;
3377         }
3378
3379         /*
3380          * If we're below the current max, allow a dispatch
3381          */
3382         return cfqq->dispatched < max_dispatch;
3383 }
3384
3385 /*
3386  * Dispatch a request from cfqq, moving them to the request queue
3387  * dispatch list.
3388  */
3389 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3390 {
3391         struct request *rq;
3392
3393         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3394
3395         if (!cfq_may_dispatch(cfqd, cfqq))
3396                 return false;
3397
3398         /*
3399          * follow expired path, else get first next available
3400          */
3401         rq = cfq_check_fifo(cfqq);
3402         if (!rq)
3403                 rq = cfqq->next_rq;
3404
3405         /*
3406          * insert request into driver dispatch list
3407          */
3408         cfq_dispatch_insert(cfqd->queue, rq);
3409
3410         if (!cfqd->active_cic) {
3411                 struct cfq_io_cq *cic = RQ_CIC(rq);
3412
3413                 atomic_long_inc(&cic->icq.ioc->refcount);
3414                 cfqd->active_cic = cic;
3415         }
3416
3417         return true;
3418 }
3419
3420 /*
3421  * Find the cfqq that we need to service and move a request from that to the
3422  * dispatch list
3423  */
3424 static int cfq_dispatch_requests(struct request_queue *q, int force)
3425 {
3426         struct cfq_data *cfqd = q->elevator->elevator_data;
3427         struct cfq_queue *cfqq;
3428
3429         if (!cfqd->busy_queues)
3430                 return 0;
3431
3432         if (unlikely(force))
3433                 return cfq_forced_dispatch(cfqd);
3434
3435         cfqq = cfq_select_queue(cfqd);
3436         if (!cfqq)
3437                 return 0;
3438
3439         /*
3440          * Dispatch a request from this cfqq, if it is allowed
3441          */
3442         if (!cfq_dispatch_request(cfqd, cfqq))
3443                 return 0;
3444
3445         cfqq->slice_dispatch++;
3446         cfq_clear_cfqq_must_dispatch(cfqq);
3447
3448         /*
3449          * expire an async queue immediately if it has used up its slice. idle
3450          * queue always expire after 1 dispatch round.
3451          */
3452         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3453             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3454             cfq_class_idle(cfqq))) {
3455                 cfqq->slice_end = jiffies + 1;
3456                 cfq_slice_expired(cfqd, 0);
3457         }
3458
3459         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3460         return 1;
3461 }
3462
3463 /*
3464  * task holds one reference to the queue, dropped when task exits. each rq
3465  * in-flight on this queue also holds a reference, dropped when rq is freed.
3466  *
3467  * Each cfq queue took a reference on the parent group. Drop it now.
3468  * queue lock must be held here.
3469  */
3470 static void cfq_put_queue(struct cfq_queue *cfqq)
3471 {
3472         struct cfq_data *cfqd = cfqq->cfqd;
3473         struct cfq_group *cfqg;
3474
3475         BUG_ON(cfqq->ref <= 0);
3476
3477         cfqq->ref--;
3478         if (cfqq->ref)
3479                 return;
3480
3481         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3482         BUG_ON(rb_first(&cfqq->sort_list));
3483         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3484         cfqg = cfqq->cfqg;
3485
3486         if (unlikely(cfqd->active_queue == cfqq)) {
3487                 __cfq_slice_expired(cfqd, cfqq, 0);
3488                 cfq_schedule_dispatch(cfqd);
3489         }
3490
3491         BUG_ON(cfq_cfqq_on_rr(cfqq));
3492         kmem_cache_free(cfq_pool, cfqq);
3493         cfqg_put(cfqg);
3494 }
3495
3496 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3497 {
3498         struct cfq_queue *__cfqq, *next;
3499
3500         /*
3501          * If this queue was scheduled to merge with another queue, be
3502          * sure to drop the reference taken on that queue (and others in
3503          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3504          */
3505         __cfqq = cfqq->new_cfqq;
3506         while (__cfqq) {
3507                 if (__cfqq == cfqq) {
3508                         WARN(1, "cfqq->new_cfqq loop detected\n");
3509                         break;
3510                 }
3511                 next = __cfqq->new_cfqq;
3512                 cfq_put_queue(__cfqq);
3513                 __cfqq = next;
3514         }
3515 }
3516
3517 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3518 {
3519         if (unlikely(cfqq == cfqd->active_queue)) {
3520                 __cfq_slice_expired(cfqd, cfqq, 0);
3521                 cfq_schedule_dispatch(cfqd);
3522         }
3523
3524         cfq_put_cooperator(cfqq);
3525
