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 cpd;
 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         /* async queue for each priority case */
 310         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 311         struct cfq_queue *async_idle_cfqq;
 312
 313 };
 314
 315 struct cfq_io_cq {
 316         struct io_cq            icq;            /* must be the first member */
 317         struct cfq_queue        *cfqq[2];
 318         struct cfq_ttime        ttime;
 319         int                     ioprio;         /* the current ioprio */
 320 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 321         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 322 #endif
 323 };
 324
 325 /*
 326  * Per block device queue structure
 327  */
 328 struct cfq_data {
 329         struct request_queue *queue;
 330         /* Root service tree for cfq_groups */
 331         struct cfq_rb_root grp_service_tree;
 332         struct cfq_group *root_group;
 333
 334         /*
 335          * The priority currently being served
 336          */
 337         enum wl_class_t serving_wl_class;
 338         enum wl_type_t serving_wl_type;
 339         unsigned long workload_expires;
 340         struct cfq_group *serving_group;
 341
 342         /*
 343          * Each priority tree is sorted by next_request position.  These
 344          * trees are used when determining if two or more queues are
 345          * interleaving requests (see cfq_close_cooperator).
 346          */
 347         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 348
 349         unsigned int busy_queues;
 350         unsigned int busy_sync_queues;
 351
 352         int rq_in_driver;
 353         int rq_in_flight[2];
 354
 355         /*
 356          * queue-depth detection
 357          */
 358         int rq_queued;
 359         int hw_tag;
 360         /*
 361          * hw_tag can be
 362          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 363          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 364          *  0 => no NCQ
 365          */
 366         int hw_tag_est_depth;
 367         unsigned int hw_tag_samples;
 368
 369         /*
 370          * idle window management
 371          */
 372         struct timer_list idle_slice_timer;
 373         struct work_struct unplug_work;
 374
 375         struct cfq_queue *active_queue;
 376         struct cfq_io_cq *active_cic;
 377
 378         sector_t last_position;
 379
 380         /*
 381          * tunables, see top of file
 382          */
 383         unsigned int cfq_quantum;
 384         unsigned int cfq_fifo_expire[2];
 385         unsigned int cfq_back_penalty;
 386         unsigned int cfq_back_max;
 387         unsigned int cfq_slice[2];
 388         unsigned int cfq_slice_async_rq;
 389         unsigned int cfq_slice_idle;
 390         unsigned int cfq_group_idle;
 391         unsigned int cfq_latency;
 392         unsigned int cfq_target_latency;
 393
 394         /*
 395          * Fallback dummy cfqq for extreme OOM conditions
 396          */
 397         struct cfq_queue oom_cfqq;
 398
 399         unsigned long last_delayed_sync;
 400 };
 401
 402 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 403 static void cfq_put_queue(struct cfq_queue *cfqq);
 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, cpd) : 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(struct blkcg_policy_data *cpd)
1572 {
1573         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1574
1575         if (cpd_to_blkcg(cpd) == &blkcg_root) {
1576                 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1577                 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1578         } else {
1579                 cgd->weight = CFQ_WEIGHT_DEFAULT;
1580                 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1581         }
1582 }
1583
1584 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1585 {
1586         struct cfq_group *cfqg;
1587
1588         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1589         if (!cfqg)
1590                 return NULL;
1591
1592         cfq_init_cfqg_base(cfqg);
1593         cfqg_stats_init(&cfqg->stats);
1594         cfqg_stats_init(&cfqg->dead_stats);
1595
1596         return &cfqg->pd;
1597 }
1598
1599 static void cfq_pd_init(struct blkg_policy_data *pd)
1600 {
1601         struct cfq_group *cfqg = pd_to_cfqg(pd);
1602         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1603
1604         cfqg->weight = cgd->weight;
1605         cfqg->leaf_weight = cgd->leaf_weight;
1606 }
1607
1608 static void cfq_pd_offline(struct blkg_policy_data *pd)
1609 {
1610         struct cfq_group *cfqg = pd_to_cfqg(pd);
1611         int i;
1612
1613         for (i = 0; i < IOPRIO_BE_NR; i++) {
1614                 if (cfqg->async_cfqq[0][i])
1615                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1616                 if (cfqg->async_cfqq[1][i])
1617                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1618         }
1619
1620         if (cfqg->async_idle_cfqq)
1621                 cfq_put_queue(cfqg->async_idle_cfqq);
1622
1623         /*
1624          * @blkg is going offline and will be ignored by
1625          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1626          * that they don't get lost.  If IOs complete after this point, the
1627          * stats for them will be lost.  Oh well...
1628          */
1629         cfqg_stats_xfer_dead(cfqg);
1630 }
1631
1632 static void cfq_pd_free(struct blkg_policy_data *pd)
1633 {
1634         return kfree(pd);
1635 }
1636
1637 /* offset delta from cfqg->stats to cfqg->dead_stats */
1638 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1639                                         offsetof(struct cfq_group, stats);
1640
1641 /* to be used by recursive prfill, sums live and dead stats recursively */
1642 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1643 {
1644         u64 sum = 0;
1645
1646         sum += blkg_stat_recursive_sum(pd, off);
1647         sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1648         return sum;
1649 }
1650
1651 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1652 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1653                                                        int off)
1654 {
1655         struct blkg_rwstat a, b;
1656
1657         a = blkg_rwstat_recursive_sum(pd, off);
1658         b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1659         blkg_rwstat_merge(&a, &b);
1660         return a;
1661 }
1662
1663 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1664 {
1665         struct cfq_group *cfqg = pd_to_cfqg(pd);
1666
1667         cfqg_stats_reset(&cfqg->stats);
1668         cfqg_stats_reset(&cfqg->dead_stats);
1669 }
1670
1671 /*
1672  * Search for the cfq group current task belongs to. request_queue lock must
1673  * be held.
1674  */
1675 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1676                                                 struct blkcg *blkcg)
1677 {
1678         struct request_queue *q = cfqd->queue;
1679         struct cfq_group *cfqg = NULL;
1680
1681         /* avoid lookup for the common case where there's no blkcg */
1682         if (blkcg == &blkcg_root) {
1683                 cfqg = cfqd->root_group;
1684         } else {
1685                 struct blkcg_gq *blkg;
1686
1687                 blkg = blkg_lookup_create(blkcg, q);
1688                 if (!IS_ERR(blkg))
1689                         cfqg = blkg_to_cfqg(blkg);
1690         }
1691
1692         return cfqg;
1693 }
1694
1695 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1696 {
1697         cfqq->cfqg = cfqg;
1698         /* cfqq reference on cfqg */
1699         cfqg_get(cfqg);
1700 }
1701
1702 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1703                                      struct blkg_policy_data *pd, int off)
1704 {
1705         struct cfq_group *cfqg = pd_to_cfqg(pd);
1706
1707         if (!cfqg->dev_weight)
1708                 return 0;
1709         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1710 }
1711
1712 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1713 {
1714         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1715                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1716                           0, false);
1717         return 0;
1718 }
1719
1720 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1721                                           struct blkg_policy_data *pd, int off)
1722 {
1723         struct cfq_group *cfqg = pd_to_cfqg(pd);
1724
1725         if (!cfqg->dev_leaf_weight)
1726                 return 0;
1727         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1728 }
1729
1730 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1731 {
1732         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1733                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1734                           0, false);
1735         return 0;
1736 }
1737
1738 static int cfq_print_weight(struct seq_file *sf, void *v)
1739 {
1740         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1741         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1742         unsigned int val = 0;
1743
1744         if (cgd)
1745                 val = cgd->weight;
1746
1747         seq_printf(sf, "%u\n", val);
1748         return 0;
1749 }
1750
1751 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1752 {
1753         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1754         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1755         unsigned int val = 0;
1756
1757         if (cgd)
1758                 val = cgd->leaf_weight;
1759
1760         seq_printf(sf, "%u\n", val);
1761         return 0;
1762 }
1763
1764 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1765                                         char *buf, size_t nbytes, loff_t off,
1766                                         bool is_leaf_weight)
1767 {
1768         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1769         struct blkg_conf_ctx ctx;
1770         struct cfq_group *cfqg;
1771         struct cfq_group_data *cfqgd;
1772         int ret;
1773
1774         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1775         if (ret)
1776                 return ret;
1777
1778         ret = -EINVAL;
1779         cfqg = blkg_to_cfqg(ctx.blkg);
1780         cfqgd = blkcg_to_cfqgd(blkcg);
1781         if (!cfqg || !cfqgd)
1782                 goto err;
1783
1784         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1785                 if (!is_leaf_weight) {
1786                         cfqg->dev_weight = ctx.v;
1787                         cfqg->new_weight = ctx.v ?: cfqgd->weight;
1788                 } else {
1789                         cfqg->dev_leaf_weight = ctx.v;
1790                         cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1791                 }
1792                 ret = 0;
1793         }
1794
1795 err:
1796         blkg_conf_finish(&ctx);
1797         return ret ?: nbytes;
1798 }
1799
1800 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1801                                       char *buf, size_t nbytes, loff_t off)
1802 {
1803         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1804 }
1805
1806 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1807                                            char *buf, size_t nbytes, loff_t off)
1808 {
1809         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1810 }
1811
1812 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1813                             u64 val, bool is_leaf_weight)
1814 {
1815         struct blkcg *blkcg = css_to_blkcg(css);
1816         struct blkcg_gq *blkg;
1817         struct cfq_group_data *cfqgd;
1818         int ret = 0;
1819
1820         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1821                 return -EINVAL;
1822
1823         spin_lock_irq(&blkcg->lock);
1824         cfqgd = blkcg_to_cfqgd(blkcg);
1825         if (!cfqgd) {
1826                 ret = -EINVAL;
1827                 goto out;
1828         }
1829
1830         if (!is_leaf_weight)
1831                 cfqgd->weight = val;
1832         else
1833                 cfqgd->leaf_weight = val;
1834
1835         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1836                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1837
1838                 if (!cfqg)
1839                         continue;
1840
1841                 if (!is_leaf_weight) {
1842                         if (!cfqg->dev_weight)
1843                                 cfqg->new_weight = cfqgd->weight;
1844                 } else {
1845                         if (!cfqg->dev_leaf_weight)
1846                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1847                 }
1848         }
1849
1850 out:
1851         spin_unlock_irq(&blkcg->lock);
1852         return ret;
1853 }
1854
1855 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1856                           u64 val)
1857 {
1858         return __cfq_set_weight(css, cft, val, false);
1859 }
1860
1861 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1862                                struct cftype *cft, u64 val)
1863 {
1864         return __cfq_set_weight(css, cft, val, true);
1865 }
1866
1867 static int cfqg_print_stat(struct seq_file *sf, void *v)
1868 {
1869         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1870                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1871         return 0;
1872 }
1873
1874 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1875 {
1876         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1877                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1878         return 0;
1879 }
1880
1881 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1882                                       struct blkg_policy_data *pd, int off)
1883 {
1884         u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1885
1886         return __blkg_prfill_u64(sf, pd, sum);
1887 }
1888
1889 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1890                                         struct blkg_policy_data *pd, int off)
1891 {
1892         struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1893
1894         return __blkg_prfill_rwstat(sf, pd, &sum);
1895 }
1896
1897 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1898 {
