block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/jiffies.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16 #include <linux/blktrace_api.h>
  17 #include <linux/blk-cgroup.h>
  18 #include "blk.h"
  19
  20 /*
  21  * tunables
  22  */
  23 /* max queue in one round of service */
  24 static const int cfq_quantum = 8;
  25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  26 /* maximum backwards seek, in KiB */
  27 static const int cfq_back_max = 16 * 1024;
  28 /* penalty of a backwards seek */
  29 static const int cfq_back_penalty = 2;
  30 static const int cfq_slice_sync = HZ / 10;
  31 static int cfq_slice_async = HZ / 25;
  32 static const int cfq_slice_async_rq = 2;
  33 static int cfq_slice_idle = HZ / 125;
  34 static int cfq_group_idle = HZ / 125;
  35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
  36 static const int cfq_hist_divisor = 4;
  37
  38 /*
  39  * offset from end of service tree
  40  */
  41 #define CFQ_IDLE_DELAY          (HZ / 5)
  42
  43 /*
  44  * below this threshold, we consider thinktime immediate
  45  */
  46 #define CFQ_MIN_TT              (2)
  47
  48 #define CFQ_SLICE_SCALE         (5)
  49 #define CFQ_HW_QUEUE_MIN        (5)
  50 #define CFQ_SERVICE_SHIFT       12
  51
  52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  56
  57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  60
  61 static struct kmem_cache *cfq_pool;
  62
  63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  69
  70 /* blkio-related constants */
  71 #define CFQ_WEIGHT_MIN          10
  72 #define CFQ_WEIGHT_MAX          1000
  73 #define CFQ_WEIGHT_DEFAULT      500
  74
  75 struct cfq_ttime {
  76         unsigned long last_end_request;
  77
  78         unsigned long ttime_total;
  79         unsigned long ttime_samples;
  80         unsigned long ttime_mean;
  81 };
  82
  83 /*
  84  * Most of our rbtree usage is for sorting with min extraction, so
  85  * if we cache the leftmost node we don't have to walk down the tree
  86  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  87  * move this into the elevator for the rq sorting as well.
  88  */
  89 struct cfq_rb_root {
  90         struct rb_root rb;
  91         struct rb_node *left;
  92         unsigned count;
  93         u64 min_vdisktime;
  94         struct cfq_ttime ttime;
  95 };
  96 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  97                         .ttime = {.last_end_request = jiffies,},}
  98
  99 /*
 100  * Per process-grouping structure
 101  */
 102 struct cfq_queue {
 103         /* reference count */
 104         int ref;
 105         /* various state flags, see below */
 106         unsigned int flags;
 107         /* parent cfq_data */
 108         struct cfq_data *cfqd;
 109         /* service_tree member */
 110         struct rb_node rb_node;
 111         /* service_tree key */
 112         unsigned long rb_key;
 113         /* prio tree member */
 114         struct rb_node p_node;
 115         /* prio tree root we belong to, if any */
 116         struct rb_root *p_root;
 117         /* sorted list of pending requests */
 118         struct rb_root sort_list;
 119         /* if fifo isn't expired, next request to serve */
 120         struct request *next_rq;
 121         /* requests queued in sort_list */
 122         int queued[2];
 123         /* currently allocated requests */
 124         int allocated[2];
 125         /* fifo list of requests in sort_list */
 126         struct list_head fifo;
 127
 128         /* time when queue got scheduled in to dispatch first request. */
 129         unsigned long dispatch_start;
 130         unsigned int allocated_slice;
 131         unsigned int slice_dispatch;
 132         /* time when first request from queue completed and slice started. */
 133         unsigned long slice_start;
 134         unsigned long slice_end;
 135         long slice_resid;
 136
 137         /* pending priority requests */
 138         int prio_pending;
 139         /* number of requests that are on the dispatch list or inside driver */
 140         int dispatched;
 141
 142         /* io prio of this group */
 143         unsigned short ioprio, org_ioprio;
 144         unsigned short ioprio_class;
 145
 146         pid_t pid;
 147
 148         u32 seek_history;
 149         sector_t last_request_pos;
 150
 151         struct cfq_rb_root *service_tree;
 152         struct cfq_queue *new_cfqq;
 153         struct cfq_group *cfqg;
 154         /* Number of sectors dispatched from queue in single dispatch round */
 155         unsigned long nr_sectors;
 156 };
 157
 158 /*
 159  * First index in the service_trees.
 160  * IDLE is handled separately, so it has negative index
 161  */
 162 enum wl_class_t {
 163         BE_WORKLOAD = 0,
 164         RT_WORKLOAD = 1,
 165         IDLE_WORKLOAD = 2,
 166         CFQ_PRIO_NR,
 167 };
 168
 169 /*
 170  * Second index in the service_trees.
 171  */
 172 enum wl_type_t {
 173         ASYNC_WORKLOAD = 0,
 174         SYNC_NOIDLE_WORKLOAD = 1,
 175         SYNC_WORKLOAD = 2
 176 };
 177
 178 struct cfqg_stats {
 179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 180         /* total bytes transferred */
 181         struct blkg_rwstat              service_bytes;
 182         /* total IOs serviced, post merge */
 183         struct blkg_rwstat              serviced;
 184         /* number of ios merged */
 185         struct blkg_rwstat              merged;
 186         /* total time spent on device in ns, may not be accurate w/ queueing */
 187         struct blkg_rwstat              service_time;
 188         /* total time spent waiting in scheduler queue in ns */
 189         struct blkg_rwstat              wait_time;
 190         /* number of IOs queued up */
 191         struct blkg_rwstat              queued;
 192         /* total sectors transferred */
 193         struct blkg_stat                sectors;
 194         /* total disk time and nr sectors dispatched by this group */
 195         struct blkg_stat                time;
 196 #ifdef CONFIG_DEBUG_BLK_CGROUP
 197         /* time not charged to this cgroup */
 198         struct blkg_stat                unaccounted_time;
 199         /* sum of number of ios queued across all samples */
 200         struct blkg_stat                avg_queue_size_sum;
 201         /* count of samples taken for average */
 202         struct blkg_stat                avg_queue_size_samples;
 203         /* how many times this group has been removed from service tree */
 204         struct blkg_stat                dequeue;
 205         /* total time spent waiting for it to be assigned a timeslice. */
 206         struct blkg_stat                group_wait_time;
 207         /* time spent idling for this blkcg_gq */
 208         struct blkg_stat                idle_time;
 209         /* total time with empty current active q with other requests queued */
 210         struct blkg_stat                empty_time;
 211         /* fields after this shouldn't be cleared on stat reset */
 212         uint64_t                        start_group_wait_time;
 213         uint64_t                        start_idle_time;
 214         uint64_t                        start_empty_time;
 215         uint16_t                        flags;
 216 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 217 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 218 };
 219
 220 /* Per-cgroup data */
 221 struct cfq_group_data {
 222         /* must be the first member */
 223         struct blkcg_policy_data cpd;
 224
 225         unsigned int weight;
 226         unsigned int leaf_weight;
 227 };
 228
 229 /* This is per cgroup per device grouping structure */
 230 struct cfq_group {
 231         /* must be the first member */
 232         struct blkg_policy_data pd;
 233
 234         /* group service_tree member */
 235         struct rb_node rb_node;
 236
 237         /* group service_tree key */
 238         u64 vdisktime;
 239
 240         /*
 241          * The number of active cfqgs and sum of their weights under this
 242          * cfqg.  This covers this cfqg's leaf_weight and all children's
 243          * weights, but does not cover weights of further descendants.