3526         cfq_put_queue(cfqq);
3527 }
3528
3529 static void cfq_init_icq(struct io_cq *icq)
3530 {
3531         struct cfq_io_cq *cic = icq_to_cic(icq);
3532
3533         cic->ttime.last_end_request = jiffies;
3534 }
3535
3536 static void cfq_exit_icq(struct io_cq *icq)
3537 {
3538         struct cfq_io_cq *cic = icq_to_cic(icq);
3539         struct cfq_data *cfqd = cic_to_cfqd(cic);
3540
3541         if (cic_to_cfqq(cic, false)) {
3542                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3543                 cic_set_cfqq(cic, NULL, false);
3544         }
3545
3546         if (cic_to_cfqq(cic, true)) {
3547                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3548                 cic_set_cfqq(cic, NULL, true);
3549         }
3550 }
3551
3552 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3553 {
3554         struct task_struct *tsk = current;
3555         int ioprio_class;
3556
3557         if (!cfq_cfqq_prio_changed(cfqq))
3558                 return;
3559
3560         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3561         switch (ioprio_class) {
3562         default:
3563                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3564         case IOPRIO_CLASS_NONE:
3565                 /*
3566                  * no prio set, inherit CPU scheduling settings
3567                  */
3568                 cfqq->ioprio = task_nice_ioprio(tsk);
3569                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3570                 break;
3571         case IOPRIO_CLASS_RT:
3572                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3573                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3574                 break;
3575         case IOPRIO_CLASS_BE:
3576                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3577                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3578                 break;
3579         case IOPRIO_CLASS_IDLE:
3580                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3581                 cfqq->ioprio = 7;
3582                 cfq_clear_cfqq_idle_window(cfqq);
3583                 break;
3584         }
3585
3586         /*
3587          * keep track of original prio settings in case we have to temporarily
3588          * elevate the priority of this queue
3589          */
3590         cfqq->org_ioprio = cfqq->ioprio;
3591         cfq_clear_cfqq_prio_changed(cfqq);
3592 }
3593
3594 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3595 {
3596         int ioprio = cic->icq.ioc->ioprio;
3597         struct cfq_data *cfqd = cic_to_cfqd(cic);
3598         struct cfq_queue *cfqq;
3599
3600         /*
3601          * Check whether ioprio has changed.  The condition may trigger
3602          * spuriously on a newly created cic but there's no harm.
3603          */
3604         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3605                 return;
3606
3607         cfqq = cic_to_cfqq(cic, false);
3608         if (cfqq) {
3609                 cfq_put_queue(cfqq);
3610                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3611                 cic_set_cfqq(cic, cfqq, false);
3612         }
3613
3614         cfqq = cic_to_cfqq(cic, true);
3615         if (cfqq)
3616                 cfq_mark_cfqq_prio_changed(cfqq);
3617
3618         cic->ioprio = ioprio;
3619 }
3620
3621 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3622                           pid_t pid, bool is_sync)
3623 {
3624         RB_CLEAR_NODE(&cfqq->rb_node);
3625         RB_CLEAR_NODE(&cfqq->p_node);
3626         INIT_LIST_HEAD(&cfqq->fifo);
3627
3628         cfqq->ref = 0;
3629         cfqq->cfqd = cfqd;
3630
3631         cfq_mark_cfqq_prio_changed(cfqq);
3632
3633         if (is_sync) {
3634                 if (!cfq_class_idle(cfqq))
3635                         cfq_mark_cfqq_idle_window(cfqq);
3636                 cfq_mark_cfqq_sync(cfqq);
3637         }
3638         cfqq->pid = pid;
3639 }
3640
3641 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3642 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3643 {
3644         struct cfq_data *cfqd = cic_to_cfqd(cic);
3645         struct cfq_queue *cfqq;
3646         uint64_t serial_nr;
3647
3648         rcu_read_lock();
3649         serial_nr = bio_blkcg(bio)->css.serial_nr;
3650         rcu_read_unlock();
3651
3652         /*
3653          * Check whether blkcg has changed.  The condition may trigger
3654          * spuriously on a newly created cic but there's no harm.
3655          */
3656         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3657                 return;
3658
3659         /*
3660          * Drop reference to queues.  New queues will be assigned in new
3661          * group upon arrival of fresh requests.