1899         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1900                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1901                           seq_cft(sf)->private, false);
1902         return 0;
1903 }
1904
1905 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1906 {
1907         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1908                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1909                           seq_cft(sf)->private, true);
1910         return 0;
1911 }
1912
1913 #ifdef CONFIG_DEBUG_BLK_CGROUP
1914 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1915                                       struct blkg_policy_data *pd, int off)
1916 {
1917         struct cfq_group *cfqg = pd_to_cfqg(pd);
1918         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1919         u64 v = 0;
1920
1921         if (samples) {
1922                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1923                 v = div64_u64(v, samples);
1924         }
1925         __blkg_prfill_u64(sf, pd, v);
1926         return 0;
1927 }
1928
1929 /* print avg_queue_size */
1930 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1931 {
1932         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1933                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1934                           0, false);
1935         return 0;
1936 }
1937 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1938
1939 static struct cftype cfq_blkcg_files[] = {
1940         /* on root, weight is mapped to leaf_weight */
1941         {
1942                 .name = "weight_device",
1943                 .flags = CFTYPE_ONLY_ON_ROOT,
1944                 .seq_show = cfqg_print_leaf_weight_device,
1945                 .write = cfqg_set_leaf_weight_device,
1946         },
1947         {
1948                 .name = "weight",
1949                 .flags = CFTYPE_ONLY_ON_ROOT,
1950                 .seq_show = cfq_print_leaf_weight,
1951                 .write_u64 = cfq_set_leaf_weight,
1952         },
1953
1954         /* no such mapping necessary for !roots */
1955         {
1956                 .name = "weight_device",
1957                 .flags = CFTYPE_NOT_ON_ROOT,
1958                 .seq_show = cfqg_print_weight_device,
1959                 .write = cfqg_set_weight_device,
1960         },
1961         {
1962                 .name = "weight",
1963                 .flags = CFTYPE_NOT_ON_ROOT,
1964                 .seq_show = cfq_print_weight,
1965                 .write_u64 = cfq_set_weight,
1966         },
1967
1968         {
1969                 .name = "leaf_weight_device",
1970                 .seq_show = cfqg_print_leaf_weight_device,
1971                 .write = cfqg_set_leaf_weight_device,
1972         },
1973         {
1974                 .name = "leaf_weight",
1975                 .seq_show = cfq_print_leaf_weight,
1976                 .write_u64 = cfq_set_leaf_weight,
1977         },
1978
1979         /* statistics, covers only the tasks in the cfqg */
1980         {
1981                 .name = "time",
1982                 .private = offsetof(struct cfq_group, stats.time),
1983                 .seq_show = cfqg_print_stat,
1984         },
1985         {
1986                 .name = "sectors",
1987                 .private = offsetof(struct cfq_group, stats.sectors),
1988                 .seq_show = cfqg_print_stat,
1989         },
1990         {
1991                 .name = "io_service_bytes",
1992                 .private = offsetof(struct cfq_group, stats.service_bytes),
1993                 .seq_show = cfqg_print_rwstat,
1994         },
1995         {
1996                 .name = "io_serviced",
1997                 .private = offsetof(struct cfq_group, stats.serviced),
1998                 .seq_show = cfqg_print_rwstat,
1999         },
2000         {
2001                 .name = "io_service_time",
2002                 .private = offsetof(struct cfq_group, stats.service_time),
2003                 .seq_show = cfqg_print_rwstat,
2004         },
2005         {
2006                 .name = "io_wait_time",
2007                 .private = offsetof(struct cfq_group, stats.wait_time),
2008                 .seq_show = cfqg_print_rwstat,
2009         },
2010         {
2011                 .name = "io_merged",
2012                 .private = offsetof(struct cfq_group, stats.merged),
2013                 .seq_show = cfqg_print_rwstat,
2014         },
2015         {
2016                 .name = "io_queued",
2017                 .private = offsetof(struct cfq_group, stats.queued),
2018                 .seq_show = cfqg_print_rwstat,
2019         },
2020
2021         /* the same statictics which cover the cfqg and its descendants */
2022         {
2023                 .name = "time_recursive",
2024                 .private = offsetof(struct cfq_group, stats.time),
2025                 .seq_show = cfqg_print_stat_recursive,
2026         },
2027         {
2028                 .name = "sectors_recursive",
2029                 .private = offsetof(struct cfq_group, stats.sectors),
2030                 .seq_show = cfqg_print_stat_recursive,
2031         },
2032         {
2033                 .name = "io_service_bytes_recursive",
2034                 .private = offsetof(struct cfq_group, stats.service_bytes),
2035                 .seq_show = cfqg_print_rwstat_recursive,
2036         },
2037         {
2038                 .name = "io_serviced_recursive",
2039                 .private = offsetof(struct cfq_group, stats.serviced),
2040                 .seq_show = cfqg_print_rwstat_recursive,
2041         },
2042         {
2043                 .name = "io_service_time_recursive",
2044                 .private = offsetof(struct cfq_group, stats.service_time),
2045                 .seq_show = cfqg_print_rwstat_recursive,
2046         },
2047         {
2048                 .name = "io_wait_time_recursive",
2049                 .private = offsetof(struct cfq_group, stats.wait_time),
2050                 .seq_show = cfqg_print_rwstat_recursive,
2051         },
2052         {
2053                 .name = "io_merged_recursive",
2054                 .private = offsetof(struct cfq_group, stats.merged),
2055                 .seq_show = cfqg_print_rwstat_recursive,
2056         },
2057         {
2058                 .name = "io_queued_recursive",
2059                 .private = offsetof(struct cfq_group, stats.queued),
2060                 .seq_show = cfqg_print_rwstat_recursive,
2061         },
2062 #ifdef CONFIG_DEBUG_BLK_CGROUP
2063         {
2064                 .name = "avg_queue_size",
2065                 .seq_show = cfqg_print_avg_queue_size,
2066         },
2067         {
2068                 .name = "group_wait_time",
2069                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2070                 .seq_show = cfqg_print_stat,
2071         },
2072         {
2073                 .name = "idle_time",
2074                 .private = offsetof(struct cfq_group, stats.idle_time),
2075                 .seq_show = cfqg_print_stat,
2076         },
2077         {
2078                 .name = "empty_time",
2079                 .private = offsetof(struct cfq_group, stats.empty_time),
2080                 .seq_show = cfqg_print_stat,
2081         },
2082         {
2083                 .name = "dequeue",
2084                 .private = offsetof(struct cfq_group, stats.dequeue),
2085                 .seq_show = cfqg_print_stat,
2086         },
2087         {
2088                 .name = "unaccounted_time",
2089                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2090                 .seq_show = cfqg_print_stat,
2091         },
2092 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2093         { }     /* terminate */
2094 };
2095 #else /* GROUP_IOSCHED */
2096 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
2097                                                 struct blkcg *blkcg)
2098 {
2099         return cfqd->root_group;
2100 }
2101
2102 static inline void
2103 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2104         cfqq->cfqg = cfqg;
2105 }
2106
2107 #endif /* GROUP_IOSCHED */
2108
2109 /*
2110  * The cfqd->service_trees holds all pending cfq_queue's that have
2111  * requests waiting to be processed. It is sorted in the order that
2112  * we will service the queues.
2113  */
2114 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2115                                  bool add_front)
2116 {
2117         struct rb_node **p, *parent;
2118         struct cfq_queue *__cfqq;
2119         unsigned long rb_key;
2120         struct cfq_rb_root *st;
2121         int left;
2122         int new_cfqq = 1;
2123
2124         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2125         if (cfq_class_idle(cfqq)) {
2126                 rb_key = CFQ_IDLE_DELAY;
2127                 parent = rb_last(&st->rb);
2128                 if (parent && parent != &cfqq->rb_node) {
2129                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2130                         rb_key += __cfqq->rb_key;
2131                 } else
2132                         rb_key += jiffies;
2133         } else if (!add_front) {
2134                 /*
2135                  * Get our rb key offset. Subtract any residual slice
2136                  * value carried from last service. A negative resid
2137                  * count indicates slice overrun, and this should position
2138                  * the next service time further away in the tree.
2139                  */
2140                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2141                 rb_key -= cfqq->slice_resid;
2142                 cfqq->slice_resid = 0;
2143         } else {
2144                 rb_key = -HZ;
2145                 __cfqq = cfq_rb_first(st);
2146                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2147         }
2148
2149         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2150                 new_cfqq = 0;
2151                 /*
2152                  * same position, nothing more to do
2153                  */
2154                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2155                         return;
2156
2157                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2158                 cfqq->service_tree = NULL;
2159         }
2160
2161         left = 1;
2162         parent = NULL;
2163         cfqq->service_tree = st;
2164         p = &st->rb.rb_node;
2165         while (*p) {
2166                 parent = *p;
2167                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2168
2169                 /*
2170                  * sort by key, that represents service time.
2171                  */
2172                 if (time_before(rb_key, __cfqq->rb_key))
2173                         p = &parent->rb_left;
2174                 else {
2175                         p = &parent->rb_right;
2176                         left = 0;
2177                 }
2178         }
2179
2180         if (left)
2181                 st->left = &cfqq->rb_node;
2182
2183         cfqq->rb_key = rb_key;
2184         rb_link_node(&cfqq->rb_node, parent, p);
2185         rb_insert_color(&cfqq->rb_node, &st->rb);
2186         st->count++;
2187         if (add_front || !new_cfqq)
2188                 return;
2189         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2190 }
2191
2192 static struct cfq_queue *
2193 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2194                      sector_t sector, struct rb_node **ret_parent,
2195                      struct rb_node ***rb_link)
2196 {
2197         struct rb_node **p, *parent;
2198         struct cfq_queue *cfqq = NULL;
2199
2200         parent = NULL;
2201         p = &root->rb_node;
2202         while (*p) {
2203                 struct rb_node **n;
2204
2205                 parent = *p;
2206                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2207
2208                 /*
2209                  * Sort strictly based on sector.  Smallest to the left,
2210                  * largest to the right.
2211                  */
2212                 if (sector > blk_rq_pos(cfqq->next_rq))
2213                         n = &(*p)->rb_right;
2214                 else if (sector < blk_rq_pos(cfqq->next_rq))
2215                         n = &(*p)->rb_left;
2216                 else
2217                         break;
2218                 p = n;
2219                 cfqq = NULL;
2220         }
2221
2222         *ret_parent = parent;
2223         if (rb_link)
2224                 *rb_link = p;
2225         return cfqq;
2226 }
2227
2228 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2229 {
2230         struct rb_node **p, *parent;
2231         struct cfq_queue *__cfqq;
2232
2233         if (cfqq->p_root) {
2234                 rb_erase(&cfqq->p_node, cfqq->p_root);
2235                 cfqq->p_root = NULL;
2236         }
2237
2238         if (cfq_class_idle(cfqq))
2239                 return;
2240         if (!cfqq->next_rq)
2241                 return;
2242
2243         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2244         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2245                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2246         if (!__cfqq) {
2247                 rb_link_node(&cfqq->p_node, parent, p);
2248                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2249         } else
2250                 cfqq->p_root = NULL;
2251 }
2252
2253 /*
2254  * Update cfqq's position in the service tree.
2255  */
2256 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2257 {
2258         /*
2259          * Resorting requires the cfqq to be on the RR list already.
2260          */
2261         if (cfq_cfqq_on_rr(cfqq)) {
2262                 cfq_service_tree_add(cfqd, cfqq, 0);
2263                 cfq_prio_tree_add(cfqd, cfqq);
2264         }
2265 }
2266
2267 /*
2268  * add to busy list of queues for service, trying to be fair in ordering
2269  * the pending list according to last request service
2270  */
2271 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2272 {
2273         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2274         BUG_ON(cfq_cfqq_on_rr(cfqq));
2275         cfq_mark_cfqq_on_rr(cfqq);
2276         cfqd->busy_queues++;
2277         if (cfq_cfqq_sync(cfqq))
2278                 cfqd->busy_sync_queues++;
2279
2280         cfq_resort_rr_list(cfqd, cfqq);
2281 }
2282
2283 /*
2284  * Called when the cfqq no longer has requests pending, remove it from
2285  * the service tree.