 244          *
 245          * If a cfqg is on the service tree, it's active.  An active cfqg
 246          * also activates its parent and contributes to the children_weight
 247          * of the parent.
 248          */
 249         int nr_active;
 250         unsigned int children_weight;
 251
 252         /*
 253          * vfraction is the fraction of vdisktime that the tasks in this
 254          * cfqg are entitled to.  This is determined by compounding the
 255          * ratios walking up from this cfqg to the root.
 256          *
 257          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 258          * vfractions on a service tree is approximately 1.  The sum may
 259          * deviate a bit due to rounding errors and fluctuations caused by
 260          * cfqgs entering and leaving the service tree.
 261          */
 262         unsigned int vfraction;
 263
 264         /*
 265          * There are two weights - (internal) weight is the weight of this
 266          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 267          * this cfqg against the child cfqgs.  For the root cfqg, both
 268          * weights are kept in sync for backward compatibility.
 269          */
 270         unsigned int weight;
 271         unsigned int new_weight;
 272         unsigned int dev_weight;
 273
 274         unsigned int leaf_weight;
 275         unsigned int new_leaf_weight;
 276         unsigned int dev_leaf_weight;
 277
 278         /* number of cfqq currently on this group */
 279         int nr_cfqq;
 280
 281         /*
 282          * Per group busy queues average. Useful for workload slice calc. We
 283          * create the array for each prio class but at run time it is used
 284          * only for RT and BE class and slot for IDLE class remains unused.
 285          * This is primarily done to avoid confusion and a gcc warning.
 286          */
 287         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 288         /*
 289          * rr lists of queues with requests. We maintain service trees for
 290          * RT and BE classes. These trees are subdivided in subclasses
 291          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 292          * class there is no subclassification and all the cfq queues go on
 293          * a single tree service_tree_idle.
 294          * Counts are embedded in the cfq_rb_root
 295          */
 296         struct cfq_rb_root service_trees[2][3];
 297         struct cfq_rb_root service_tree_idle;
 298
 299         unsigned long saved_wl_slice;
 300         enum wl_type_t saved_wl_type;
 301         enum wl_class_t saved_wl_class;
 302
 303         /* number of requests that are on the dispatch list or inside driver */
 304         int dispatched;
 305         struct cfq_ttime ttime;
 306         struct cfqg_stats stats;        /* stats for this cfqg */
 307         struct cfqg_stats dead_stats;   /* stats pushed from dead children */
 308
 309         /* async queue for each priority case */
 310         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 311         struct cfq_queue *async_idle_cfqq;
 312
 313 };
 314
 315 struct cfq_io_cq {
 316         struct io_cq            icq;            /* must be the first member */
 317         struct cfq_queue        *cfqq[2];
 318         struct cfq_ttime        ttime;
 319         int                     ioprio;         /* the current ioprio */
 320 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 321         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 322 #endif
 323 };
 324
 325 /*
 326  * Per block device queue structure
 327  */
 328 struct cfq_data {
 329         struct request_queue *queue;
 330         /* Root service tree for cfq_groups */
 331         struct cfq_rb_root grp_service_tree;
 332         struct cfq_group *root_group;
 333
 334         /*
 335          * The priority currently being served
 336          */
 337         enum wl_class_t serving_wl_class;
 338         enum wl_type_t serving_wl_type;
 339         unsigned long workload_expires;
 340         struct cfq_group *serving_group;
 341
 342         /*
 343          * Each priority tree is sorted by next_request position.  These
 344          * trees are used when determining if two or more queues are
 345          * interleaving requests (see cfq_close_cooperator).
 346          */
 347         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 348
 349         unsigned int busy_queues;
 350         unsigned int busy_sync_queues;
 351
 352         int rq_in_driver;
 353         int rq_in_flight[2];
 354
 355         /*
 356          * queue-depth detection
 357          */
 358         int rq_queued;
 359         int hw_tag;
 360         /*
 361          * hw_tag can be
 362          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 363          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 364          *  0 => no NCQ
 365          */
 366         int hw_tag_est_depth;
 367         unsigned int hw_tag_samples;
 368
 369         /*
 370          * idle window management
 371          */
 372         struct timer_list idle_slice_timer;
 373         struct work_struct unplug_work;
 374
 375         struct cfq_queue *active_queue;
 376         struct cfq_io_cq *active_cic;
 377
 378         sector_t last_position;
 379
 380         /*
 381          * tunables, see top of file
 382          */
 383         unsigned int cfq_quantum;
 384         unsigned int cfq_fifo_expire[2];
 385         unsigned int cfq_back_penalty;
 386         unsigned int cfq_back_max;
 387         unsigned int cfq_slice[2];
 388         unsigned int cfq_slice_async_rq;
 389         unsigned int cfq_slice_idle;
 390         unsigned int cfq_group_idle;
 391         unsigned int cfq_latency;
 392         unsigned int cfq_target_latency;
 393
 394         /*
 395          * Fallback dummy cfqq for extreme OOM conditions
 396          */
 397         struct cfq_queue oom_cfqq;
 398
 399         unsigned long last_delayed_sync;
 400 };
 401
 402 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 403 static void cfq_put_queue(struct cfq_queue *cfqq);
 404
 405 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 406                                             enum wl_class_t class,
 407                                             enum wl_type_t type)
 408 {
 409         if (!cfqg)
 410                 return NULL;
 411
 412         if (class == IDLE_WORKLOAD)
 413                 return &cfqg->service_tree_idle;
 414
 415         return &cfqg->service_trees[class][type];
 416 }
 417
 418 enum cfqq_state_flags {
 419         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 420         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 421         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 422         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 423         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 424         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 425         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 426         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 427         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 428         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 429         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 430         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 431         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 432 };
 433
 434 #define CFQ_CFQQ_FNS(name)                                              \
 435 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 436 {                                                                       \
 437         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 438 }                                                                       \
 439 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 440 {                                                                       \
 441         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 442 }                                                                       \
 443 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 444 {                                                                       \
 445         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 446 }
 447
 448 CFQ_CFQQ_FNS(on_rr);
 449 CFQ_CFQQ_FNS(wait_request);
 450 CFQ_CFQQ_FNS(must_dispatch);
 451 CFQ_CFQQ_FNS(must_alloc_slice);
 452 CFQ_CFQQ_FNS(fifo_expire);
 453 CFQ_CFQQ_FNS(idle_window);
 454 CFQ_CFQQ_FNS(prio_changed);
 455 CFQ_CFQQ_FNS(slice_new);
 456 CFQ_CFQQ_FNS(sync);
 457 CFQ_CFQQ_FNS(coop);
 458 CFQ_CFQQ_FNS(split_coop);
 459 CFQ_CFQQ_FNS(deep);
 460 CFQ_CFQQ_FNS(wait_busy);
 461 #undef CFQ_CFQQ_FNS
 462
 463 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 464
 465 /* cfqg stats flags */
 466 enum cfqg_stats_flags {
 467         CFQG_stats_waiting = 0,
 468         CFQG_stats_idling,
 469         CFQG_stats_empty,
 470 };
 471
 472 #define CFQG_FLAG_FNS(name)                                             \
 473 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 474 {                                                                       \
 475         stats->flags |= (1 << CFQG_stats_##name);                       \
 476 }                                                                       \
 477 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 478 {                                                                       \
 479         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 480 }                                                                       \
 481 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 482 {                                                                       \
 483         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 484 }                                                                       \
 485
 486 CFQG_FLAG_FNS(waiting)
 487 CFQG_FLAG_FNS(idling)
 488 CFQG_FLAG_FNS(empty)
 489 #undef CFQG_FLAG_FNS
 490
 491 /* This should be called with the queue_lock held. */
 492 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 493 {
 494         unsigned long long now;
 495