3662          */
3663         cfqq = cic_to_cfqq(cic, false);
3664         if (cfqq) {
3665                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3666                 cic_set_cfqq(cic, NULL, false);
3667                 cfq_put_queue(cfqq);
3668         }
3669
3670         cfqq = cic_to_cfqq(cic, true);
3671         if (cfqq) {
3672                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3673                 cic_set_cfqq(cic, NULL, true);
3674                 cfq_put_queue(cfqq);
3675         }
3676
3677         cic->blkcg_serial_nr = serial_nr;
3678 }
3679 #else
3680 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3681 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3682
3683 static struct cfq_queue **
3684 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3685 {
3686         switch (ioprio_class) {
3687         case IOPRIO_CLASS_RT:
3688                 return &cfqg->async_cfqq[0][ioprio];
3689         case IOPRIO_CLASS_NONE:
3690                 ioprio = IOPRIO_NORM;
3691                 /* fall through */
3692         case IOPRIO_CLASS_BE:
3693                 return &cfqg->async_cfqq[1][ioprio];
3694         case IOPRIO_CLASS_IDLE:
3695                 return &cfqg->async_idle_cfqq;
3696         default:
3697                 BUG();
3698         }
3699 }
3700
3701 static struct cfq_queue *
3702 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3703               struct bio *bio)
3704 {
3705         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3706         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3707         struct cfq_queue **async_cfqq = NULL;
3708         struct cfq_queue *cfqq;
3709         struct cfq_group *cfqg;
3710
3711         rcu_read_lock();
3712         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3713         if (!cfqg) {
3714                 cfqq = &cfqd->oom_cfqq;
3715                 goto out;
3716         }
3717
3718         if (!is_sync) {
3719                 if (!ioprio_valid(cic->ioprio)) {
3720                         struct task_struct *tsk = current;
3721                         ioprio = task_nice_ioprio(tsk);
3722                         ioprio_class = task_nice_ioclass(tsk);
3723                 }
3724                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3725                 cfqq = *async_cfqq;
3726                 if (cfqq)
3727                         goto out;
3728         }
3729
3730         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3731                                      cfqd->queue->node);
3732         if (!cfqq) {
3733                 cfqq = &cfqd->oom_cfqq;
3734                 goto out;
3735         }
3736
3737         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3738         cfq_init_prio_data(cfqq, cic);
3739         cfq_link_cfqq_cfqg(cfqq, cfqg);
3740         cfq_log_cfqq(cfqd, cfqq, "alloced");
3741
3742         if (async_cfqq) {
3743                 /* a new async queue is created, pin and remember */
3744                 cfqq->ref++;
3745                 *async_cfqq = cfqq;
3746         }
3747 out:
3748         cfqq->ref++;
3749         rcu_read_unlock();
3750         return cfqq;
3751 }
3752
3753 static void
3754 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3755 {
3756         unsigned long elapsed = jiffies - ttime->last_end_request;
3757         elapsed = min(elapsed, 2UL * slice_idle);
3758
3759         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3760         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3761         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3762 }
3763
3764 static void
3765 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3766                         struct cfq_io_cq *cic)
3767 {
3768         if (cfq_cfqq_sync(cfqq)) {
3769                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3770                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3771                         cfqd->cfq_slice_idle);
3772         }
3773 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3774         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3775 #endif
3776 }
3777
3778 static void
3779 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3780                        struct request *rq)
3781 {
3782         sector_t sdist = 0;
3783         sector_t n_sec = blk_rq_sectors(rq);
3784         if (cfqq->last_request_pos) {
3785                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3786                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3787                 else
3788                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3789         }
3790
3791         cfqq->seek_history <<= 1;
3792         if (blk_queue_nonrot(cfqd->queue))
3793                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3794         else
3795                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3796 }
3797
3798 /*
3799  * Disable idle window if the process thinks too long or seeks so much that
3800  * it doesn't matter
3801  */
3802 static void
3803 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3804                        struct cfq_io_cq *cic)
3805 {
3806         int old_idle, enable_idle;
3807
3808         /*
3809          * Don't idle for async or idle io prio class
3810          */
3811         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3812                 return;
3813
3814         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3815
3816         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3817                 cfq_mark_cfqq_deep(cfqq);
3818
3819         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3820                 enable_idle = 0;
3821         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3822                  !cfqd->cfq_slice_idle ||
3823                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3824                 enable_idle = 0;
3825         else if (sample_valid(cic->ttime.ttime_samples)) {
3826                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3827                         enable_idle = 0;
3828                 else
3829                         enable_idle = 1;
3830         }
3831
3832         if (old_idle != enable_idle) {
3833                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3834                 if (enable_idle)
3835                         cfq_mark_cfqq_idle_window(cfqq);
3836                 else
3837                         cfq_clear_cfqq_idle_window(cfqq);
3838         }
3839 }
3840
3841 /*
3842  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3843  * no or if we aren't sure, a 1 will cause a preempt.
3844  */
3845 static bool
3846 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3847                    struct request *rq)
3848 {
3849         struct cfq_queue *cfqq;
3850
3851         cfqq = cfqd->active_queue;
3852         if (!cfqq)
3853                 return false;
3854
3855         if (cfq_class_idle(new_cfqq))
3856                 return false;
3857
3858         if (cfq_class_idle(cfqq))
3859                 return true;
3860
3861         /*
3862          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3863          */
3864         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3865                 return false;
3866
3867         /*
3868          * if the new request is sync, but the currently running queue is
3869          * not, let the sync request have priority.
3870          */
3871         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3872                 return true;
3873
3874         if (new_cfqq->cfqg != cfqq->cfqg)
3875                 return false;
3876
3877         if (cfq_slice_used(cfqq))
3878                 return true;
3879
3880         /* Allow preemption only if we are idling on sync-noidle tree */
3881         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3882             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3883             new_cfqq->service_tree->count == 2 &&
3884             RB_EMPTY_ROOT(&cfqq->sort_list))
3885                 return true;
3886
3887         /*
3888          * So both queues are sync. Let the new request get disk time if
3889          * it's a metadata request and the current queue is doing regular IO.
3890          */
3891         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3892                 return true;
3893
3894         /*
3895          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3896          */
3897         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3898                 return true;
3899
3900         /* An idle queue should not be idle now for some reason */
3901         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3902                 return true;
3903
3904         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3905                 return false;
3906
3907         /*
3908          * if this request is as-good as one we would expect from the
3909          * current cfqq, let it preempt
3910          */
3911         if (cfq_rq_close(cfqd, cfqq, rq))
3912                 return true;
3913
3914         return false;
3915 }
3916
3917 /*
3918  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3919  * let it have half of its nominal slice.