2286  */
2287 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2288 {
2289         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2290         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2291         cfq_clear_cfqq_on_rr(cfqq);
2292
2293         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2294                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2295                 cfqq->service_tree = NULL;
2296         }
2297         if (cfqq->p_root) {
2298                 rb_erase(&cfqq->p_node, cfqq->p_root);
2299                 cfqq->p_root = NULL;
2300         }
2301
2302         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2303         BUG_ON(!cfqd->busy_queues);
2304         cfqd->busy_queues--;
2305         if (cfq_cfqq_sync(cfqq))
2306                 cfqd->busy_sync_queues--;
2307 }
2308
2309 /*
2310  * rb tree support functions
2311  */
2312 static void cfq_del_rq_rb(struct request *rq)
2313 {
2314         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2315         const int sync = rq_is_sync(rq);
2316
2317         BUG_ON(!cfqq->queued[sync]);
2318         cfqq->queued[sync]--;
2319
2320         elv_rb_del(&cfqq->sort_list, rq);
2321
2322         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2323                 /*
2324                  * Queue will be deleted from service tree when we actually
2325                  * expire it later. Right now just remove it from prio tree
2326                  * as it is empty.
2327                  */
2328                 if (cfqq->p_root) {
2329                         rb_erase(&cfqq->p_node, cfqq->p_root);
2330                         cfqq->p_root = NULL;
2331                 }
2332         }
2333 }
2334
2335 static void cfq_add_rq_rb(struct request *rq)
2336 {
2337         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2338         struct cfq_data *cfqd = cfqq->cfqd;
2339         struct request *prev;
2340
2341         cfqq->queued[rq_is_sync(rq)]++;
2342
2343         elv_rb_add(&cfqq->sort_list, rq);
2344
2345         if (!cfq_cfqq_on_rr(cfqq))
2346                 cfq_add_cfqq_rr(cfqd, cfqq);
2347
2348         /*
2349          * check if this request is a better next-serve candidate
2350          */
2351         prev = cfqq->next_rq;
2352         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2353
2354         /*
2355          * adjust priority tree position, if ->next_rq changes
2356          */
2357         if (prev != cfqq->next_rq)
2358                 cfq_prio_tree_add(cfqd, cfqq);
2359
2360         BUG_ON(!cfqq->next_rq);
2361 }
2362
2363 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2364 {
2365         elv_rb_del(&cfqq->sort_list, rq);
2366         cfqq->queued[rq_is_sync(rq)]--;
2367         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2368         cfq_add_rq_rb(rq);
2369         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2370                                  rq->cmd_flags);
2371 }
2372
2373 static struct request *
2374 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2375 {
2376         struct task_struct *tsk = current;
2377         struct cfq_io_cq *cic;
2378         struct cfq_queue *cfqq;
2379
2380         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2381         if (!cic)
2382                 return NULL;
2383
2384         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2385         if (cfqq)
2386                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2387
2388         return NULL;
2389 }
2390
2391 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2392 {
2393         struct cfq_data *cfqd = q->elevator->elevator_data;
2394
2395         cfqd->rq_in_driver++;
2396         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2397                                                 cfqd->rq_in_driver);
2398
2399         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2400 }
2401
2402 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2403 {
2404         struct cfq_data *cfqd = q->elevator->elevator_data;
2405
2406         WARN_ON(!cfqd->rq_in_driver);
2407         cfqd->rq_in_driver--;
2408         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2409                                                 cfqd->rq_in_driver);
2410 }
2411
2412 static void cfq_remove_request(struct request *rq)
2413 {
2414         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2415
2416         if (cfqq->next_rq == rq)
2417                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2418
2419         list_del_init(&rq->queuelist);
2420         cfq_del_rq_rb(rq);
2421
2422         cfqq->cfqd->rq_queued--;
2423         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2424         if (rq->cmd_flags & REQ_PRIO) {
2425                 WARN_ON(!cfqq->prio_pending);
2426                 cfqq->prio_pending--;
2427         }
2428 }
2429
2430 static int cfq_merge(struct request_queue *q, struct request **req,
2431                      struct bio *bio)
2432 {
2433         struct cfq_data *cfqd = q->elevator->elevator_data;
2434         struct request *__rq;
2435
2436         __rq = cfq_find_rq_fmerge(cfqd, bio);
2437         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2438                 *req = __rq;
2439                 return ELEVATOR_FRONT_MERGE;
2440         }
2441
2442         return ELEVATOR_NO_MERGE;
2443 }
2444
2445 static void cfq_merged_request(struct request_queue *q, struct request *req,
2446                                int type)
2447 {
2448         if (type == ELEVATOR_FRONT_MERGE) {
2449                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2450
2451                 cfq_reposition_rq_rb(cfqq, req);
2452         }
2453 }
2454
2455 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2456                                 struct bio *bio)
2457 {
2458         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2459 }
2460
2461 static void
2462 cfq_merged_requests(struct request_queue *q, struct request *rq,
2463                     struct request *next)
2464 {
2465         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2466         struct cfq_data *cfqd = q->elevator->elevator_data;
2467
2468         /*
2469          * reposition in fifo if next is older than rq
2470          */
2471         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2472             time_before(next->fifo_time, rq->fifo_time) &&
2473             cfqq == RQ_CFQQ(next)) {
2474                 list_move(&rq->queuelist, &next->queuelist);
2475                 rq->fifo_time = next->fifo_time;
2476         }
2477
2478         if (cfqq->next_rq == next)
2479                 cfqq->next_rq = rq;
2480         cfq_remove_request(next);
2481         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2482
2483         cfqq = RQ_CFQQ(next);
2484         /*
2485          * all requests of this queue are merged to other queues, delete it
2486          * from the service tree. If it's the active_queue,
2487          * cfq_dispatch_requests() will choose to expire it or do idle
2488          */
2489         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2490             cfqq != cfqd->active_queue)
2491                 cfq_del_cfqq_rr(cfqd, cfqq);
2492 }
2493
2494 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2495                            struct bio *bio)
2496 {
2497         struct cfq_data *cfqd = q->elevator->elevator_data;
2498         struct cfq_io_cq *cic;
2499         struct cfq_queue *cfqq;
2500
2501         /*
2502          * Disallow merge of a sync bio into an async request.
2503          */
2504         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2505                 return false;
2506
2507         /*
2508          * Lookup the cfqq that this bio will be queued with and allow
2509          * merge only if rq is queued there.
2510          */
2511         cic = cfq_cic_lookup(cfqd, current->io_context);
2512         if (!cic)
2513                 return false;
2514
2515         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2516         return cfqq == RQ_CFQQ(rq);
2517 }
2518
2519 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2520 {
2521         del_timer(&cfqd->idle_slice_timer);
2522         cfqg_stats_update_idle_time(cfqq->cfqg);
2523 }
2524
2525 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2526                                    struct cfq_queue *cfqq)
2527 {
2528         if (cfqq) {
2529                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2530                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2531                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2532                 cfqq->slice_start = 0;
2533                 cfqq->dispatch_start = jiffies;
2534                 cfqq->allocated_slice = 0;
2535                 cfqq->slice_end = 0;
2536                 cfqq->slice_dispatch = 0;
2537                 cfqq->nr_sectors = 0;
2538
2539                 cfq_clear_cfqq_wait_request(cfqq);
2540                 cfq_clear_cfqq_must_dispatch(cfqq);
2541                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2542                 cfq_clear_cfqq_fifo_expire(cfqq);
2543                 cfq_mark_cfqq_slice_new(cfqq);
2544
2545                 cfq_del_timer(cfqd, cfqq);
2546         }
2547
2548         cfqd->active_queue = cfqq;
2549 }
2550
2551 /*
2552  * current cfqq expired its slice (or was too idle), select new one
2553  */
2554 static void
2555 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2556                     bool timed_out)
2557 {
2558         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2559
2560         if (cfq_cfqq_wait_request(cfqq))
2561                 cfq_del_timer(cfqd, cfqq);
2562
2563         cfq_clear_cfqq_wait_request(cfqq);
2564         cfq_clear_cfqq_wait_busy(cfqq);
2565
2566         /*
2567          * If this cfqq is shared between multiple processes, check to
2568          * make sure that those processes are still issuing I/Os within
2569          * the mean seek distance.  If not, it may be time to break the
2570          * queues apart again.
2571          */
2572         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2573                 cfq_mark_cfqq_split_coop(cfqq);
2574
2575         /*
2576          * store what was left of this slice, if the queue idled/timed out
2577          */
2578         if (timed_out) {
2579                 if (cfq_cfqq_slice_new(cfqq))
2580                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2581                 else
2582                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2583                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2584         }
2585
2586         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2587
2588         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2589                 cfq_del_cfqq_rr(cfqd, cfqq);
2590
2591         cfq_resort_rr_list(cfqd, cfqq);
2592
2593         if (cfqq == cfqd->active_queue)
2594                 cfqd->active_queue = NULL;
2595
2596         if (cfqd->active_cic) {
2597                 put_io_context(cfqd->active_cic->icq.ioc);
2598                 cfqd->active_cic = NULL;
2599         }
2600 }
2601
2602 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2603 {
2604         struct cfq_queue *cfqq = cfqd->active_queue;
2605
2606         if (cfqq)
2607                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2608 }
2609
2610 /*
2611  * Get next queue for service. Unless we have a queue preemption,
2612  * we'll simply select the first cfqq in the service tree.
2613  */
2614 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2615 {
2616         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2617                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2618
2619         if (!cfqd->rq_queued)
2620                 return NULL;
2621
2622         /* There is nothing to dispatch */
2623         if (!st)
2624                 return NULL;
2625         if (RB_EMPTY_ROOT(&st->rb))
2626                 return NULL;
2627         return cfq_rb_first(st);
2628 }
2629
2630 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2631 {
2632         struct cfq_group *cfqg;
2633         struct cfq_queue *cfqq;
2634         int i, j;
2635         struct cfq_rb_root *st;
2636
2637         if (!cfqd->rq_queued)
2638                 return NULL;
2639
2640         cfqg = cfq_get_next_cfqg(cfqd);
2641         if (!cfqg)
2642                 return NULL;
2643
2644         for_each_cfqg_st(cfqg, i, j, st)
2645                 if ((cfqq = cfq_rb_first(st)) != NULL)
2646                         return cfqq;
2647         return NULL;
2648 }
2649
2650 /*
2651  * Get and set a new active queue for service.
2652  */
2653 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2654                                               struct cfq_queue *cfqq)
2655 {
2656         if (!cfqq)
2657                 cfqq = cfq_get_next_queue(cfqd);
2658
2659         __cfq_set_active_queue(cfqd, cfqq);
2660         return cfqq;
2661 }
2662
2663 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2664                                           struct request *rq)
2665 {
2666         if (blk_rq_pos(rq) >= cfqd->last_position)
2667                 return blk_rq_pos(rq) - cfqd->last_position;
2668         else
2669                 return cfqd->last_position - blk_rq_pos(rq);
2670 }
2671
2672 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2673                                struct request *rq)
2674 {
2675         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2676 }
2677
2678 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2679                                     struct cfq_queue *cur_cfqq)
2680 {
2681         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2682         struct rb_node *parent, *node;
2683         struct cfq_queue *__cfqq;
2684         sector_t sector = cfqd->last_position;
2685
2686         if (RB_EMPTY_ROOT(root))
2687                 return NULL;
2688
2689         /*
2690          * First, if we find a request starting at the end of the last
2691          * request, choose it.
2692          */
2693         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2694         if (__cfqq)
2695                 return __cfqq;
2696
2697         /*
2698          * If the exact sector wasn't found, the parent of the NULL leaf
2699          * will contain the closest sector.