 496         if (!cfqg_stats_waiting(stats))
 497                 return;
 498
 499         now = sched_clock();
 500         if (time_after64(now, stats->start_group_wait_time))
 501                 blkg_stat_add(&stats->group_wait_time,
 502                               now - stats->start_group_wait_time);
 503         cfqg_stats_clear_waiting(stats);
 504 }
 505
 506 /* This should be called with the queue_lock held. */
 507 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 508                                                  struct cfq_group *curr_cfqg)
 509 {
 510         struct cfqg_stats *stats = &cfqg->stats;
 511
 512         if (cfqg_stats_waiting(stats))
 513                 return;
 514         if (cfqg == curr_cfqg)
 515                 return;
 516         stats->start_group_wait_time = sched_clock();
 517         cfqg_stats_mark_waiting(stats);
 518 }
 519
 520 /* This should be called with the queue_lock held. */
 521 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 522 {
 523         unsigned long long now;
 524
 525         if (!cfqg_stats_empty(stats))
 526                 return;
 527
 528         now = sched_clock();
 529         if (time_after64(now, stats->start_empty_time))
 530                 blkg_stat_add(&stats->empty_time,
 531                               now - stats->start_empty_time);
 532         cfqg_stats_clear_empty(stats);
 533 }
 534
 535 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 536 {
 537         blkg_stat_add(&cfqg->stats.dequeue, 1);
 538 }
 539
 540 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 541 {
 542         struct cfqg_stats *stats = &cfqg->stats;
 543
 544         if (blkg_rwstat_total(&stats->queued))
 545                 return;
 546
 547         /*
 548          * group is already marked empty. This can happen if cfqq got new
 549          * request in parent group and moved to this group while being added
 550          * to service tree. Just ignore the event and move on.
 551          */
 552         if (cfqg_stats_empty(stats))
 553                 return;
 554
 555         stats->start_empty_time = sched_clock();
 556         cfqg_stats_mark_empty(stats);
 557 }
 558
 559 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 560 {
 561         struct cfqg_stats *stats = &cfqg->stats;
 562
 563         if (cfqg_stats_idling(stats)) {
 564                 unsigned long long now = sched_clock();
 565
 566                 if (time_after64(now, stats->start_idle_time))
 567                         blkg_stat_add(&stats->idle_time,
 568                                       now - stats->start_idle_time);
 569                 cfqg_stats_clear_idling(stats);
 570         }
 571 }
 572
 573 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 574 {
 575         struct cfqg_stats *stats = &cfqg->stats;
 576
 577         BUG_ON(cfqg_stats_idling(stats));
 578
 579         stats->start_idle_time = sched_clock();
 580         cfqg_stats_mark_idling(stats);
 581 }
 582
 583 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 584 {
 585         struct cfqg_stats *stats = &cfqg->stats;
 586
 587         blkg_stat_add(&stats->avg_queue_size_sum,
 588                       blkg_rwstat_total(&stats->queued));
 589         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 590         cfqg_stats_update_group_wait_time(stats);
 591 }
 592
 593 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 594
 595 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 596 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 597 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 598 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 599 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 600 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 601 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 602
 603 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 604
 605 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 606
 607 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 608 {
 609         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 610 }
 611
 612 static struct cfq_group_data
 613 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 614 {
 615         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 616 }
 617
 618 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 619 {
 620         return pd_to_blkg(&cfqg->pd);
 621 }
 622
 623 static struct blkcg_policy blkcg_policy_cfq;
 624
 625 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 626 {
 627         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 628 }
 629
 630 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 631 {
 632         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 633 }
 634
 635 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 636 {
 637         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 638
 639         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 640 }
 641
 642 static inline void cfqg_get(struct cfq_group *cfqg)
 643 {
 644         return blkg_get(cfqg_to_blkg(cfqg));
 645 }
 646
 647 static inline void cfqg_put(struct cfq_group *cfqg)
 648 {
 649         return blkg_put(cfqg_to_blkg(cfqg));
 650 }
 651
 652 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 653         char __pbuf[128];                                               \
 654                                                                         \
 655         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 656         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 657                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 658                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 659                           __pbuf, ##args);                              \
 660 } while (0)
 661
 662 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 663         char __pbuf[128];                                               \
 664                                                                         \
 665         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 666         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 667 } while (0)
 668
 669 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 670                                             struct cfq_group *curr_cfqg, int rw)
 671 {
 672         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 673         cfqg_stats_end_empty_time(&cfqg->stats);
 674         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 675 }
 676
 677 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 678                         unsigned long time, unsigned long unaccounted_time)
 679 {
 680         blkg_stat_add(&cfqg->stats.time, time);
 681 #ifdef CONFIG_DEBUG_BLK_CGROUP
 682         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 683 #endif
 684 }
 685
 686 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 687 {
 688         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 689 }
 690
 691 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 692 {
 693         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 694 }
 695
 696 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 697                                               uint64_t bytes, int rw)
 698 {
 699         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
 700         blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
 701         blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
 702 }
 703
 704 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 705                         uint64_t start_time, uint64_t io_start_time, int rw)
 706 {
 707         struct cfqg_stats *stats = &cfqg->stats;
 708         unsigned long long now = sched_clock();
 709
 710         if (time_after64(now, io_start_time))
 711                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 712         if (time_after64(io_start_time, start_time))
 713                 blkg_rwstat_add(&stats->wait_time, rw,
 714                                 io_start_time - start_time);
 715 }
 716
 717 /* @stats = 0 */
 718 static void cfqg_stats_reset(struct cfqg_stats *stats)
 719 {
 720         /* queued stats shouldn't be cleared */
 721         blkg_rwstat_reset(&stats->service_bytes);
 722         blkg_rwstat_reset(&stats->serviced);
 723         blkg_rwstat_reset(&stats->merged);
 724         blkg_rwstat_reset(&stats->service_time);
 725         blkg_rwstat_reset(&stats->wait_time);
 726         blkg_stat_reset(&stats->time);
 727 #ifdef CONFIG_DEBUG_BLK_CGROUP
 728         blkg_stat_reset(&stats->unaccounted_time);
 729         blkg_stat_reset(&stats->avg_queue_size_sum);
 730         blkg_stat_reset(&stats->avg_queue_size_samples);
 731         blkg_stat_reset(&stats->dequeue);
 732         blkg_stat_reset(&stats->group_wait_time);
 733         blkg_stat_reset(&stats->idle_time);
 734         blkg_stat_reset(&stats->empty_time);
 735 #endif
 736 }
 737
 738 /* @to += @from */
 739 static void cfqg_stats_merge(struct cfqg_stats *to, struct cfqg_stats *from)
 740 {
 741         /* queued stats shouldn't be cleared */
 742         blkg_rwstat_merge(&to->service_bytes, &from->service_bytes);
 743         blkg_rwstat_merge(&to->serviced, &from->serviced);
 744         blkg_rwstat_merge(&to->merged, &from->merged);
 745         blkg_rwstat_merge(&to->service_time, &from->service_time);
 746         blkg_rwstat_merge(&to->wait_time, &from->wait_time);
 747         blkg_stat_merge(&from->time, &from->time);
 748 #ifdef CONFIG_DEBUG_BLK_CGROUP
 749         blkg_stat_merge(&to->unaccounted_time, &from->unaccounted_time);
 750         blkg_stat_merge(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 751         blkg_stat_merge(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 752         blkg_stat_merge(&to->dequeue, &from->dequeue);
 753         blkg_stat_merge(&to->group_wait_time, &from->group_wait_time);
 754         blkg_stat_merge(&to->idle_time, &from->idle_time);
 755         blkg_stat_merge(&to->empty_time, &from->empty_time);
 756 #endif
 757 }
 758
 759 /*
 760  * Transfer @cfqg's stats to its parent's dead_stats so that the ancestors'
 761  * recursive stats can still account for the amount used by this cfqg after
 762  * it's gone.