3920  */
3921 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3922 {
3923         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3924
3925         cfq_log_cfqq(cfqd, cfqq, "preempt");
3926         cfq_slice_expired(cfqd, 1);
3927
3928         /*
3929          * workload type is changed, don't save slice, otherwise preempt
3930          * doesn't happen
3931          */
3932         if (old_type != cfqq_type(cfqq))
3933                 cfqq->cfqg->saved_wl_slice = 0;
3934
3935         /*
3936          * Put the new queue at the front of the of the current list,
3937          * so we know that it will be selected next.
3938          */
3939         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3940
3941         cfq_service_tree_add(cfqd, cfqq, 1);
3942
3943         cfqq->slice_end = 0;
3944         cfq_mark_cfqq_slice_new(cfqq);
3945 }
3946
3947 /*
3948  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3949  * something we should do about it
3950  */
3951 static void
3952 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3953                 struct request *rq)
3954 {
3955         struct cfq_io_cq *cic = RQ_CIC(rq);
3956
3957         cfqd->rq_queued++;
3958         if (rq->cmd_flags & REQ_PRIO)
3959                 cfqq->prio_pending++;
3960
3961         cfq_update_io_thinktime(cfqd, cfqq, cic);
3962         cfq_update_io_seektime(cfqd, cfqq, rq);
3963         cfq_update_idle_window(cfqd, cfqq, cic);
3964
3965         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3966
3967         if (cfqq == cfqd->active_queue) {
3968                 /*
3969                  * Remember that we saw a request from this process, but
3970                  * don't start queuing just yet. Otherwise we risk seeing lots
3971                  * of tiny requests, because we disrupt the normal plugging
3972                  * and merging. If the request is already larger than a single
3973                  * page, let it rip immediately. For that case we assume that
3974                  * merging is already done. Ditto for a busy system that
3975                  * has other work pending, don't risk delaying until the
3976                  * idle timer unplug to continue working.
3977                  */
3978                 if (cfq_cfqq_wait_request(cfqq)) {
3979                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3980                             cfqd->busy_queues > 1) {
3981                                 cfq_del_timer(cfqd, cfqq);
3982                                 cfq_clear_cfqq_wait_request(cfqq);
3983                                 __blk_run_queue(cfqd->queue);
3984                         } else {
3985                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3986                                 cfq_mark_cfqq_must_dispatch(cfqq);
3987                         }
3988                 }
3989         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3990                 /*
3991                  * not the active queue - expire current slice if it is
3992                  * idle and has expired it's mean thinktime or this new queue
3993                  * has some old slice time left and is of higher priority or
3994                  * this new queue is RT and the current one is BE
3995                  */
3996                 cfq_preempt_queue(cfqd, cfqq);
3997                 __blk_run_queue(cfqd->queue);
3998         }
3999 }
4000
4001 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4002 {
4003         struct cfq_data *cfqd = q->elevator->elevator_data;
4004         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4005
4006         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4007         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4008
4009         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4010         list_add_tail(&rq->queuelist, &cfqq->fifo);
4011         cfq_add_rq_rb(rq);
4012         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4013                                  rq->cmd_flags);
4014         cfq_rq_enqueued(cfqd, cfqq, rq);
4015 }
4016
4017 /*
4018  * Update hw_tag based on peak queue depth over 50 samples under
4019  * sufficient load.
4020  */
4021 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4022 {
4023         struct cfq_queue *cfqq = cfqd->active_queue;
4024
4025         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4026                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4027
4028         if (cfqd->hw_tag == 1)
4029                 return;
4030
4031         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4032             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4033                 return;
4034
4035         /*
4036          * If active queue hasn't enough requests and can idle, cfq might not
4037          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4038          * case
4039          */
4040         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4041             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4042             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4043                 return;
4044
4045         if (cfqd->hw_tag_samples++ < 50)
4046                 return;
4047
4048         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4049                 cfqd->hw_tag = 1;
4050         else
4051                 cfqd->hw_tag = 0;
4052 }
4053
4054 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4055 {
4056         struct cfq_io_cq *cic = cfqd->active_cic;
4057
4058         /* If the queue already has requests, don't wait */
4059         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4060                 return false;
4061
4062         /* If there are other queues in the group, don't wait */
4063         if (cfqq->cfqg->nr_cfqq > 1)
4064                 return false;
4065
4066         /* the only queue in the group, but think time is big */
4067         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4068                 return false;
4069
4070         if (cfq_slice_used(cfqq))
4071                 return true;
4072
4073         /* if slice left is less than think time, wait busy */
4074         if (cic && sample_valid(cic->ttime.ttime_samples)
4075             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4076                 return true;
4077
4078         /*
4079          * If think times is less than a jiffy than ttime_mean=0 and above
4080          * will not be true. It might happen that slice has not expired yet
4081          * but will expire soon (4-5 ns) during select_queue(). To cover the
4082          * case where think time is less than a jiffy, mark the queue wait
4083          * busy if only 1 jiffy is left in the slice.