2700          */
2701         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2702         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2703                 return __cfqq;
2704
2705         if (blk_rq_pos(__cfqq->next_rq) < sector)
2706                 node = rb_next(&__cfqq->p_node);
2707         else
2708                 node = rb_prev(&__cfqq->p_node);
2709         if (!node)
2710                 return NULL;
2711
2712         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2713         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2714                 return __cfqq;
2715
2716         return NULL;
2717 }
2718
2719 /*
2720  * cfqd - obvious
2721  * cur_cfqq - passed in so that we don't decide that the current queue is
2722  *            closely cooperating with itself.
2723  *
2724  * So, basically we're assuming that that cur_cfqq has dispatched at least
2725  * one request, and that cfqd->last_position reflects a position on the disk
2726  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2727  * assumption.
2728  */
2729 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2730                                               struct cfq_queue *cur_cfqq)
2731 {
2732         struct cfq_queue *cfqq;
2733
2734         if (cfq_class_idle(cur_cfqq))
2735                 return NULL;
2736         if (!cfq_cfqq_sync(cur_cfqq))
2737                 return NULL;
2738         if (CFQQ_SEEKY(cur_cfqq))
2739                 return NULL;
2740
2741         /*
2742          * Don't search priority tree if it's the only queue in the group.
2743          */
2744         if (cur_cfqq->cfqg->nr_cfqq == 1)
2745                 return NULL;
2746
2747         /*
2748          * We should notice if some of the queues are cooperating, eg
2749          * working closely on the same area of the disk. In that case,
2750          * we can group them together and don't waste time idling.
2751          */
2752         cfqq = cfqq_close(cfqd, cur_cfqq);
2753         if (!cfqq)
2754                 return NULL;
2755
2756         /* If new queue belongs to different cfq_group, don't choose it */
2757         if (cur_cfqq->cfqg != cfqq->cfqg)
2758                 return NULL;
2759
2760         /*
2761          * It only makes sense to merge sync queues.
2762          */
2763         if (!cfq_cfqq_sync(cfqq))
2764                 return NULL;
2765         if (CFQQ_SEEKY(cfqq))
2766                 return NULL;
2767
2768         /*
2769          * Do not merge queues of different priority classes
2770          */
2771         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2772                 return NULL;
2773
2774         return cfqq;
2775 }
2776
2777 /*
2778  * Determine whether we should enforce idle window for this queue.
2779  */
2780
2781 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2782 {
2783         enum wl_class_t wl_class = cfqq_class(cfqq);
2784         struct cfq_rb_root *st = cfqq->service_tree;
2785
2786         BUG_ON(!st);
2787         BUG_ON(!st->count);
2788
2789         if (!cfqd->cfq_slice_idle)
2790                 return false;
2791
2792         /* We never do for idle class queues. */
2793         if (wl_class == IDLE_WORKLOAD)
2794                 return false;
2795
2796         /* We do for queues that were marked with idle window flag. */
2797         if (cfq_cfqq_idle_window(cfqq) &&
2798            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2799                 return true;
2800
2801         /*
2802          * Otherwise, we do only if they are the last ones
2803          * in their service tree.
2804          */
2805         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2806            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2807                 return true;
2808         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2809         return false;
2810 }
2811
2812 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2813 {
2814         struct cfq_queue *cfqq = cfqd->active_queue;
2815         struct cfq_io_cq *cic;
2816         unsigned long sl, group_idle = 0;
2817
2818         /*
2819          * SSD device without seek penalty, disable idling. But only do so
2820          * for devices that support queuing, otherwise we still have a problem
2821          * with sync vs async workloads.
2822          */
2823         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2824                 return;
2825
2826         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2827         WARN_ON(cfq_cfqq_slice_new(cfqq));
2828
2829         /*
2830          * idle is disabled, either manually or by past process history
2831          */
2832         if (!cfq_should_idle(cfqd, cfqq)) {
2833                 /* no queue idling. Check for group idling */
2834                 if (cfqd->cfq_group_idle)
2835                         group_idle = cfqd->cfq_group_idle;
2836                 else
2837                         return;
2838         }
2839
2840         /*
2841          * still active requests from this queue, don't idle
2842          */
2843         if (cfqq->dispatched)
2844                 return;
2845
2846         /*
2847          * task has exited, don't wait
2848          */
2849         cic = cfqd->active_cic;
2850         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2851                 return;
2852
2853         /*
2854          * If our average think time is larger than the remaining time
2855          * slice, then don't idle. This avoids overrunning the allotted
2856          * time slice.
2857          */
2858         if (sample_valid(cic->ttime.ttime_samples) &&
2859             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2860                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2861                              cic->ttime.ttime_mean);
2862                 return;
2863         }
2864
2865         /* There are other queues in the group, don't do group idle */
2866         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2867                 return;
2868
2869         cfq_mark_cfqq_wait_request(cfqq);
2870
2871         if (group_idle)
2872                 sl = cfqd->cfq_group_idle;
2873         else
2874                 sl = cfqd->cfq_slice_idle;
2875
2876         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2877         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2878         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2879                         group_idle ? 1 : 0);
2880 }
2881
2882 /*
2883  * Move request from internal lists to the request queue dispatch list.
2884  */
2885 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2886 {
2887         struct cfq_data *cfqd = q->elevator->elevator_data;
2888         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2889
2890         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2891
2892         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2893         cfq_remove_request(rq);
2894         cfqq->dispatched++;
2895         (RQ_CFQG(rq))->dispatched++;
2896         elv_dispatch_sort(q, rq);
2897
2898         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2899         cfqq->nr_sectors += blk_rq_sectors(rq);
2900         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2901 }
2902
2903 /*
2904  * return expired entry, or NULL to just start from scratch in rbtree
2905  */
2906 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2907 {
2908         struct request *rq = NULL;
2909
2910         if (cfq_cfqq_fifo_expire(cfqq))
2911                 return NULL;
2912
2913         cfq_mark_cfqq_fifo_expire(cfqq);
2914
2915         if (list_empty(&cfqq->fifo))
2916                 return NULL;
2917
2918         rq = rq_entry_fifo(cfqq->fifo.next);
2919         if (time_before(jiffies, rq->fifo_time))
2920                 rq = NULL;
2921
2922         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2923         return rq;
2924 }
2925
2926 static inline int
2927 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2928 {
2929         const int base_rq = cfqd->cfq_slice_async_rq;
2930
2931         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2932
2933         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2934 }
2935
2936 /*
2937  * Must be called with the queue_lock held.
2938  */
2939 static int cfqq_process_refs(struct cfq_queue *cfqq)
2940 {
2941         int process_refs, io_refs;
2942
2943         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2944         process_refs = cfqq->ref - io_refs;
2945         BUG_ON(process_refs < 0);
2946         return process_refs;
2947 }
2948
2949 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2950 {
2951         int process_refs, new_process_refs;
2952         struct cfq_queue *__cfqq;
2953
2954         /*
2955          * If there are no process references on the new_cfqq, then it is
2956          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2957          * chain may have dropped their last reference (not just their
2958          * last process reference).
2959          */
2960         if (!cfqq_process_refs(new_cfqq))
2961                 return;
2962
2963         /* Avoid a circular list and skip interim queue merges */
2964         while ((__cfqq = new_cfqq->new_cfqq)) {
2965                 if (__cfqq == cfqq)
2966                         return;
2967                 new_cfqq = __cfqq;
2968         }
2969
2970         process_refs = cfqq_process_refs(cfqq);
2971         new_process_refs = cfqq_process_refs(new_cfqq);
2972         /*
2973          * If the process for the cfqq has gone away, there is no
2974          * sense in merging the queues.
2975          */
2976         if (process_refs == 0 || new_process_refs == 0)
2977                 return;
2978
2979         /*
2980          * Merge in the direction of the lesser amount of work.
2981          */
2982         if (new_process_refs >= process_refs) {
2983                 cfqq->new_cfqq = new_cfqq;
2984                 new_cfqq->ref += process_refs;
2985         } else {
2986                 new_cfqq->new_cfqq = cfqq;
2987                 cfqq->ref += new_process_refs;
2988         }
2989 }
2990
2991 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2992                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2993 {
2994         struct cfq_queue *queue;
2995         int i;
2996         bool key_valid = false;
2997         unsigned long lowest_key = 0;
2998         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2999
3000         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3001                 /* select the one with lowest rb_key */
3002                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3003                 if (queue &&
3004                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
3005                         lowest_key = queue->rb_key;
3006                         cur_best = i;
3007                         key_valid = true;
3008                 }
3009         }
3010
3011         return cur_best;
3012 }
3013
3014 static void
3015 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3016 {
3017         unsigned slice;
3018         unsigned count;
3019         struct cfq_rb_root *st;
3020         unsigned group_slice;
3021         enum wl_class_t original_class = cfqd->serving_wl_class;
3022
3023         /* Choose next priority. RT > BE > IDLE */
3024         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3025                 cfqd->serving_wl_class = RT_WORKLOAD;
3026         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3027                 cfqd->serving_wl_class = BE_WORKLOAD;
3028         else {
3029                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3030                 cfqd->workload_expires = jiffies + 1;
3031                 return;
3032         }
3033
3034         if (original_class != cfqd->serving_wl_class)
3035                 goto new_workload;
3036
3037         /*
3038          * For RT and BE, we have to choose also the type
3039          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3040          * expiration time
3041          */
3042         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3043         count = st->count;
3044
3045         /*
3046          * check workload expiration, and that we still have other queues ready
3047          */
3048         if (count && !time_after(jiffies, cfqd->workload_expires))
3049                 return;
3050
3051 new_workload:
3052         /* otherwise select new workload type */
3053         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3054                                         cfqd->serving_wl_class);
3055         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3056         count = st->count;
3057
3058         /*
3059          * the workload slice is computed as a fraction of target latency
3060          * proportional to the number of queues in that workload, over
3061          * all the queues in the same priority class
3062          */
3063         group_slice = cfq_group_slice(cfqd, cfqg);
3064
3065         slice = group_slice * count /
3066                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3067                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3068                                         cfqg));
3069
3070         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3071                 unsigned int tmp;
3072
3073                 /*
3074                  * Async queues are currently system wide. Just taking
3075                  * proportion of queues with-in same group will lead to higher
3076                  * async ratio system wide as generally root group is going
3077                  * to have higher weight. A more accurate thing would be to
3078                  * calculate system wide asnc/sync ratio.
3079                  */
3080                 tmp = cfqd->cfq_target_latency *
3081                         cfqg_busy_async_queues(cfqd, cfqg);
3082                 tmp = tmp/cfqd->busy_queues;
3083                 slice = min_t(unsigned, slice, tmp);
3084
3085                 /* async workload slice is scaled down according to
3086                  * the sync/async slice ratio. */
3087                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3088         } else
3089                 /* sync workload slice is at least 2 * cfq_slice_idle */
3090                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3091
3092         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3093         cfq_log(cfqd, "workload slice:%d", slice);
3094         cfqd->workload_expires = jiffies + slice;
3095 }
3096
3097 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3098 {
3099         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3100         struct cfq_group *cfqg;
3101
3102         if (RB_EMPTY_ROOT(&st->rb))
3103                 return NULL;
3104         cfqg = cfq_rb_first_group(st);
3105         update_min_vdisktime(st);
3106         return cfqg;
3107 }
3108
3109 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3110 {
3111         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3112
3113         cfqd->serving_group = cfqg;
3114
3115         /* Restore the workload type data */
3116         if (cfqg->saved_wl_slice) {
3117                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3118                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3119                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3120         } else
3121                 cfqd->workload_expires = jiffies - 1;
3122
3123         choose_wl_class_and_type(cfqd, cfqg);
3124 }
3125
3126 /*
3127  * Select a queue for service. If we have a current active queue,
3128  * check whether to continue servicing it, or retrieve and set a new one.
