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