4084          */
4085         if (cfqq->slice_end - jiffies == 1)
4086                 return true;
4087
4088         return false;
4089 }
4090
4091 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4092 {
4093         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4094         struct cfq_data *cfqd = cfqq->cfqd;
4095         const int sync = rq_is_sync(rq);
4096         unsigned long now;
4097
4098         now = jiffies;
4099         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4100                      !!(rq->cmd_flags & REQ_NOIDLE));
4101
4102         cfq_update_hw_tag(cfqd);
4103
4104         WARN_ON(!cfqd->rq_in_driver);
4105         WARN_ON(!cfqq->dispatched);
4106         cfqd->rq_in_driver--;
4107         cfqq->dispatched--;
4108         (RQ_CFQG(rq))->dispatched--;
4109         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4110                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4111
4112         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4113
4114         if (sync) {
4115                 struct cfq_rb_root *st;
4116
4117                 RQ_CIC(rq)->ttime.last_end_request = now;
4118
4119                 if (cfq_cfqq_on_rr(cfqq))
4120                         st = cfqq->service_tree;
4121                 else
4122                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4123                                         cfqq_type(cfqq));
4124
4125                 st->ttime.last_end_request = now;
4126                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4127                         cfqd->last_delayed_sync = now;
4128         }
4129
4130 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4131         cfqq->cfqg->ttime.last_end_request = now;
4132 #endif
4133
4134         /*
4135          * If this is the active queue, check if it needs to be expired,
4136          * or if we want to idle in case it has no pending requests.
4137          */
4138         if (cfqd->active_queue == cfqq) {
4139                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4140
4141                 if (cfq_cfqq_slice_new(cfqq)) {
4142                         cfq_set_prio_slice(cfqd, cfqq);
4143                         cfq_clear_cfqq_slice_new(cfqq);
4144                 }
4145
4146                 /*
4147                  * Should we wait for next request to come in before we expire
4148                  * the queue.
4149                  */
4150                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4151                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4152                         if (!cfqd->cfq_slice_idle)
4153                                 extend_sl = cfqd->cfq_group_idle;
4154                         cfqq->slice_end = jiffies + extend_sl;
4155                         cfq_mark_cfqq_wait_busy(cfqq);
4156                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4157                 }
4158
4159                 /*
4160                  * Idling is not enabled on:
4161                  * - expired queues
4162                  * - idle-priority queues
4163                  * - async queues
4164                  * - queues with still some requests queued
4165                  * - when there is a close cooperator
4166                  */
4167                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4168                         cfq_slice_expired(cfqd, 1);
4169                 else if (sync && cfqq_empty &&
4170                          !cfq_close_cooperator(cfqd, cfqq)) {
4171                         cfq_arm_slice_timer(cfqd);
4172                 }
4173         }
4174
4175         if (!cfqd->rq_in_driver)
4176                 cfq_schedule_dispatch(cfqd);
4177 }
4178
4179 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4180 {
4181         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4182                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4183                 return ELV_MQUEUE_MUST;
4184         }
4185
4186         return ELV_MQUEUE_MAY;
4187 }
4188
4189 static int cfq_may_queue(struct request_queue *q, int rw)
4190 {
4191         struct cfq_data *cfqd = q->elevator->elevator_data;
4192         struct task_struct *tsk = current;
4193         struct cfq_io_cq *cic;
4194         struct cfq_queue *cfqq;
4195
4196         /*
4197          * don't force setup of a queue from here, as a call to may_queue
4198          * does not necessarily imply that a request actually will be queued.
4199          * so just lookup a possibly existing queue, or return 'may queue'
4200          * if that fails
4201          */
4202         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4203         if (!cic)
4204                 return ELV_MQUEUE_MAY;
4205
4206         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4207         if (cfqq) {
4208                 cfq_init_prio_data(cfqq, cic);
4209
4210                 return __cfq_may_queue(cfqq);
4211         }
4212
4213         return ELV_MQUEUE_MAY;
4214 }
4215
4216 /*
4217  * queue lock held here
4218  */
4219 static void cfq_put_request(struct request *rq)
4220 {
4221         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4222
4223         if (cfqq) {
4224                 const int rw = rq_data_dir(rq);
4225
4226                 BUG_ON(!cfqq->allocated[rw]);
4227                 cfqq->allocated[rw]--;
4228
4229                 /* Put down rq reference on cfqg */
4230                 cfqg_put(RQ_CFQG(rq));
4231                 rq->elv.priv[0] = NULL;
4232                 rq->elv.priv[1] = NULL;
4233
4234                 cfq_put_queue(cfqq);
4235         }
4236 }
4237
4238 static struct cfq_queue *
4239 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4240                 struct cfq_queue *cfqq)
4241 {
4242         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4243         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4244         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4245         cfq_put_queue(cfqq);
4246         return cic_to_cfqq(cic, 1);
4247 }
4248
4249 /*
4250  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4251  * was the last process referring to said cfqq.