3129  */
3130 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3131 {
3132         struct cfq_queue *cfqq, *new_cfqq = NULL;
3133
3134         cfqq = cfqd->active_queue;
3135         if (!cfqq)
3136                 goto new_queue;
3137
3138         if (!cfqd->rq_queued)
3139                 return NULL;
3140
3141         /*
3142          * We were waiting for group to get backlogged. Expire the queue
3143          */
3144         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3145                 goto expire;
3146
3147         /*
3148          * The active queue has run out of time, expire it and select new.
3149          */
3150         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3151                 /*
3152                  * If slice had not expired at the completion of last request
3153                  * we might not have turned on wait_busy flag. Don't expire
3154                  * the queue yet. Allow the group to get backlogged.
3155                  *
3156                  * The very fact that we have used the slice, that means we
3157                  * have been idling all along on this queue and it should be
3158                  * ok to wait for this request to complete.
3159                  */
3160                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3161                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3162                         cfqq = NULL;
3163                         goto keep_queue;
3164                 } else
3165                         goto check_group_idle;
3166         }
3167
3168         /*
3169          * The active queue has requests and isn't expired, allow it to
3170          * dispatch.
3171          */
3172         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3173                 goto keep_queue;
3174
3175         /*
3176          * If another queue has a request waiting within our mean seek
3177          * distance, let it run.  The expire code will check for close
3178          * cooperators and put the close queue at the front of the service
3179          * tree.  If possible, merge the expiring queue with the new cfqq.
3180          */
3181         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3182         if (new_cfqq) {
3183                 if (!cfqq->new_cfqq)
3184                         cfq_setup_merge(cfqq, new_cfqq);
3185                 goto expire;
3186         }
3187
3188         /*
3189          * No requests pending. If the active queue still has requests in
3190          * flight or is idling for a new request, allow either of these
3191          * conditions to happen (or time out) before selecting a new queue.
3192          */
3193         if (timer_pending(&cfqd->idle_slice_timer)) {
3194                 cfqq = NULL;
3195                 goto keep_queue;
3196         }
3197
3198         /*
3199          * This is a deep seek queue, but the device is much faster than
3200          * the queue can deliver, don't idle
3201          **/
3202         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3203             (cfq_cfqq_slice_new(cfqq) ||
3204             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3205                 cfq_clear_cfqq_deep(cfqq);
3206                 cfq_clear_cfqq_idle_window(cfqq);
3207         }
3208
3209         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3210                 cfqq = NULL;
3211                 goto keep_queue;
3212         }
3213
3214         /*
3215          * If group idle is enabled and there are requests dispatched from
3216          * this group, wait for requests to complete.
3217          */
3218 check_group_idle:
3219         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3220             cfqq->cfqg->dispatched &&
3221             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3222                 cfqq = NULL;
3223                 goto keep_queue;
3224         }
3225
3226 expire:
3227         cfq_slice_expired(cfqd, 0);
3228 new_queue:
3229         /*
3230          * Current queue expired. Check if we have to switch to a new
3231          * service tree
3232          */
3233         if (!new_cfqq)
3234                 cfq_choose_cfqg(cfqd);
3235
3236         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3237 keep_queue:
3238         return cfqq;
3239 }
3240
3241 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3242 {
3243         int dispatched = 0;
3244
3245         while (cfqq->next_rq) {
3246                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3247                 dispatched++;
3248         }
3249
3250         BUG_ON(!list_empty(&cfqq->fifo));
3251
3252         /* By default cfqq is not expired if it is empty. Do it explicitly */
3253         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3254         return dispatched;
3255 }
3256
3257 /*
3258  * Drain our current requests. Used for barriers and when switching
3259  * io schedulers on-the-fly.
3260  */
3261 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3262 {
3263         struct cfq_queue *cfqq;
3264         int dispatched = 0;
3265
3266         /* Expire the timeslice of the current active queue first */
3267         cfq_slice_expired(cfqd, 0);
3268         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3269                 __cfq_set_active_queue(cfqd, cfqq);
3270                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3271         }
3272
3273         BUG_ON(cfqd->busy_queues);
3274
3275         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3276         return dispatched;
3277 }
3278
3279 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3280         struct cfq_queue *cfqq)
3281 {
3282         /* the queue hasn't finished any request, can't estimate */
3283         if (cfq_cfqq_slice_new(cfqq))
3284                 return true;
3285         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3286                 cfqq->slice_end))
3287                 return true;
3288
3289         return false;
3290 }
3291
3292 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3293 {
3294         unsigned int max_dispatch;
3295
3296         /*
3297          * Drain async requests before we start sync IO
3298          */
3299         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3300                 return false;
3301
3302         /*
3303          * If this is an async queue and we have sync IO in flight, let it wait
3304          */
3305         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3306                 return false;
3307
3308         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3309         if (cfq_class_idle(cfqq))
3310                 max_dispatch = 1;
3311
3312         /*
3313          * Does this cfqq already have too much IO in flight?
3314          */
3315         if (cfqq->dispatched >= max_dispatch) {
3316                 bool promote_sync = false;
3317                 /*
3318                  * idle queue must always only have a single IO in flight
3319                  */
3320                 if (cfq_class_idle(cfqq))
3321                         return false;
3322
3323                 /*
3324                  * If there is only one sync queue
3325                  * we can ignore async queue here and give the sync
3326                  * queue no dispatch limit. The reason is a sync queue can
3327                  * preempt async queue, limiting the sync queue doesn't make
3328                  * sense. This is useful for aiostress test.
3329                  */
3330                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3331                         promote_sync = true;
3332
3333                 /*
3334                  * We have other queues, don't allow more IO from this one
3335                  */
3336                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3337                                 !promote_sync)
3338                         return false;
3339
3340                 /*
3341                  * Sole queue user, no limit
3342                  */
3343                 if (cfqd->busy_queues == 1 || promote_sync)
3344                         max_dispatch = -1;
3345                 else
3346                         /*
3347                          * Normally we start throttling cfqq when cfq_quantum/2
3348                          * requests have been dispatched. But we can drive
3349                          * deeper queue depths at the beginning of slice
3350                          * subjected to upper limit of cfq_quantum.
3351                          * */
3352                         max_dispatch = cfqd->cfq_quantum;
3353         }
3354
3355         /*
3356          * Async queues must wait a bit before being allowed dispatch.
3357          * We also ramp up the dispatch depth gradually for async IO,
3358          * based on the last sync IO we serviced
3359          */
3360         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3361                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3362                 unsigned int depth;
3363
3364                 depth = last_sync / cfqd->cfq_slice[1];
3365                 if (!depth && !cfqq->dispatched)
3366                         depth = 1;
3367                 if (depth < max_dispatch)
3368                         max_dispatch = depth;
3369         }
3370
3371         /*
3372          * If we're below the current max, allow a dispatch
3373          */
3374         return cfqq->dispatched < max_dispatch;
3375 }
3376
3377 /*
3378  * Dispatch a request from cfqq, moving them to the request queue
3379  * dispatch list.
3380  */
3381 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3382 {
3383         struct request *rq;
3384
3385         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3386
3387         if (!cfq_may_dispatch(cfqd, cfqq))
3388                 return false;
3389
3390         /*
3391          * follow expired path, else get first next available
3392          */
3393         rq = cfq_check_fifo(cfqq);
3394         if (!rq)
3395                 rq = cfqq->next_rq;
3396
3397         /*
3398          * insert request into driver dispatch list
3399          */
3400         cfq_dispatch_insert(cfqd->queue, rq);
3401
3402         if (!cfqd->active_cic) {
3403                 struct cfq_io_cq *cic = RQ_CIC(rq);
3404
3405                 atomic_long_inc(&cic->icq.ioc->refcount);
3406                 cfqd->active_cic = cic;
3407         }
3408
3409         return true;
3410 }
3411
3412 /*
3413  * Find the cfqq that we need to service and move a request from that to the
3414  * dispatch list
3415  */
3416 static int cfq_dispatch_requests(struct request_queue *q, int force)
3417 {
3418         struct cfq_data *cfqd = q->elevator->elevator_data;
3419         struct cfq_queue *cfqq;
3420
3421         if (!cfqd->busy_queues)
3422                 return 0;
3423
3424         if (unlikely(force))
3425                 return cfq_forced_dispatch(cfqd);
3426
3427         cfqq = cfq_select_queue(cfqd);
3428         if (!cfqq)
3429                 return 0;
3430
3431         /*
3432          * Dispatch a request from this cfqq, if it is allowed
3433          */
3434         if (!cfq_dispatch_request(cfqd, cfqq))
3435                 return 0;
3436
3437         cfqq->slice_dispatch++;
3438         cfq_clear_cfqq_must_dispatch(cfqq);
3439
3440         /*
3441          * expire an async queue immediately if it has used up its slice. idle
3442          * queue always expire after 1 dispatch round.
3443          */
3444         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3445             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3446             cfq_class_idle(cfqq))) {
3447                 cfqq->slice_end = jiffies + 1;
3448                 cfq_slice_expired(cfqd, 0);
3449         }
3450
3451         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3452         return 1;
3453 }
3454
3455 /*
3456  * task holds one reference to the queue, dropped when task exits. each rq
3457  * in-flight on this queue also holds a reference, dropped when rq is freed.
3458  *
3459  * Each cfq queue took a reference on the parent group. Drop it now.
3460  * queue lock must be held here.
3461  */
3462 static void cfq_put_queue(struct cfq_queue *cfqq)
3463 {
3464         struct cfq_data *cfqd = cfqq->cfqd;
3465         struct cfq_group *cfqg;
3466
3467         BUG_ON(cfqq->ref <= 0);
3468
3469         cfqq->ref--;
3470         if (cfqq->ref)
3471                 return;
3472
3473         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3474         BUG_ON(rb_first(&cfqq->sort_list));
3475         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3476         cfqg = cfqq->cfqg;
3477
3478         if (unlikely(cfqd->active_queue == cfqq)) {
3479                 __cfq_slice_expired(cfqd, cfqq, 0);
3480                 cfq_schedule_dispatch(cfqd);
3481         }
3482
3483         BUG_ON(cfq_cfqq_on_rr(cfqq));
3484         kmem_cache_free(cfq_pool, cfqq);
3485         cfqg_put(cfqg);
3486 }
3487
3488 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3489 {
3490         struct cfq_queue *__cfqq, *next;
3491
3492         /*
3493          * If this queue was scheduled to merge with another queue, be
3494          * sure to drop the reference taken on that queue (and others in
3495          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3496          */
3497         __cfqq = cfqq->new_cfqq;
3498         while (__cfqq) {
3499                 if (__cfqq == cfqq) {
3500                         WARN(1, "cfqq->new_cfqq loop detected\n");
3501                         break;
3502                 }
3503                 next = __cfqq->new_cfqq;
3504                 cfq_put_queue(__cfqq);
3505                 __cfqq = next;
3506         }
3507 }
3508
3509 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3510 {
3511         if (unlikely(cfqq == cfqd->active_queue)) {
3512                 __cfq_slice_expired(cfqd, cfqq, 0);
3513                 cfq_schedule_dispatch(cfqd);
3514         }
3515
3516         cfq_put_cooperator(cfqq);
3517
3518         cfq_put_queue(cfqq);
3519 }
3520
3521 static void cfq_init_icq(struct io_cq *icq)
3522 {
3523         struct cfq_io_cq *cic = icq_to_cic(icq);
3524
3525         cic->ttime.last_end_request = jiffies;
3526 }
3527
3528 static void cfq_exit_icq(struct io_cq *icq)
3529 {
3530         struct cfq_io_cq *cic = icq_to_cic(icq);
3531         struct cfq_data *cfqd = cic_to_cfqd(cic);
3532
3533         if (cic_to_cfqq(cic, false)) {
3534                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3535                 cic_set_cfqq(cic, NULL, false);
3536         }
3537
3538         if (cic_to_cfqq(cic, true)) {
3539                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3540                 cic_set_cfqq(cic, NULL, true);
3541         }
3542 }
3543
3544 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3545 {
3546         struct task_struct *tsk = current;
3547         int ioprio_class;
3548
3549         if (!cfq_cfqq_prio_changed(cfqq))
3550                 return;
3551
3552         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3553         switch (ioprio_class) {
3554         default:
3555                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3556         case IOPRIO_CLASS_NONE:
3557                 /*
3558                  * no prio set, inherit CPU scheduling settings
3559                  */
3560                 cfqq->ioprio = task_nice_ioprio(tsk);
3561                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3562                 break;
3563         case IOPRIO_CLASS_RT:
3564                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3565                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3566                 break;
3567         case IOPRIO_CLASS_BE:
3568                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3569                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3570                 break;
3571         case IOPRIO_CLASS_IDLE:
3572                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3573                 cfqq->ioprio = 7;
3574                 cfq_clear_cfqq_idle_window(cfqq);
3575                 break;
3576         }
3577
3578         /*
3579          * keep track of original prio settings in case we have to temporarily
3580          * elevate the priority of this queue
3581          */
3582         cfqq->org_ioprio = cfqq->ioprio;
3583         cfq_clear_cfqq_prio_changed(cfqq);
3584 }
3585
3586 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3587 {
3588         int ioprio = cic->icq.ioc->ioprio;
3589         struct cfq_data *cfqd = cic_to_cfqd(cic);
3590         struct cfq_queue *cfqq;
3591
3592         /*
3593          * Check whether ioprio has changed.  The condition may trigger
3594          * spuriously on a newly created cic but there's no harm.