 763  */
 764 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 765 {
 766         struct cfq_group *parent = cfqg_parent(cfqg);
 767
 768         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 769
 770         if (unlikely(!parent))
 771                 return;
 772
 773         cfqg_stats_merge(&parent->dead_stats, &cfqg->stats);
 774         cfqg_stats_merge(&parent->dead_stats, &cfqg->dead_stats);
 775         cfqg_stats_reset(&cfqg->stats);
 776         cfqg_stats_reset(&cfqg->dead_stats);
 777 }
 778
 779 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 780
 781 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 782 static inline void cfqg_get(struct cfq_group *cfqg) { }
 783 static inline void cfqg_put(struct cfq_group *cfqg) { }
 784
 785 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 786         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 787                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 788                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 789                                 ##args)
 790 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 791
 792 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 793                         struct cfq_group *curr_cfqg, int rw) { }
 794 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 795                         unsigned long time, unsigned long unaccounted_time) { }
 796 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 797 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 798 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 799                                               uint64_t bytes, int rw) { }
 800 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 801                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 802
 803 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 804
 805 #define cfq_log(cfqd, fmt, args...)     \
 806         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 807
 808 /* Traverses through cfq group service trees */
 809 #define for_each_cfqg_st(cfqg, i, j, st) \
 810         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 811                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 812                         : &cfqg->service_tree_idle; \
 813                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 814                         (i == IDLE_WORKLOAD && j == 0); \
 815                         j++, st = i < IDLE_WORKLOAD ? \
 816                         &cfqg->service_trees[i][j]: NULL) \
 817
 818 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 819         struct cfq_ttime *ttime, bool group_idle)
 820 {
 821         unsigned long slice;
 822         if (!sample_valid(ttime->ttime_samples))
 823                 return false;
 824         if (group_idle)
 825                 slice = cfqd->cfq_group_idle;
 826         else
 827                 slice = cfqd->cfq_slice_idle;
 828         return ttime->ttime_mean > slice;
 829 }
 830
 831 static inline bool iops_mode(struct cfq_data *cfqd)
 832 {
 833         /*
 834          * If we are not idling on queues and it is a NCQ drive, parallel
 835          * execution of requests is on and measuring time is not possible
 836          * in most of the cases until and unless we drive shallower queue
 837          * depths and that becomes a performance bottleneck. In such cases
 838          * switch to start providing fairness in terms of number of IOs.
 839          */
 840         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 841                 return true;
 842         else
 843                 return false;
 844 }
 845
 846 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 847 {
 848         if (cfq_class_idle(cfqq))
 849                 return IDLE_WORKLOAD;
 850         if (cfq_class_rt(cfqq))
 851                 return RT_WORKLOAD;
 852         return BE_WORKLOAD;
 853 }
 854
 855
 856 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 857 {
 858         if (!cfq_cfqq_sync(cfqq))
 859                 return ASYNC_WORKLOAD;
 860         if (!cfq_cfqq_idle_window(cfqq))
 861                 return SYNC_NOIDLE_WORKLOAD;
 862         return SYNC_WORKLOAD;
 863 }
 864
 865 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 866                                         struct cfq_data *cfqd,
 867                                         struct cfq_group *cfqg)
 868 {
 869         if (wl_class == IDLE_WORKLOAD)
 870                 return cfqg->service_tree_idle.count;
 871
 872         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 873                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 874                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 875 }
 876
 877 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 878                                         struct cfq_group *cfqg)
 879 {
 880         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 881                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 882 }
 883
 884 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 885 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 886                                        struct cfq_io_cq *cic, struct bio *bio);
 887
 888 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 889 {
 890         /* cic->icq is the first member, %NULL will convert to %NULL */
 891         return container_of(icq, struct cfq_io_cq, icq);
 892 }
 893
 894 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 895                                                struct io_context *ioc)
 896 {
 897         if (ioc)
 898                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 899         return NULL;
 900 }
 901
 902 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 903 {
 904         return cic->cfqq[is_sync];
 905 }
 906
 907 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 908                                 bool is_sync)
 909 {
 910         cic->cfqq[is_sync] = cfqq;
 911 }
 912
 913 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 914 {
 915         return cic->icq.q->elevator->elevator_data;
 916 }
 917
 918 /*
 919  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 920  * set (in which case it could also be direct WRITE).
 921  */
 922 static inline bool cfq_bio_sync(struct bio *bio)
 923 {
 924         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 925 }
 926
 927 /*
 928  * scheduler run of queue, if there are requests pending and no one in the
 929  * driver that will restart queueing
 930  */
 931 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 932 {
 933         if (cfqd->busy_queues) {
 934                 cfq_log(cfqd, "schedule dispatch");
 935                 kblockd_schedule_work(&cfqd->unplug_work);
 936         }
 937 }
 938
 939 /*
 940  * Scale schedule slice based on io priority. Use the sync time slice only
 941  * if a queue is marked sync and has sync io queued. A sync queue with async
 942  * io only, should not get full sync slice length.
 943  */
 944 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 945                                  unsigned short prio)
 946 {
 947         const int base_slice = cfqd->cfq_slice[sync];
 948
 949         WARN_ON(prio >= IOPRIO_BE_NR);
 950
 951         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 952 }
 953
 954 static inline int
 955 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 956 {
 957         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 958 }
 959
 960 /**
 961  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 962  * @charge: disk time being charged
 963  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 964  *
 965  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 966  * scaling is inversely proportional.
 967  *
 968  * scaled = charge / vfraction
 969  *
 970  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 971  */
 972 static inline u64 cfqg_scale_charge(unsigned long charge,
 973                                     unsigned int vfraction)
 974 {
 975         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 976
 977         /* charge / vfraction */
 978         c <<= CFQ_SERVICE_SHIFT;
 979         do_div(c, vfraction);
 980         return c;
 981 }
 982
 983 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 984 {
 985         s64 delta = (s64)(vdisktime - min_vdisktime);
 986         if (delta > 0)
 987                 min_vdisktime = vdisktime;
 988
 989         return min_vdisktime;
 990 }
 991
 992 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 993 {
 994         s64 delta = (s64)(vdisktime - min_vdisktime);
 995         if (delta < 0)
 996                 min_vdisktime = vdisktime;
 997
 998         return min_vdisktime;
 999 }
1000
1001 static void update_min_vdisktime(struct cfq_rb_root *st)
1002 {
1003         struct cfq_group *cfqg;
1004
1005         if (st->left) {
1006                 cfqg = rb_entry_cfqg(st->left);
1007                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1008                                                   cfqg->vdisktime);
1009         }
1010 }
1011
1012 /*
1013  * get averaged number of queues of RT/BE priority.