4252  */
4253 static struct cfq_queue *
4254 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4255 {
4256         if (cfqq_process_refs(cfqq) == 1) {
4257                 cfqq->pid = current->pid;
4258                 cfq_clear_cfqq_coop(cfqq);
4259                 cfq_clear_cfqq_split_coop(cfqq);
4260                 return cfqq;
4261         }
4262
4263         cic_set_cfqq(cic, NULL, 1);
4264
4265         cfq_put_cooperator(cfqq);
4266
4267         cfq_put_queue(cfqq);
4268         return NULL;
4269 }
4270 /*
4271  * Allocate cfq data structures associated with this request.
4272  */
4273 static int
4274 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4275                 gfp_t gfp_mask)
4276 {
4277         struct cfq_data *cfqd = q->elevator->elevator_data;
4278         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4279         const int rw = rq_data_dir(rq);
4280         const bool is_sync = rq_is_sync(rq);
4281         struct cfq_queue *cfqq;
4282
4283         spin_lock_irq(q->queue_lock);
4284
4285         check_ioprio_changed(cic, bio);
4286         check_blkcg_changed(cic, bio);
4287 new_queue:
4288         cfqq = cic_to_cfqq(cic, is_sync);
4289         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4290                 if (cfqq)
4291                         cfq_put_queue(cfqq);
4292                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4293                 cic_set_cfqq(cic, cfqq, is_sync);
4294         } else {
4295                 /*
4296                  * If the queue was seeky for too long, break it apart.
4297                  */
4298                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4299                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4300                         cfqq = split_cfqq(cic, cfqq);
4301                         if (!cfqq)
4302                                 goto new_queue;
4303                 }
4304
4305                 /*
4306                  * Check to see if this queue is scheduled to merge with
4307                  * another, closely cooperating queue.  The merging of
4308                  * queues happens here as it must be done in process context.
4309                  * The reference on new_cfqq was taken in merge_cfqqs.
4310                  */
4311                 if (cfqq->new_cfqq)
4312                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4313         }
4314
4315         cfqq->allocated[rw]++;
4316
4317         cfqq->ref++;
4318         cfqg_get(cfqq->cfqg);
4319         rq->elv.priv[0] = cfqq;
4320         rq->elv.priv[1] = cfqq->cfqg;
4321         spin_unlock_irq(q->queue_lock);
4322         return 0;
4323 }
4324
4325 static void cfq_kick_queue(struct work_struct *work)
4326 {
4327         struct cfq_data *cfqd =
4328                 container_of(work, struct cfq_data, unplug_work);
4329         struct request_queue *q = cfqd->queue;
4330
4331         spin_lock_irq(q->queue_lock);
4332         __blk_run_queue(cfqd->queue);
4333         spin_unlock_irq(q->queue_lock);
4334 }
4335
4336 /*
4337  * Timer running if the active_queue is currently idling inside its time slice
4338  */
4339 static void cfq_idle_slice_timer(unsigned long data)
4340 {
4341         struct cfq_data *cfqd = (struct cfq_data *) data;
4342         struct cfq_queue *cfqq;
4343         unsigned long flags;
4344         int timed_out = 1;
4345
4346         cfq_log(cfqd, "idle timer fired");
4347
4348         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4349
4350         cfqq = cfqd->active_queue;
4351         if (cfqq) {
4352                 timed_out = 0;
4353
4354                 /*
4355                  * We saw a request before the queue expired, let it through
4356                  */
4357                 if (cfq_cfqq_must_dispatch(cfqq))
4358                         goto out_kick;
4359
4360                 /*
4361                  * expired
4362                  */
4363                 if (cfq_slice_used(cfqq))
4364                         goto expire;
4365
4366                 /*
4367                  * only expire and reinvoke request handler, if there are
4368                  * other queues with pending requests
4369                  */
4370                 if (!cfqd->busy_queues)
4371                         goto out_cont;
4372
4373                 /*
4374                  * not expired and it has a request pending, let it dispatch
4375                  */
4376                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4377                         goto out_kick;
4378
4379                 /*
4380                  * Queue depth flag is reset only when the idle didn't succeed
4381                  */
4382                 cfq_clear_cfqq_deep(cfqq);
4383         }
4384 expire:
4385         cfq_slice_expired(cfqd, timed_out);
4386 out_kick:
4387         cfq_schedule_dispatch(cfqd);
4388 out_cont:
4389         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4390 }
4391
4392 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4393 {
4394         del_timer_sync(&cfqd->idle_slice_timer);
4395         cancel_work_sync(&cfqd->unplug_work);
4396 }
4397
4398 static void cfq_exit_queue(struct elevator_queue *e)
4399 {
4400         struct cfq_data *cfqd = e->elevator_data;
4401         struct request_queue *q = cfqd->queue;
4402
4403         cfq_shutdown_timer_wq(cfqd);
4404
4405         spin_lock_irq(q->queue_lock);
4406
4407         if (cfqd->active_queue)
4408                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4409
4410         spin_unlock_irq(q->queue_lock);
4411
4412         cfq_shutdown_timer_wq(cfqd);
4413
4414 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4415         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4416 #else
4417         kfree(cfqd->root_group);
4418 #endif
4419         kfree(cfqd);
4420 }
4421
4422 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4423 {
4424         struct cfq_data *cfqd;
4425         struct blkcg_gq *blkg __maybe_unused;
4426         int i, ret;
4427         struct elevator_queue *eq;
4428
4429         eq = elevator_alloc(q, e);
4430         if (!eq)
4431                 return -ENOMEM;
4432