3595          */
3596         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3597                 return;
3598
3599         cfqq = cic_to_cfqq(cic, false);
3600         if (cfqq) {
3601                 cfq_put_queue(cfqq);
3602                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3603                 cic_set_cfqq(cic, cfqq, false);
3604         }
3605
3606         cfqq = cic_to_cfqq(cic, true);
3607         if (cfqq)
3608                 cfq_mark_cfqq_prio_changed(cfqq);
3609
3610         cic->ioprio = ioprio;
3611 }
3612
3613 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3614                           pid_t pid, bool is_sync)
3615 {
3616         RB_CLEAR_NODE(&cfqq->rb_node);
3617         RB_CLEAR_NODE(&cfqq->p_node);
3618         INIT_LIST_HEAD(&cfqq->fifo);
3619
3620         cfqq->ref = 0;
3621         cfqq->cfqd = cfqd;
3622
3623         cfq_mark_cfqq_prio_changed(cfqq);
3624
3625         if (is_sync) {
3626                 if (!cfq_class_idle(cfqq))
3627                         cfq_mark_cfqq_idle_window(cfqq);
3628                 cfq_mark_cfqq_sync(cfqq);
3629         }
3630         cfqq->pid = pid;
3631 }
3632
3633 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3634 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3635 {
3636         struct cfq_data *cfqd = cic_to_cfqd(cic);
3637         struct cfq_queue *cfqq;
3638         uint64_t serial_nr;
3639
3640         rcu_read_lock();
3641         serial_nr = bio_blkcg(bio)->css.serial_nr;
3642         rcu_read_unlock();
3643
3644         /*
3645          * Check whether blkcg has changed.  The condition may trigger
3646          * spuriously on a newly created cic but there's no harm.
3647          */
3648         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3649                 return;
3650
3651         /*
3652          * Drop reference to queues.  New queues will be assigned in new
3653          * group upon arrival of fresh requests.
3654          */
3655         cfqq = cic_to_cfqq(cic, false);
3656         if (cfqq) {
3657                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3658                 cic_set_cfqq(cic, NULL, false);
3659                 cfq_put_queue(cfqq);
3660         }
3661
3662         cfqq = cic_to_cfqq(cic, true);
3663         if (cfqq) {
3664                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3665                 cic_set_cfqq(cic, NULL, true);
3666                 cfq_put_queue(cfqq);
3667         }
3668
3669         cic->blkcg_serial_nr = serial_nr;
3670 }
3671 #else
3672 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3673 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3674
3675 static struct cfq_queue **
3676 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3677 {
3678         switch (ioprio_class) {
3679         case IOPRIO_CLASS_RT:
3680                 return &cfqg->async_cfqq[0][ioprio];
3681         case IOPRIO_CLASS_NONE:
3682                 ioprio = IOPRIO_NORM;
3683                 /* fall through */
3684         case IOPRIO_CLASS_BE:
3685                 return &cfqg->async_cfqq[1][ioprio];
3686         case IOPRIO_CLASS_IDLE:
3687                 return &cfqg->async_idle_cfqq;
3688         default:
3689                 BUG();
3690         }
3691 }
3692
3693 static struct cfq_queue *
3694 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3695               struct bio *bio)
3696 {
3697         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3698         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3699         struct cfq_queue **async_cfqq = NULL;
3700         struct cfq_queue *cfqq;
3701         struct cfq_group *cfqg;
3702
3703         rcu_read_lock();
3704         cfqg = cfq_lookup_create_cfqg(cfqd, bio_blkcg(bio));
3705         if (!cfqg) {
3706                 cfqq = &cfqd->oom_cfqq;
3707                 goto out;
3708         }
3709
3710         if (!is_sync) {
3711                 if (!ioprio_valid(cic->ioprio)) {
3712                         struct task_struct *tsk = current;
3713                         ioprio = task_nice_ioprio(tsk);
3714                         ioprio_class = task_nice_ioclass(tsk);
3715                 }
3716                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3717                 cfqq = *async_cfqq;
3718                 if (cfqq)
3719                         goto out;
3720         }
3721
3722         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3723                                      cfqd->queue->node);
3724         if (!cfqq) {
3725                 cfqq = &cfqd->oom_cfqq;
3726                 goto out;
3727         }
3728
3729         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3730         cfq_init_prio_data(cfqq, cic);
3731         cfq_link_cfqq_cfqg(cfqq, cfqg);
3732         cfq_log_cfqq(cfqd, cfqq, "alloced");
3733
3734         if (async_cfqq) {
3735                 /* a new async queue is created, pin and remember */
3736                 cfqq->ref++;
3737                 *async_cfqq = cfqq;
3738         }
3739 out:
3740         cfqq->ref++;
3741         rcu_read_unlock();
3742         return cfqq;
3743 }
3744
3745 static void
3746 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3747 {
3748         unsigned long elapsed = jiffies - ttime->last_end_request;
3749         elapsed = min(elapsed, 2UL * slice_idle);
3750
3751         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3752         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3753         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3754 }
3755
3756 static void
3757 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3758                         struct cfq_io_cq *cic)
3759 {
3760         if (cfq_cfqq_sync(cfqq)) {
3761                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3762                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3763                         cfqd->cfq_slice_idle);
3764         }
3765 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3766         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3767 #endif
3768 }
3769
3770 static void
3771 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3772                        struct request *rq)
3773 {
3774         sector_t sdist = 0;
3775         sector_t n_sec = blk_rq_sectors(rq);
3776         if (cfqq->last_request_pos) {
3777                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3778                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3779                 else
3780                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3781         }
3782
3783         cfqq->seek_history <<= 1;
3784         if (blk_queue_nonrot(cfqd->queue))
3785                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3786         else
3787                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3788 }
3789
3790 /*
3791  * Disable idle window if the process thinks too long or seeks so much that
3792  * it doesn't matter
3793  */
3794 static void
3795 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3796                        struct cfq_io_cq *cic)
3797 {
3798         int old_idle, enable_idle;
3799
3800         /*
3801          * Don't idle for async or idle io prio class
3802          */
3803         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3804                 return;
3805
3806         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3807
3808         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3809                 cfq_mark_cfqq_deep(cfqq);
3810
3811         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3812                 enable_idle = 0;
3813         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3814                  !cfqd->cfq_slice_idle ||
3815                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3816                 enable_idle = 0;
3817         else if (sample_valid(cic->ttime.ttime_samples)) {
3818                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3819                         enable_idle = 0;
3820                 else
3821                         enable_idle = 1;
3822         }
3823
3824         if (old_idle != enable_idle) {
3825                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3826                 if (enable_idle)
3827                         cfq_mark_cfqq_idle_window(cfqq);
3828                 else
3829                         cfq_clear_cfqq_idle_window(cfqq);
3830         }
3831 }
3832
3833 /*
3834  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3835  * no or if we aren't sure, a 1 will cause a preempt.
3836  */
3837 static bool
3838 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3839                    struct request *rq)
3840 {
3841         struct cfq_queue *cfqq;
3842
3843         cfqq = cfqd->active_queue;
3844         if (!cfqq)
3845                 return false;
3846
3847         if (cfq_class_idle(new_cfqq))
3848                 return false;
3849
3850         if (cfq_class_idle(cfqq))
3851                 return true;
3852
3853         /*
3854          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3855          */
3856         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3857                 return false;
3858
3859         /*
3860          * if the new request is sync, but the currently running queue is
3861          * not, let the sync request have priority.
3862          */
3863         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3864                 return true;
3865
3866         if (new_cfqq->cfqg != cfqq->cfqg)
3867                 return false;
3868
3869         if (cfq_slice_used(cfqq))
3870                 return true;
3871
3872         /* Allow preemption only if we are idling on sync-noidle tree */
3873         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3874             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3875             new_cfqq->service_tree->count == 2 &&
3876             RB_EMPTY_ROOT(&cfqq->sort_list))
3877                 return true;
3878
3879         /*
3880          * So both queues are sync. Let the new request get disk time if
3881          * it's a metadata request and the current queue is doing regular IO.
3882          */
3883         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3884                 return true;
3885
3886         /*
3887          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3888          */
3889         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3890                 return true;
3891
3892         /* An idle queue should not be idle now for some reason */
3893         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3894                 return true;
3895
3896         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3897                 return false;
3898
3899         /*
3900          * if this request is as-good as one we would expect from the
3901          * current cfqq, let it preempt
3902          */
3903         if (cfq_rq_close(cfqd, cfqq, rq))
3904                 return true;
3905
3906         return false;
3907 }
3908
3909 /*
3910  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3911  * let it have half of its nominal slice.
3912  */
3913 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3914 {
3915         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3916
3917         cfq_log_cfqq(cfqd, cfqq, "preempt");
3918         cfq_slice_expired(cfqd, 1);
3919
3920         /*
3921          * workload type is changed, don't save slice, otherwise preempt
3922          * doesn't happen
3923          */
3924         if (old_type != cfqq_type(cfqq))
3925                 cfqq->cfqg->saved_wl_slice = 0;
3926
3927         /*
3928          * Put the new queue at the front of the of the current list,
3929          * so we know that it will be selected next.
3930          */
3931         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3932
3933         cfq_service_tree_add(cfqd, cfqq, 1);
3934
3935         cfqq->slice_end = 0;
3936         cfq_mark_cfqq_slice_new(cfqq);
3937 }
3938
3939 /*
3940  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3941  * something we should do about it
3942  */
3943 static void
3944 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3945                 struct request *rq)
3946 {
3947         struct cfq_io_cq *cic = RQ_CIC(rq);
3948
3949         cfqd->rq_queued++;
3950         if (rq->cmd_flags & REQ_PRIO)
3951                 cfqq->prio_pending++;
3952
3953         cfq_update_io_thinktime(cfqd, cfqq, cic);
3954         cfq_update_io_seektime(cfqd, cfqq, rq);
3955         cfq_update_idle_window(cfqd, cfqq, cic);
3956
3957         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3958
3959         if (cfqq == cfqd->active_queue) {
3960                 /*
3961                  * Remember that we saw a request from this process, but
3962                  * don't start queuing just yet. Otherwise we risk seeing lots
3963                  * of tiny requests, because we disrupt the normal plugging
3964                  * and merging. If the request is already larger than a single
3965                  * page, let it rip immediately. For that case we assume that
3966                  * merging is already done. Ditto for a busy system that
3967                  * has other work pending, don't risk delaying until the
3968                  * idle timer unplug to continue working.