1014  * average is updated, with a formula that gives more weight to higher numbers,
1015  * to quickly follows sudden increases and decrease slowly
1016  */
1017
1018 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1019                                         struct cfq_group *cfqg, bool rt)
1020 {
1021         unsigned min_q, max_q;
1022         unsigned mult  = cfq_hist_divisor - 1;
1023         unsigned round = cfq_hist_divisor / 2;
1024         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1025
1026         min_q = min(cfqg->busy_queues_avg[rt], busy);
1027         max_q = max(cfqg->busy_queues_avg[rt], busy);
1028         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1029                 cfq_hist_divisor;
1030         return cfqg->busy_queues_avg[rt];
1031 }
1032
1033 static inline unsigned
1034 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1035 {
1036         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1037 }
1038
1039 static inline unsigned
1040 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1041 {
1042         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1043         if (cfqd->cfq_latency) {
1044                 /*
1045                  * interested queues (we consider only the ones with the same
1046                  * priority class in the cfq group)
1047                  */
1048                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1049                                                 cfq_class_rt(cfqq));
1050                 unsigned sync_slice = cfqd->cfq_slice[1];
1051                 unsigned expect_latency = sync_slice * iq;
1052                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1053
1054                 if (expect_latency > group_slice) {
1055                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1056                         /* scale low_slice according to IO priority
1057                          * and sync vs async */
1058                         unsigned low_slice =
1059                                 min(slice, base_low_slice * slice / sync_slice);
1060                         /* the adapted slice value is scaled to fit all iqs
1061                          * into the target latency */
1062                         slice = max(slice * group_slice / expect_latency,
1063                                     low_slice);
1064                 }
1065         }
1066         return slice;
1067 }
1068
1069 static inline void
1070 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1071 {
1072         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1073
1074         cfqq->slice_start = jiffies;
1075         cfqq->slice_end = jiffies + slice;
1076         cfqq->allocated_slice = slice;
1077         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1078 }
1079
1080 /*
1081  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1082  * isn't valid until the first request from the dispatch is activated
1083  * and the slice time set.
1084  */
1085 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1086 {
1087         if (cfq_cfqq_slice_new(cfqq))
1088                 return false;
1089         if (time_before(jiffies, cfqq->slice_end))
1090                 return false;
1091
1092         return true;
1093 }
1094
1095 /*
1096  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1097  * We choose the request that is closest to the head right now. Distance
1098  * behind the head is penalized and only allowed to a certain extent.
1099  */
1100 static struct request *
1101 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1102 {
1103         sector_t s1, s2, d1 = 0, d2 = 0;
1104         unsigned long back_max;
1105 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1106 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1107         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1108
1109         if (rq1 == NULL || rq1 == rq2)
1110                 return rq2;
1111         if (rq2 == NULL)
1112                 return rq1;
1113
1114         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1115                 return rq_is_sync(rq1) ? rq1 : rq2;
1116
1117         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1118                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1119
1120         s1 = blk_rq_pos(rq1);
1121         s2 = blk_rq_pos(rq2);
1122
1123         /*
1124          * by definition, 1KiB is 2 sectors
1125          */
1126         back_max = cfqd->cfq_back_max * 2;
1127
1128         /*
1129          * Strict one way elevator _except_ in the case where we allow
1130          * short backward seeks which are biased as twice the cost of a
1131          * similar forward seek.
1132          */
1133         if (s1 >= last)
1134                 d1 = s1 - last;
1135         else if (s1 + back_max >= last)
1136                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1137         else
1138                 wrap |= CFQ_RQ1_WRAP;
1139
1140         if (s2 >= last)
1141                 d2 = s2 - last;
1142         else if (s2 + back_max >= last)
1143                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1144         else
1145                 wrap |= CFQ_RQ2_WRAP;
1146
1147         /* Found required data */
1148
1149         /*
1150          * By doing switch() on the bit mask "wrap" we avoid having to
1151          * check two variables for all permutations: --> faster!
1152          */
1153         switch (wrap) {
1154         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1155                 if (d1 < d2)
1156                         return rq1;
1157                 else if (d2 < d1)
1158                         return rq2;
1159                 else {
1160                         if (s1 >= s2)
1161                                 return rq1;
1162                         else
1163                                 return rq2;
1164                 }
1165
1166         case CFQ_RQ2_WRAP:
1167                 return rq1;
1168         case CFQ_RQ1_WRAP:
1169                 return rq2;
1170         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1171         default:
1172                 /*
1173                  * Since both rqs are wrapped,
1174                  * start with the one that's further behind head
1175                  * (--> only *one* back seek required),
1176                  * since back seek takes more time than forward.
1177                  */
1178                 if (s1 <= s2)
1179                         return rq1;
1180                 else
1181                         return rq2;
1182         }
1183 }
1184
1185 /*
1186  * The below is leftmost cache rbtree addon
1187  */
1188 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1189 {
1190         /* Service tree is empty */
1191         if (!root->count)
1192                 return NULL;
1193
1194         if (!root->left)
1195                 root->left = rb_first(&root->rb);
1196
1197         if (root->left)
1198                 return rb_entry(root->left, struct cfq_queue, rb_node);
1199
1200         return NULL;
1201 }
1202
1203 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1204 {
1205         if (!root->left)
1206                 root->left = rb_first(&root->rb);
1207
1208         if (root->left)
1209                 return rb_entry_cfqg(root->left);
1210
1211         return NULL;
1212 }
1213
1214 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1215 {
1216         rb_erase(n, root);
1217         RB_CLEAR_NODE(n);
1218 }
1219
1220 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1221 {
1222         if (root->left == n)
1223                 root->left = NULL;
1224         rb_erase_init(n, &root->rb);
1225         --root->count;
1226 }
1227
1228 /*
1229  * would be nice to take fifo expire time into account as well
1230  */
1231 static struct request *
1232 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1233                   struct request *last)
1234 {
1235         struct rb_node *rbnext = rb_next(&last->rb_node);
1236         struct rb_node *rbprev = rb_prev(&last->rb_node);
1237         struct request *next = NULL, *prev = NULL;
1238
1239         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1240
1241         if (rbprev)
1242                 prev = rb_entry_rq(rbprev);
1243
1244         if (rbnext)
1245                 next = rb_entry_rq(rbnext);
1246         else {
1247                 rbnext = rb_first(&cfqq->sort_list);
1248                 if (rbnext && rbnext != &last->rb_node)
1249                         next = rb_entry_rq(rbnext);
1250         }
1251
1252         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1253 }
1254
1255 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1256                                       struct cfq_queue *cfqq)
1257 {
1258         /*
1259          * just an approximation, should be ok.