4433         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4434         if (!cfqd) {
4435                 kobject_put(&eq->kobj);
4436                 return -ENOMEM;
4437         }
4438         eq->elevator_data = cfqd;
4439
4440         cfqd->queue = q;
4441         spin_lock_irq(q->queue_lock);
4442         q->elevator = eq;
4443         spin_unlock_irq(q->queue_lock);
4444
4445         /* Init root service tree */
4446         cfqd->grp_service_tree = CFQ_RB_ROOT;
4447
4448         /* Init root group and prefer root group over other groups by default */
4449 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4450         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4451         if (ret)
4452                 goto out_free;
4453
4454         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4455 #else
4456         ret = -ENOMEM;
4457         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4458                                         GFP_KERNEL, cfqd->queue->node);
4459         if (!cfqd->root_group)
4460                 goto out_free;
4461
4462         cfq_init_cfqg_base(cfqd->root_group);
4463 #endif
4464         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4465         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4466
4467         /*
4468          * Not strictly needed (since RB_ROOT just clears the node and we
4469          * zeroed cfqd on alloc), but better be safe in case someone decides
4470          * to add magic to the rb code
4471          */
4472         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4473                 cfqd->prio_trees[i] = RB_ROOT;
4474
4475         /*
4476          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4477          * Grab a permanent reference to it, so that the normal code flow
4478          * will not attempt to free it.  oom_cfqq is linked to root_group
4479          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4480          * the reference from linking right away.
4481          */
4482         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4483         cfqd->oom_cfqq.ref++;
4484
4485         spin_lock_irq(q->queue_lock);
4486         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4487         cfqg_put(cfqd->root_group);
4488         spin_unlock_irq(q->queue_lock);
4489
4490         init_timer(&cfqd->idle_slice_timer);
4491         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4492         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4493
4494         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4495
4496         cfqd->cfq_quantum = cfq_quantum;
4497         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4498         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4499         cfqd->cfq_back_max = cfq_back_max;
4500         cfqd->cfq_back_penalty = cfq_back_penalty;
4501         cfqd->cfq_slice[0] = cfq_slice_async;
4502         cfqd->cfq_slice[1] = cfq_slice_sync;
4503         cfqd->cfq_target_latency = cfq_target_latency;
4504         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4505         cfqd->cfq_slice_idle = cfq_slice_idle;
4506         cfqd->cfq_group_idle = cfq_group_idle;
4507         cfqd->cfq_latency = 1;
4508         cfqd->hw_tag = -1;
4509         /*
4510          * we optimistically start assuming sync ops weren't delayed in last
4511          * second, in order to have larger depth for async operations.
4512          */
4513         cfqd->last_delayed_sync = jiffies - HZ;
4514         return 0;
4515
4516 out_free:
4517         kfree(cfqd);
4518         kobject_put(&eq->kobj);
4519         return ret;
4520 }
4521
4522 static void cfq_registered_queue(struct request_queue *q)
4523 {
4524         struct elevator_queue *e = q->elevator;
4525         struct cfq_data *cfqd = e->elevator_data;
4526
4527         /*
4528          * Default to IOPS mode with no idling for SSDs
4529          */
4530         if (blk_queue_nonrot(q))
4531                 cfqd->cfq_slice_idle = 0;
4532 }
4533
4534 /*
4535  * sysfs parts below -->
4536  */
4537 static ssize_t
4538 cfq_var_show(unsigned int var, char *page)
4539 {
4540         return sprintf(page, "%u\n", var);
4541 }
4542
4543 static ssize_t
4544 cfq_var_store(unsigned int *var, const char *page, size_t count)
4545 {
4546         char *p = (char *) page;
4547
4548         *var = simple_strtoul(p, &p, 10);
4549         return count;
4550 }
4551
4552 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4553 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4554 {                                                                       \
4555         struct cfq_data *cfqd = e->elevator_data;                       \
4556         unsigned int __data = __VAR;                                    \
4557         if (__CONV)                                                     \
4558                 __data = jiffies_to_msecs(__data);                      \
4559         return cfq_var_show(__data, (page));                            \
4560 }
4561 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4562 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4563 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4564 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4565 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4566 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4567 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4568 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4569 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4570 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4571 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4572 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4573 #undef SHOW_FUNCTION
4574
4575 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4576 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4577 {                                                                       \
4578         struct cfq_data *cfqd = e->elevator_data;                       \
4579         unsigned int __data;                                            \
4580         int ret = cfq_var_store(&__data, (page), count);                \
4581         if (__data < (MIN))                                             \