3969                  */
3970                 if (cfq_cfqq_wait_request(cfqq)) {
3971                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3972                             cfqd->busy_queues > 1) {
3973                                 cfq_del_timer(cfqd, cfqq);
3974                                 cfq_clear_cfqq_wait_request(cfqq);
3975                                 __blk_run_queue(cfqd->queue);
3976                         } else {
3977                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3978                                 cfq_mark_cfqq_must_dispatch(cfqq);
3979                         }
3980                 }
3981         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3982                 /*
3983                  * not the active queue - expire current slice if it is
3984                  * idle and has expired it's mean thinktime or this new queue
3985                  * has some old slice time left and is of higher priority or
3986                  * this new queue is RT and the current one is BE
3987                  */
3988                 cfq_preempt_queue(cfqd, cfqq);
3989                 __blk_run_queue(cfqd->queue);
3990         }
3991 }
3992
3993 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3994 {
3995         struct cfq_data *cfqd = q->elevator->elevator_data;
3996         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3997
3998         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3999         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4000
4001         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4002         list_add_tail(&rq->queuelist, &cfqq->fifo);
4003         cfq_add_rq_rb(rq);
4004         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4005                                  rq->cmd_flags);
4006         cfq_rq_enqueued(cfqd, cfqq, rq);
4007 }
4008
4009 /*
4010  * Update hw_tag based on peak queue depth over 50 samples under
4011  * sufficient load.
4012  */
4013 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4014 {
4015         struct cfq_queue *cfqq = cfqd->active_queue;
4016
4017         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4018                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4019
4020         if (cfqd->hw_tag == 1)
4021                 return;
4022
4023         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4024             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4025                 return;
4026
4027         /*
4028          * If active queue hasn't enough requests and can idle, cfq might not
4029          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4030          * case
4031          */
4032         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4033             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4034             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4035                 return;
4036
4037         if (cfqd->hw_tag_samples++ < 50)
4038                 return;
4039
4040         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4041                 cfqd->hw_tag = 1;
4042         else
4043                 cfqd->hw_tag = 0;
4044 }
4045
4046 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4047 {
4048         struct cfq_io_cq *cic = cfqd->active_cic;
4049
4050         /* If the queue already has requests, don't wait */
4051         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4052                 return false;
4053
4054         /* If there are other queues in the group, don't wait */
4055         if (cfqq->cfqg->nr_cfqq > 1)
4056                 return false;
4057
4058         /* the only queue in the group, but think time is big */
4059         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4060                 return false;
4061
4062         if (cfq_slice_used(cfqq))
4063                 return true;
4064
4065         /* if slice left is less than think time, wait busy */
4066         if (cic && sample_valid(cic->ttime.ttime_samples)
4067             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4068                 return true;
4069
4070         /*
4071          * If think times is less than a jiffy than ttime_mean=0 and above
4072          * will not be true. It might happen that slice has not expired yet
4073          * but will expire soon (4-5 ns) during select_queue(). To cover the
4074          * case where think time is less than a jiffy, mark the queue wait
4075          * busy if only 1 jiffy is left in the slice.
4076          */
4077         if (cfqq->slice_end - jiffies == 1)
4078                 return true;
4079
4080         return false;
4081 }
4082
4083 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4084 {
4085         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4086         struct cfq_data *cfqd = cfqq->cfqd;
4087         const int sync = rq_is_sync(rq);
4088         unsigned long now;
4089
4090         now = jiffies;
4091         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4092                      !!(rq->cmd_flags & REQ_NOIDLE));
4093
4094         cfq_update_hw_tag(cfqd);
4095
4096         WARN_ON(!cfqd->rq_in_driver);
4097         WARN_ON(!cfqq->dispatched);
4098         cfqd->rq_in_driver--;
4099         cfqq->dispatched--;
4100         (RQ_CFQG(rq))->dispatched--;
4101         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4102                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4103
4104         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4105
4106         if (sync) {
4107                 struct cfq_rb_root *st;
4108
4109                 RQ_CIC(rq)->ttime.last_end_request = now;
4110
4111                 if (cfq_cfqq_on_rr(cfqq))
4112                         st = cfqq->service_tree;
4113                 else
4114                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4115                                         cfqq_type(cfqq));
4116
4117                 st->ttime.last_end_request = now;
4118                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4119                         cfqd->last_delayed_sync = now;
4120         }
4121
4122 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4123         cfqq->cfqg->ttime.last_end_request = now;
4124 #endif
4125
4126         /*
4127          * If this is the active queue, check if it needs to be expired,
4128          * or if we want to idle in case it has no pending requests.
4129          */
4130         if (cfqd->active_queue == cfqq) {
4131                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4132
4133                 if (cfq_cfqq_slice_new(cfqq)) {
4134                         cfq_set_prio_slice(cfqd, cfqq);
4135                         cfq_clear_cfqq_slice_new(cfqq);
4136                 }
4137
4138                 /*
4139                  * Should we wait for next request to come in before we expire
4140                  * the queue.
4141                  */
4142                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4143                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4144                         if (!cfqd->cfq_slice_idle)
4145                                 extend_sl = cfqd->cfq_group_idle;
4146                         cfqq->slice_end = jiffies + extend_sl;
4147                         cfq_mark_cfqq_wait_busy(cfqq);
4148                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4149                 }
4150
4151                 /*
4152                  * Idling is not enabled on:
4153                  * - expired queues
4154                  * - idle-priority queues
4155                  * - async queues
4156                  * - queues with still some requests queued
4157                  * - when there is a close cooperator
4158                  */
4159                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4160                         cfq_slice_expired(cfqd, 1);
4161                 else if (sync && cfqq_empty &&
4162                          !cfq_close_cooperator(cfqd, cfqq)) {
4163                         cfq_arm_slice_timer(cfqd);
4164                 }
4165         }
4166
4167         if (!cfqd->rq_in_driver)
4168                 cfq_schedule_dispatch(cfqd);
4169 }
4170
4171 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4172 {
4173         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4174                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4175                 return ELV_MQUEUE_MUST;
4176         }
4177
4178         return ELV_MQUEUE_MAY;
4179 }
4180
4181 static int cfq_may_queue(struct request_queue *q, int rw)
4182 {
4183         struct cfq_data *cfqd = q->elevator->elevator_data;
4184         struct task_struct *tsk = current;
4185         struct cfq_io_cq *cic;
4186         struct cfq_queue *cfqq;
4187
4188         /*
4189          * don't force setup of a queue from here, as a call to may_queue
4190          * does not necessarily imply that a request actually will be queued.
4191          * so just lookup a possibly existing queue, or return 'may queue'
4192          * if that fails
4193          */
4194         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4195         if (!cic)
4196                 return ELV_MQUEUE_MAY;
4197
4198         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4199         if (cfqq) {
4200                 cfq_init_prio_data(cfqq, cic);
4201
4202                 return __cfq_may_queue(cfqq);
4203         }
4204
4205         return ELV_MQUEUE_MAY;
4206 }
4207
4208 /*
4209  * queue lock held here
4210  */
4211 static void cfq_put_request(struct request *rq)
4212 {
4213         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4214
4215         if (cfqq) {
4216                 const int rw = rq_data_dir(rq);
4217
4218                 BUG_ON(!cfqq->allocated[rw]);
4219                 cfqq->allocated[rw]--;
4220
4221                 /* Put down rq reference on cfqg */
4222                 cfqg_put(RQ_CFQG(rq));
4223                 rq->elv.priv[0] = NULL;
4224                 rq->elv.priv[1] = NULL;
4225
4226                 cfq_put_queue(cfqq);
4227         }
4228 }
4229
4230 static struct cfq_queue *
4231 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4232                 struct cfq_queue *cfqq)
4233 {
4234         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4235         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4236         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4237         cfq_put_queue(cfqq);
4238         return cic_to_cfqq(cic, 1);
4239 }
4240
4241 /*
4242  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4243  * was the last process referring to said cfqq.
4244  */
4245 static struct cfq_queue *
4246 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4247 {
4248         if (cfqq_process_refs(cfqq) == 1) {
4249                 cfqq->pid = current->pid;
4250                 cfq_clear_cfqq_coop(cfqq);
4251                 cfq_clear_cfqq_split_coop(cfqq);
4252                 return cfqq;
4253         }
4254
4255         cic_set_cfqq(cic, NULL, 1);
4256
4257         cfq_put_cooperator(cfqq);
4258
4259         cfq_put_queue(cfqq);
4260         return NULL;
4261 }
4262 /*
4263  * Allocate cfq data structures associated with this request.
4264  */
4265 static int
4266 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4267                 gfp_t gfp_mask)
4268 {
4269         struct cfq_data *cfqd = q->elevator->elevator_data;
4270         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4271         const int rw = rq_data_dir(rq);
4272         const bool is_sync = rq_is_sync(rq);
4273         struct cfq_queue *cfqq;
4274
4275         spin_lock_irq(q->queue_lock);
4276
4277         check_ioprio_changed(cic, bio);
4278         check_blkcg_changed(cic, bio);
4279 new_queue:
4280         cfqq = cic_to_cfqq(cic, is_sync);
4281         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4282                 if (cfqq)
4283                         cfq_put_queue(cfqq);
4284                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4285                 cic_set_cfqq(cic, cfqq, is_sync);
4286         } else {
4287                 /*
4288                  * If the queue was seeky for too long, break it apart.
4289                  */
4290                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4291                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4292                         cfqq = split_cfqq(cic, cfqq);
4293                         if (!cfqq)
4294                                 goto new_queue;
4295                 }
4296
4297                 /*
4298                  * Check to see if this queue is scheduled to merge with
4299                  * another, closely cooperating queue.  The merging of
4300                  * queues happens here as it must be done in process context.
4301                  * The reference on new_cfqq was taken in merge_cfqqs.