1260          */
1261         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1262                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1263 }
1264
1265 static inline s64
1266 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1267 {
1268         return cfqg->vdisktime - st->min_vdisktime;
1269 }
1270
1271 static void
1272 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1273 {
1274         struct rb_node **node = &st->rb.rb_node;
1275         struct rb_node *parent = NULL;
1276         struct cfq_group *__cfqg;
1277         s64 key = cfqg_key(st, cfqg);
1278         int left = 1;
1279
1280         while (*node != NULL) {
1281                 parent = *node;
1282                 __cfqg = rb_entry_cfqg(parent);
1283
1284                 if (key < cfqg_key(st, __cfqg))
1285                         node = &parent->rb_left;
1286                 else {
1287                         node = &parent->rb_right;
1288                         left = 0;
1289                 }
1290         }
1291
1292         if (left)
1293                 st->left = &cfqg->rb_node;
1294
1295         rb_link_node(&cfqg->rb_node, parent, node);
1296         rb_insert_color(&cfqg->rb_node, &st->rb);
1297 }
1298
1299 /*
1300  * This has to be called only on activation of cfqg
1301  */
1302 static void
1303 cfq_update_group_weight(struct cfq_group *cfqg)
1304 {
1305         if (cfqg->new_weight) {
1306                 cfqg->weight = cfqg->new_weight;
1307                 cfqg->new_weight = 0;
1308         }
1309 }
1310
1311 static void
1312 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1313 {
1314         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1315
1316         if (cfqg->new_leaf_weight) {
1317                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1318                 cfqg->new_leaf_weight = 0;
1319         }
1320 }
1321
1322 static void
1323 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1324 {
1325         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1326         struct cfq_group *pos = cfqg;
1327         struct cfq_group *parent;
1328         bool propagate;
1329
1330         /* add to the service tree */
1331         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1332
1333         /*
1334          * Update leaf_weight.  We cannot update weight at this point
1335          * because cfqg might already have been activated and is
1336          * contributing its current weight to the parent's child_weight.
1337          */
1338         cfq_update_group_leaf_weight(cfqg);
1339         __cfq_group_service_tree_add(st, cfqg);
1340
1341         /*
1342          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1343          * entitled to.  vfraction is calculated by walking the tree
1344          * towards the root calculating the fraction it has at each level.
1345          * The compounded ratio is how much vfraction @cfqg owns.
1346          *
1347          * Start with the proportion tasks in this cfqg has against active
1348          * children cfqgs - its leaf_weight against children_weight.
1349          */
1350         propagate = !pos->nr_active++;
1351         pos->children_weight += pos->leaf_weight;
1352         vfr = vfr * pos->leaf_weight / pos->children_weight;
1353
1354         /*
1355          * Compound ->weight walking up the tree.  Both activation and
1356          * vfraction calculation are done in the same loop.  Propagation
1357          * stops once an already activated node is met.  vfraction
1358          * calculation should always continue to the root.
1359          */
1360         while ((parent = cfqg_parent(pos))) {
1361                 if (propagate) {
1362                         cfq_update_group_weight(pos);
1363                         propagate = !parent->nr_active++;
1364                         parent->children_weight += pos->weight;
1365                 }
1366                 vfr = vfr * pos->weight / parent->children_weight;
1367                 pos = parent;
1368         }
1369
1370         cfqg->vfraction = max_t(unsigned, vfr, 1);
1371 }
1372
1373 static void
1374 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1375 {
1376         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1377         struct cfq_group *__cfqg;
1378         struct rb_node *n;
1379
1380         cfqg->nr_cfqq++;
1381         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1382                 return;
1383
1384         /*
1385          * Currently put the group at the end. Later implement something
1386          * so that groups get lesser vtime based on their weights, so that
1387          * if group does not loose all if it was not continuously backlogged.
1388          */
1389         n = rb_last(&st->rb);
1390         if (n) {
1391                 __cfqg = rb_entry_cfqg(n);
1392                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1393         } else
1394                 cfqg->vdisktime = st->min_vdisktime;
1395         cfq_group_service_tree_add(st, cfqg);
1396 }
1397
1398 static void
1399 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1400 {
1401         struct cfq_group *pos = cfqg;
1402         bool propagate;
1403
1404         /*
1405          * Undo activation from cfq_group_service_tree_add().  Deactivate
1406          * @cfqg and propagate deactivation upwards.
1407          */
1408         propagate = !--pos->nr_active;
1409         pos->children_weight -= pos->leaf_weight;
1410
1411         while (propagate) {
1412                 struct cfq_group *parent = cfqg_parent(pos);
1413
1414                 /* @pos has 0 nr_active at this point */
1415                 WARN_ON_ONCE(pos->children_weight);
1416                 pos->vfraction = 0;
1417
1418                 if (!parent)
1419                         break;
1420
1421                 propagate = !--parent->nr_active;
1422                 parent->children_weight -= pos->weight;
1423                 pos = parent;
1424         }
1425
1426         /* remove from the service tree */
1427         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1428                 cfq_rb_erase(&cfqg->rb_node, st);
1429 }
1430
1431 static void
1432 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1433 {
1434         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1435
1436         BUG_ON(cfqg->nr_cfqq < 1);
1437         cfqg->nr_cfqq--;
1438
1439         /* If there are other cfq queues under this group, don't delete it */
1440         if (cfqg->nr_cfqq)
1441                 return;
1442
1443         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1444         cfq_group_service_tree_del(st, cfqg);
1445         cfqg->saved_wl_slice = 0;
1446         cfqg_stats_update_dequeue(cfqg);
1447 }
1448
1449 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1450                                                 unsigned int *unaccounted_time)
1451 {
1452         unsigned int slice_used;
1453
1454         /*
1455          * Queue got expired before even a single request completed or
1456          * got expired immediately after first request completion.
1457          */
1458         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1459                 /*
1460                  * Also charge the seek time incurred to the group, otherwise
1461                  * if there are mutiple queues in the group, each can dispatch
1462                  * a single request on seeky media and cause lots of seek time
1463                  * and group will never know it.
1464                  */
1465                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1466                                         1);
1467         } else {
1468                 slice_used = jiffies - cfqq->slice_start;
1469                 if (slice_used > cfqq->allocated_slice) {
1470                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1471                         slice_used = cfqq->allocated_slice;
1472                 }
1473                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1474                         *unaccounted_time += cfqq->slice_start -
1475                                         cfqq->dispatch_start;
1476         }
1477
1478         return slice_used;
1479 }
1480
1481 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1482                                 struct cfq_queue *cfqq)
1483 {
1484         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1485         unsigned int used_sl, charge, unaccounted_sl = 0;
1486         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1487                         - cfqg->service_tree_idle.count;
1488         unsigned int vfr;
1489
1490         BUG_ON(nr_sync < 0);
1491         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1492
1493         if (iops_mode(cfqd))
1494                 charge = cfqq->slice_dispatch;
1495         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1496                 charge = cfqq->allocated_slice;
1497
1498         /*
1499          * Can't update vdisktime while on service tree and cfqg->vfraction
1500          * is valid only while on it.  Cache vfr, leave the service tree,
1501          * update vdisktime and go back on.  The re-addition to the tree
1502          * will also update the weights as necessary.