4582                 __data = (MIN);                                         \
4583         else if (__data > (MAX))                                        \
4584                 __data = (MAX);                                         \
4585         if (__CONV)                                                     \
4586                 *(__PTR) = msecs_to_jiffies(__data);                    \
4587         else                                                            \
4588                 *(__PTR) = __data;                                      \
4589         return ret;                                                     \
4590 }
4591 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4592 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4593                 UINT_MAX, 1);
4594 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4595                 UINT_MAX, 1);
4596 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4597 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4598                 UINT_MAX, 0);
4599 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4600 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4601 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4602 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4603 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4604                 UINT_MAX, 0);
4605 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4606 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4607 #undef STORE_FUNCTION
4608
4609 #define CFQ_ATTR(name) \
4610         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4611
4612 static struct elv_fs_entry cfq_attrs[] = {
4613         CFQ_ATTR(quantum),
4614         CFQ_ATTR(fifo_expire_sync),
4615         CFQ_ATTR(fifo_expire_async),
4616         CFQ_ATTR(back_seek_max),
4617         CFQ_ATTR(back_seek_penalty),
4618         CFQ_ATTR(slice_sync),
4619         CFQ_ATTR(slice_async),
4620         CFQ_ATTR(slice_async_rq),
4621         CFQ_ATTR(slice_idle),
4622         CFQ_ATTR(group_idle),
4623         CFQ_ATTR(low_latency),
4624         CFQ_ATTR(target_latency),
4625         __ATTR_NULL
4626 };
4627
4628 static struct elevator_type iosched_cfq = {
4629         .ops = {
4630                 .elevator_merge_fn =            cfq_merge,
4631                 .elevator_merged_fn =           cfq_merged_request,
4632                 .elevator_merge_req_fn =        cfq_merged_requests,
4633                 .elevator_allow_merge_fn =      cfq_allow_merge,
4634                 .elevator_bio_merged_fn =       cfq_bio_merged,
4635                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4636                 .elevator_add_req_fn =          cfq_insert_request,
4637                 .elevator_activate_req_fn =     cfq_activate_request,
4638                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4639                 .elevator_completed_req_fn =    cfq_completed_request,
4640                 .elevator_former_req_fn =       elv_rb_former_request,
4641                 .elevator_latter_req_fn =       elv_rb_latter_request,
4642                 .elevator_init_icq_fn =         cfq_init_icq,
4643                 .elevator_exit_icq_fn =         cfq_exit_icq,
4644                 .elevator_set_req_fn =          cfq_set_request,
4645                 .elevator_put_req_fn =          cfq_put_request,
4646                 .elevator_may_queue_fn =        cfq_may_queue,
4647                 .elevator_init_fn =             cfq_init_queue,
4648                 .elevator_exit_fn =             cfq_exit_queue,
4649                 .elevator_registered_fn =       cfq_registered_queue,
4650         },
4651         .icq_size       =       sizeof(struct cfq_io_cq),
4652         .icq_align      =       __alignof__(struct cfq_io_cq),
4653         .elevator_attrs =       cfq_attrs,
4654         .elevator_name  =       "cfq",
4655         .elevator_owner =       THIS_MODULE,
4656 };
4657
4658 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4659 static struct blkcg_policy blkcg_policy_cfq = {
4660         .cftypes                = cfq_blkcg_files,
4661
4662         .cpd_alloc_fn           = cfq_cpd_alloc,
4663         .cpd_init_fn            = cfq_cpd_init,
4664         .cpd_free_fn            = cfq_cpd_free,
4665
4666         .pd_alloc_fn            = cfq_pd_alloc,
4667         .pd_init_fn             = cfq_pd_init,
4668         .pd_offline_fn          = cfq_pd_offline,
4669         .pd_free_fn             = cfq_pd_free,
4670         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4671 };
4672 #endif
4673
4674 static int __init cfq_init(void)
4675 {
4676         int ret;
4677
4678         /*
4679          * could be 0 on HZ < 1000 setups
4680          */
4681         if (!cfq_slice_async)
4682                 cfq_slice_async = 1;
4683         if (!cfq_slice_idle)
4684                 cfq_slice_idle = 1;
4685
4686 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4687         if (!cfq_group_idle)
4688                 cfq_group_idle = 1;
4689
4690         ret = blkcg_policy_register(&blkcg_policy_cfq);
4691         if (ret)
4692                 return ret;
4693 #else
4694         cfq_group_idle = 0;
4695 #endif
4696
4697         ret = -ENOMEM;
4698         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4699         if (!cfq_pool)
4700                 goto err_pol_unreg;
4701
4702         ret = elv_register(&iosched_cfq);
4703         if (ret)
4704                 goto err_free_pool;
4705
4706         return 0;
4707
4708 err_free_pool:
4709         kmem_cache_destroy(cfq_pool);
4710 err_pol_unreg:
4711 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4712         blkcg_policy_unregister(&blkcg_policy_cfq);
4713 #endif
4714         return ret;
4715 }
4716
4717 static void __exit cfq_exit(void)
4718 {
4719 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4720         blkcg_policy_unregister(&blkcg_policy_cfq);
4721 #endif
4722         elv_unregister(&iosched_cfq);
4723         kmem_cache_destroy(cfq_pool);
4724 }
4725
4726 module_init(cfq_init);
4727 module_exit(cfq_exit);
4728
4729 MODULE_AUTHOR("Jens Axboe");
4730 MODULE_LICENSE("GPL");
4731 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");