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