1503          */
1504         vfr = cfqg->vfraction;
1505         cfq_group_service_tree_del(st, cfqg);
1506         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1507         cfq_group_service_tree_add(st, cfqg);
1508
1509         /* This group is being expired. Save the context */
1510         if (time_after(cfqd->workload_expires, jiffies)) {
1511                 cfqg->saved_wl_slice = cfqd->workload_expires
1512                                                 - jiffies;
1513                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1514                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1515         } else
1516                 cfqg->saved_wl_slice = 0;
1517
1518         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1519                                         st->min_vdisktime);
1520         cfq_log_cfqq(cfqq->cfqd, cfqq,
1521                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1522                      used_sl, cfqq->slice_dispatch, charge,
1523                      iops_mode(cfqd), cfqq->nr_sectors);
1524         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1525         cfqg_stats_set_start_empty_time(cfqg);
1526 }
1527
1528 /**
1529  * cfq_init_cfqg_base - initialize base part of a cfq_group
1530  * @cfqg: cfq_group to initialize
1531  *
1532  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1533  * is enabled or not.
1534  */
1535 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1536 {
1537         struct cfq_rb_root *st;
1538         int i, j;
1539
1540         for_each_cfqg_st(cfqg, i, j, st)
1541                 *st = CFQ_RB_ROOT;
1542         RB_CLEAR_NODE(&cfqg->rb_node);
1543
1544         cfqg->ttime.last_end_request = jiffies;
1545 }
1546
1547 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1548 static void cfqg_stats_init(struct cfqg_stats *stats)
1549 {
1550         blkg_rwstat_init(&stats->service_bytes);
1551         blkg_rwstat_init(&stats->serviced);
1552         blkg_rwstat_init(&stats->merged);
1553         blkg_rwstat_init(&stats->service_time);
1554         blkg_rwstat_init(&stats->wait_time);
1555         blkg_rwstat_init(&stats->queued);
1556
1557         blkg_stat_init(&stats->sectors);
1558         blkg_stat_init(&stats->time);
1559
1560 #ifdef CONFIG_DEBUG_BLK_CGROUP
1561         blkg_stat_init(&stats->unaccounted_time);
1562         blkg_stat_init(&stats->avg_queue_size_sum);
1563         blkg_stat_init(&stats->avg_queue_size_samples);
1564         blkg_stat_init(&stats->dequeue);
1565         blkg_stat_init(&stats->group_wait_time);
1566         blkg_stat_init(&stats->idle_time);
1567         blkg_stat_init(&stats->empty_time);
1568 #endif
1569 }
1570
1571 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1572 {
1573         struct cfq_group_data *cgd;
1574
1575         cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1576         if (!cgd)
1577                 return NULL;
1578         return &cgd->cpd;
1579 }
1580
1581 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1582 {
1583         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1584
1585         if (cpd_to_blkcg(cpd) == &blkcg_root) {
1586                 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1587                 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1588         } else {
1589                 cgd->weight = CFQ_WEIGHT_DEFAULT;
1590                 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1591         }
1592 }
1593
1594 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1595 {
1596         kfree(cpd_to_cfqgd(cpd));
1597 }
1598
1599 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1600 {
1601         struct cfq_group *cfqg;
1602
1603         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1604         if (!cfqg)
1605                 return NULL;
1606
1607         cfq_init_cfqg_base(cfqg);
1608         cfqg_stats_init(&cfqg->stats);
1609         cfqg_stats_init(&cfqg->dead_stats);
1610
1611         return &cfqg->pd;
1612 }
1613
1614 static void cfq_pd_init(struct blkg_policy_data *pd)
1615 {
1616         struct cfq_group *cfqg = pd_to_cfqg(pd);
1617         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1618
1619         cfqg->weight = cgd->weight;
1620         cfqg->leaf_weight = cgd->leaf_weight;
1621 }
1622
1623 static void cfq_pd_offline(struct blkg_policy_data *pd)
1624 {
1625         struct cfq_group *cfqg = pd_to_cfqg(pd);
1626         int i;
1627
1628         for (i = 0; i < IOPRIO_BE_NR; i++) {
1629                 if (cfqg->async_cfqq[0][i])
1630                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1631                 if (cfqg->async_cfqq[1][i])
1632                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1633         }
1634
1635         if (cfqg->async_idle_cfqq)
1636                 cfq_put_queue(cfqg->async_idle_cfqq);
1637
1638         /*
1639          * @blkg is going offline and will be ignored by
1640          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1641          * that they don't get lost.  If IOs complete after this point, the
1642          * stats for them will be lost.  Oh well...
1643          */
1644         cfqg_stats_xfer_dead(cfqg);
1645 }
1646
1647 static void cfq_pd_free(struct blkg_policy_data *pd)
1648 {
1649         return kfree(pd);
1650 }
1651
1652 /* offset delta from cfqg->stats to cfqg->dead_stats */
1653 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1654                                         offsetof(struct cfq_group, stats);
1655
1656 /* to be used by recursive prfill, sums live and dead stats recursively */
1657 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1658 {
1659         u64 sum = 0;
1660
1661         sum += blkg_stat_recursive_sum(pd, off);
1662         sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1663         return sum;
1664 }
1665
1666 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1667 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1668                                                        int off)
1669 {
1670         struct blkg_rwstat a, b;
1671
1672         a = blkg_rwstat_recursive_sum(pd, off);
1673         b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1674         blkg_rwstat_merge(&a, &b);
1675         return a;
1676 }
1677
1678 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1679 {
1680         struct cfq_group *cfqg = pd_to_cfqg(pd);
1681
1682         cfqg_stats_reset(&cfqg->stats);
1683         cfqg_stats_reset(&cfqg->dead_stats);
1684 }
1685
1686 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1687                                          struct blkcg *blkcg)
1688 {
1689         struct blkcg_gq *blkg;
1690
1691         blkg = blkg_lookup(blkcg, cfqd->queue);
1692         if (likely(blkg))
1693                 return blkg_to_cfqg(blkg);
1694         return NULL;
1695 }
1696
1697 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1698 {
1699         cfqq->cfqg = cfqg;
1700         /* cfqq reference on cfqg */
1701         cfqg_get(cfqg);
1702 }
1703
1704 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1705                                      struct blkg_policy_data *pd, int off)
1706 {
1707         struct cfq_group *cfqg = pd_to_cfqg(pd);
1708
1709         if (!cfqg->dev_weight)
1710                 return 0;
1711         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1712 }
1713
1714 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1715 {
1716         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1717                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1718                           0, false);
1719         return 0;
1720 }
1721
1722 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1723                                           struct blkg_policy_data *pd, int off)
1724 {
1725         struct cfq_group *cfqg = pd_to_cfqg(pd);
1726
1727         if (!cfqg->dev_leaf_weight)
1728                 return 0;
1729         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1730 }
1731
1732 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1733 {
1734         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1735                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1736                           0, false);
1737         return 0;
1738 }
1739
1740 static int cfq_print_weight(struct seq_file *sf, void *v)
1741 {
1742         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1743         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1744         unsigned int val = 0;
1745
1746         if (cgd)
1747                 val = cgd->weight;
1748
1749         seq_printf(sf, "%u\n", val);
1750         return 0;
1751 }
1752
1753 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1754 {
1755         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1756         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1757         unsigned int val = 0;
1758
1759         if (cgd)
1760                 val = cgd->leaf_weight;
1761
1762         seq_printf(sf, "%u\n", val);
1763         return 0;
1764 }
1765
1766 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1767                                         char *buf, size_t nbytes, loff_t off,
1768                                         bool is_leaf_weight)
1769 {
1770         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1771         struct blkg_conf_ctx ctx;
