block/cfq-iosched.c

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