4302                  */
4303                 if (cfqq->new_cfqq)
4304                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4305         }
4306
4307         cfqq->allocated[rw]++;
4308
4309         cfqq->ref++;
4310         cfqg_get(cfqq->cfqg);
4311         rq->elv.priv[0] = cfqq;
4312         rq->elv.priv[1] = cfqq->cfqg;
4313         spin_unlock_irq(q->queue_lock);
4314         return 0;
4315 }
4316
4317 static void cfq_kick_queue(struct work_struct *work)
4318 {
4319         struct cfq_data *cfqd =
4320                 container_of(work, struct cfq_data, unplug_work);
4321         struct request_queue *q = cfqd->queue;
4322
4323         spin_lock_irq(q->queue_lock);
4324         __blk_run_queue(cfqd->queue);
4325         spin_unlock_irq(q->queue_lock);
4326 }
4327
4328 /*
4329  * Timer running if the active_queue is currently idling inside its time slice
4330  */
4331 static void cfq_idle_slice_timer(unsigned long data)
4332 {
4333         struct cfq_data *cfqd = (struct cfq_data *) data;
4334         struct cfq_queue *cfqq;
4335         unsigned long flags;
4336         int timed_out = 1;
4337
4338         cfq_log(cfqd, "idle timer fired");
4339
4340         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4341
4342         cfqq = cfqd->active_queue;
4343         if (cfqq) {
4344                 timed_out = 0;
4345
4346                 /*
4347                  * We saw a request before the queue expired, let it through
4348                  */
4349                 if (cfq_cfqq_must_dispatch(cfqq))
4350                         goto out_kick;
4351
4352                 /*
4353                  * expired
4354                  */
4355                 if (cfq_slice_used(cfqq))
4356                         goto expire;
4357
4358                 /*
4359                  * only expire and reinvoke request handler, if there are
4360                  * other queues with pending requests
4361                  */
4362                 if (!cfqd->busy_queues)
4363                         goto out_cont;
4364
4365                 /*
4366                  * not expired and it has a request pending, let it dispatch
4367                  */
4368                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4369                         goto out_kick;
4370
4371                 /*
4372                  * Queue depth flag is reset only when the idle didn't succeed
4373                  */
4374                 cfq_clear_cfqq_deep(cfqq);
4375         }
4376 expire:
4377         cfq_slice_expired(cfqd, timed_out);
4378 out_kick:
4379         cfq_schedule_dispatch(cfqd);
4380 out_cont:
4381         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4382 }
4383
4384 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4385 {
4386         del_timer_sync(&cfqd->idle_slice_timer);
4387         cancel_work_sync(&cfqd->unplug_work);
4388 }
4389
4390 static void cfq_exit_queue(struct elevator_queue *e)
4391 {
4392         struct cfq_data *cfqd = e->elevator_data;
4393         struct request_queue *q = cfqd->queue;
4394
4395         cfq_shutdown_timer_wq(cfqd);
4396
4397         spin_lock_irq(q->queue_lock);
4398
4399         if (cfqd->active_queue)
4400                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4401
4402         spin_unlock_irq(q->queue_lock);
4403
4404         cfq_shutdown_timer_wq(cfqd);
4405
4406 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4407         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4408 #else
4409         kfree(cfqd->root_group);
4410 #endif
4411         kfree(cfqd);
4412 }
4413
4414 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4415 {
4416         struct cfq_data *cfqd;
4417         struct blkcg_gq *blkg __maybe_unused;
4418         int i, ret;
4419         struct elevator_queue *eq;
4420
4421         eq = elevator_alloc(q, e);
4422         if (!eq)
4423                 return -ENOMEM;
4424
4425         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4426         if (!cfqd) {
4427                 kobject_put(&eq->kobj);
4428                 return -ENOMEM;
4429         }
4430         eq->elevator_data = cfqd;
4431
4432         cfqd->queue = q;
4433         spin_lock_irq(q->queue_lock);
4434         q->elevator = eq;
4435         spin_unlock_irq(q->queue_lock);
4436
4437         /* Init root service tree */
4438         cfqd->grp_service_tree = CFQ_RB_ROOT;
4439
4440         /* Init root group and prefer root group over other groups by default */
4441 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4442         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4443         if (ret)
4444                 goto out_free;
4445
4446         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4447 #else
4448         ret = -ENOMEM;
4449         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4450                                         GFP_KERNEL, cfqd->queue->node);
4451         if (!cfqd->root_group)
4452                 goto out_free;
4453
4454         cfq_init_cfqg_base(cfqd->root_group);
4455 #endif
4456         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4457         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4458
4459         /*
4460          * Not strictly needed (since RB_ROOT just clears the node and we
4461          * zeroed cfqd on alloc), but better be safe in case someone decides
4462          * to add magic to the rb code
4463          */
4464         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4465                 cfqd->prio_trees[i] = RB_ROOT;
4466
4467         /*
4468          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4469          * Grab a permanent reference to it, so that the normal code flow
4470          * will not attempt to free it.  oom_cfqq is linked to root_group
4471          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4472          * the reference from linking right away.
4473          */
4474         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4475         cfqd->oom_cfqq.ref++;
4476
4477         spin_lock_irq(q->queue_lock);
4478         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4479         cfqg_put(cfqd->root_group);
4480         spin_unlock_irq(q->queue_lock);
4481
4482         init_timer(&cfqd->idle_slice_timer);
4483         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4484         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4485
4486         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4487
4488         cfqd->cfq_quantum = cfq_quantum;
4489         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4490         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4491         cfqd->cfq_back_max = cfq_back_max;
4492         cfqd->cfq_back_penalty = cfq_back_penalty;
4493         cfqd->cfq_slice[0] = cfq_slice_async;
4494         cfqd->cfq_slice[1] = cfq_slice_sync;
4495         cfqd->cfq_target_latency = cfq_target_latency;
4496         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4497         cfqd->cfq_slice_idle = cfq_slice_idle;
4498         cfqd->cfq_group_idle = cfq_group_idle;
4499         cfqd->cfq_latency = 1;
4500         cfqd->hw_tag = -1;
4501         /*
4502          * we optimistically start assuming sync ops weren't delayed in last
4503          * second, in order to have larger depth for async operations.
4504          */
4505         cfqd->last_delayed_sync = jiffies - HZ;
4506         return 0;
4507
4508 out_free:
4509         kfree(cfqd);
4510         kobject_put(&eq->kobj);
4511         return ret;
4512 }
4513
4514 static void cfq_registered_queue(struct request_queue *q)
4515 {
4516         struct elevator_queue *e = q->elevator;
4517         struct cfq_data *cfqd = e->elevator_data;
4518
4519         /*
4520          * Default to IOPS mode with no idling for SSDs
4521          */
4522         if (blk_queue_nonrot(q))
4523                 cfqd->cfq_slice_idle = 0;
4524 }
4525
4526 /*
4527  * sysfs parts below -->
4528  */
4529 static ssize_t
4530 cfq_var_show(unsigned int var, char *page)
4531 {
4532         return sprintf(page, "%u\n", var);
4533 }
4534
4535 static ssize_t
4536 cfq_var_store(unsigned int *var, const char *page, size_t count)
4537 {
4538         char *p = (char *) page;
4539
4540         *var = simple_strtoul(p, &p, 10);
4541         return count;
4542 }
4543
4544 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4545 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4546 {                                                                       \
4547         struct cfq_data *cfqd = e->elevator_data;                       \
4548         unsigned int __data = __VAR;                                    \
4549         if (__CONV)                                                     \
4550                 __data = jiffies_to_msecs(__data);                      \
4551         return cfq_var_show(__data, (page));                            \
4552 }
4553 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4554 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4555 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4556 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4557 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4558 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4559 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4560 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4561 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4562 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4563 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4564 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4565 #undef SHOW_FUNCTION
4566
4567 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4568 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4569 {                                                                       \
4570         struct cfq_data *cfqd = e->elevator_data;                       \
4571         unsigned int __data;                                            \
4572         int ret = cfq_var_store(&__data, (page), count);                \
4573         if (__data < (MIN))                                             \
4574                 __data = (MIN);                                         \
4575         else if (__data > (MAX))                                        \
4576                 __data = (MAX);                                         \
4577         if (__CONV)                                                     \
4578                 *(__PTR) = msecs_to_jiffies(__data);                    \
4579         else                                                            \
4580                 *(__PTR) = __data;                                      \
4581         return ret;                                                     \
4582 }
4583 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4584 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4585                 UINT_MAX, 1);
4586 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4587                 UINT_MAX, 1);
4588 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4589 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4590                 UINT_MAX, 0);
4591 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4592 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4593 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4594 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4595 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4596                 UINT_MAX, 0);
4597 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4598 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4599 #undef STORE_FUNCTION
4600
4601 #define CFQ_ATTR(name) \
4602         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4603
4604 static struct elv_fs_entry cfq_attrs[] = {
4605         CFQ_ATTR(quantum),
4606         CFQ_ATTR(fifo_expire_sync),
4607         CFQ_ATTR(fifo_expire_async),
4608         CFQ_ATTR(back_seek_max),
4609         CFQ_ATTR(back_seek_penalty),
4610         CFQ_ATTR(slice_sync),
4611         CFQ_ATTR(slice_async),
4612         CFQ_ATTR(slice_async_rq),
4613         CFQ_ATTR(slice_idle),
4614         CFQ_ATTR(group_idle),
4615         CFQ_ATTR(low_latency),
4616         CFQ_ATTR(target_latency),
4617         __ATTR_NULL
4618 };
4619
4620 static struct elevator_type iosched_cfq = {
4621         .ops = {
4622                 .elevator_merge_fn =            cfq_merge,
4623                 .elevator_merged_fn =           cfq_merged_request,
4624                 .elevator_merge_req_fn =        cfq_merged_requests,
4625                 .elevator_allow_merge_fn =      cfq_allow_merge,
4626                 .elevator_bio_merged_fn =       cfq_bio_merged,
4627                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4628                 .elevator_add_req_fn =          cfq_insert_request,
4629                 .elevator_activate_req_fn =     cfq_activate_request,
4630                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4631                 .elevator_completed_req_fn =    cfq_completed_request,
4632                 .elevator_former_req_fn =       elv_rb_former_request,
4633                 .elevator_latter_req_fn =       elv_rb_latter_request,
4634                 .elevator_init_icq_fn =         cfq_init_icq,
4635                 .elevator_exit_icq_fn =         cfq_exit_icq,
4636                 .elevator_set_req_fn =          cfq_set_request,
4637                 .elevator_put_req_fn =          cfq_put_request,
4638                 .elevator_may_queue_fn =        cfq_may_queue,
4639                 .elevator_init_fn =             cfq_init_queue,
4640                 .elevator_exit_fn =             cfq_exit_queue,
4641                 .elevator_registered_fn =       cfq_registered_queue,
4642         },
4643         .icq_size       =       sizeof(struct cfq_io_cq),
4644         .icq_align      =       __alignof__(struct cfq_io_cq),
4645         .elevator_attrs =       cfq_attrs,
4646         .elevator_name  =       "cfq",
4647         .elevator_owner =       THIS_MODULE,
4648 };
4649
4650 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4651 static struct blkcg_policy blkcg_policy_cfq = {
4652         .cpd_size               = sizeof(struct cfq_group_data),
4653         .cftypes                = cfq_blkcg_files,
4654
4655         .cpd_init_fn            = cfq_cpd_init,
4656         .pd_alloc_fn            = cfq_pd_alloc,
4657         .pd_init_fn             = cfq_pd_init,
4658         .pd_offline_fn          = cfq_pd_offline,
4659         .pd_free_fn             = cfq_pd_free,
4660         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4661 };
4662 #endif
4663
4664 static int __init cfq_init(void)
4665 {
4666         int ret;
4667
4668         /*
4669          * could be 0 on HZ < 1000 setups
4670          */
4671         if (!cfq_slice_async)
4672                 cfq_slice_async = 1;
4673         if (!cfq_slice_idle)
4674                 cfq_slice_idle = 1;
4675
4676 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4677         if (!cfq_group_idle)
4678                 cfq_group_idle = 1;
4679
4680         ret = blkcg_policy_register(&blkcg_policy_cfq);
4681         if (ret)
4682                 return ret;
4683 #else
4684         cfq_group_idle = 0;
4685 #endif
4686
4687         ret = -ENOMEM;
4688         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4689         if (!cfq_pool)
4690                 goto err_pol_unreg;
4691
4692         ret = elv_register(&iosched_cfq);
4693         if (ret)
4694                 goto err_free_pool;
4695
4696         return 0;
4697
4698 err_free_pool:
4699         kmem_cache_destroy(cfq_pool);
4700 err_pol_unreg:
4701 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4702         blkcg_policy_unregister(&blkcg_policy_cfq);
4703 #endif
4704         return ret;
4705 }
4706
4707 static void __exit cfq_exit(void)
4708 {
4709 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4710         blkcg_policy_unregister(&blkcg_policy_cfq);
4711 #endif
4712         elv_unregister(&iosched_cfq);
4713         kmem_cache_destroy(cfq_pool);
4714 }
4715
4716 module_init(cfq_init);
4717 module_exit(cfq_exit);
4718
4719 MODULE_AUTHOR("Jens Axboe");
4720 MODULE_LICENSE("GPL");
4721 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");