1772         struct cfq_group *cfqg;
1773         struct cfq_group_data *cfqgd;
1774         int ret;
1775
1776         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1777         if (ret)
1778                 return ret;
1779
1780         ret = -EINVAL;
1781         cfqg = blkg_to_cfqg(ctx.blkg);
1782         cfqgd = blkcg_to_cfqgd(blkcg);
1783         if (!cfqg || !cfqgd)
1784                 goto err;
1785
1786         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1787                 if (!is_leaf_weight) {
1788                         cfqg->dev_weight = ctx.v;
1789                         cfqg->new_weight = ctx.v ?: cfqgd->weight;
1790                 } else {
1791                         cfqg->dev_leaf_weight = ctx.v;
1792                         cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1793                 }
1794                 ret = 0;
1795         }
1796
1797 err:
1798         blkg_conf_finish(&ctx);
1799         return ret ?: nbytes;
1800 }
1801
1802 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1803                                       char *buf, size_t nbytes, loff_t off)
1804 {
1805         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1806 }
1807
1808 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1809                                            char *buf, size_t nbytes, loff_t off)
1810 {
1811         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1812 }
1813
1814 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1815                             u64 val, bool is_leaf_weight)
1816 {
1817         struct blkcg *blkcg = css_to_blkcg(css);
1818         struct blkcg_gq *blkg;
1819         struct cfq_group_data *cfqgd;
1820         int ret = 0;
1821
1822         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1823                 return -EINVAL;
1824
1825         spin_lock_irq(&blkcg->lock);
1826         cfqgd = blkcg_to_cfqgd(blkcg);
1827         if (!cfqgd) {
1828                 ret = -EINVAL;
1829                 goto out;
1830         }
1831
1832         if (!is_leaf_weight)
1833                 cfqgd->weight = val;
1834         else
1835                 cfqgd->leaf_weight = val;
1836
1837         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1838                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1839
1840                 if (!cfqg)
1841                         continue;
1842
1843                 if (!is_leaf_weight) {
1844                         if (!cfqg->dev_weight)
1845                                 cfqg->new_weight = cfqgd->weight;
1846                 } else {
1847                         if (!cfqg->dev_leaf_weight)
1848                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1849                 }
1850         }
1851
1852 out:
1853         spin_unlock_irq(&blkcg->lock);
1854         return ret;
1855 }
1856
1857 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1858                           u64 val)
1859 {
1860         return __cfq_set_weight(css, cft, val, false);
1861 }
1862
1863 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1864                                struct cftype *cft, u64 val)
1865 {
1866         return __cfq_set_weight(css, cft, val, true);
1867 }
1868
1869 static int cfqg_print_stat(struct seq_file *sf, void *v)
1870 {
1871         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1872                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1873         return 0;
1874 }
1875
1876 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1877 {
1878         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1879                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1880         return 0;
1881 }
1882
1883 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1884                                       struct blkg_policy_data *pd, int off)
1885 {
1886         u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1887
1888         return __blkg_prfill_u64(sf, pd, sum);
1889 }
1890
1891 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1892                                         struct blkg_policy_data *pd, int off)
1893 {
1894         struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1895
1896         return __blkg_prfill_rwstat(sf, pd, &sum);
1897 }
1898
1899 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1900 {
1901         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1902                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1903                           seq_cft(sf)->private, false);
1904         return 0;
1905 }
1906
1907 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1908 {
1909         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1910                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1911                           seq_cft(sf)->private, true);
1912         return 0;
1913 }
1914
1915 #ifdef CONFIG_DEBUG_BLK_CGROUP
1916 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1917                                       struct blkg_policy_data *pd, int off)
1918 {
1919         struct cfq_group *cfqg = pd_to_cfqg(pd);
1920         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1921         u64 v = 0;
1922
1923         if (samples) {
1924                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1925                 v = div64_u64(v, samples);
1926         }
1927         __blkg_prfill_u64(sf, pd, v);
1928         return 0;
1929 }
1930
1931 /* print avg_queue_size */
1932 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1933 {
1934         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1935                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1936                           0, false);
1937         return 0;
1938 }
1939 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1940
1941 static struct cftype cfq_blkcg_files[] = {
1942         /* on root, weight is mapped to leaf_weight */
1943         {
1944                 .name = "weight_device",
1945                 .flags = CFTYPE_ONLY_ON_ROOT,
1946                 .seq_show = cfqg_print_leaf_weight_device,
1947                 .write = cfqg_set_leaf_weight_device,
1948         },
1949         {
1950                 .name = "weight",
1951                 .flags = CFTYPE_ONLY_ON_ROOT,
1952                 .seq_show = cfq_print_leaf_weight,
1953                 .write_u64 = cfq_set_leaf_weight,
1954         },
1955
1956         /* no such mapping necessary for !roots */
1957         {
1958                 .name = "weight_device",
1959                 .flags = CFTYPE_NOT_ON_ROOT,
1960                 .seq_show = cfqg_print_weight_device,
1961                 .write = cfqg_set_weight_device,
1962         },
1963         {
1964                 .name = "weight",
1965                 .flags = CFTYPE_NOT_ON_ROOT,
1966                 .seq_show = cfq_print_weight,
1967                 .write_u64 = cfq_set_weight,
1968         },
1969
1970         {
1971                 .name = "leaf_weight_device",
1972                 .seq_show = cfqg_print_leaf_weight_device,
1973                 .write = cfqg_set_leaf_weight_device,
1974         },
1975         {
1976                 .name = "leaf_weight",
1977                 .seq_show = cfq_print_leaf_weight,
1978                 .write_u64 = cfq_set_leaf_weight,
1979         },
1980
1981         /* statistics, covers only the tasks in the cfqg */
1982         {
1983                 .name = "time",
1984                 .private = offsetof(struct cfq_group, stats.time),
1985                 .seq_show = cfqg_print_stat,
1986         },
1987         {
1988                 .name = "sectors",
1989                 .private = offsetof(struct cfq_group, stats.sectors),
1990                 .seq_show = cfqg_print_stat,
1991         },
1992         {
1993                 .name = "io_service_bytes",
1994                 .private = offsetof(struct cfq_group, stats.service_bytes),
1995                 .seq_show = cfqg_print_rwstat,
1996         },
1997         {
1998                 .name = "io_serviced",
1999                 .private = offsetof(struct cfq_group, stats.serviced),
2000                 .seq_show = cfqg_print_rwstat,
2001         },
2002         {
2003                 .name = "io_service_time",
2004                 .private = offsetof(struct cfq_group, stats.service_time),
2005                 .seq_show = cfqg_print_rwstat,
2006         },
2007         {
2008                 .name = "io_wait_time",
2009                 .private = offsetof(struct cfq_group, stats.wait_time),
2010                 .seq_show = cfqg_print_rwstat,
2011         },
2012         {
2013                 .name = "io_merged",
2014                 .private = offsetof(struct cfq_group, stats.merged),
2015                 .seq_show = cfqg_print_rwstat,
2016         },
2017         {
2018                 .name = "io_queued",
2019                 .private = offsetof(struct cfq_group, stats.queued),
2020                 .seq_show = cfqg_print_rwstat,
2021         },
2022
2023         /* the same statictics which cover the cfqg and its descendants */
2024         {
2025                 .name = "time_recursive",
2026                 .private = offsetof(struct cfq_group, stats.time),
2027                 .seq_show = cfqg_print_stat_recursive,
2028         },
2029         {
2030                 .name = "sectors_recursive",
2031                 .private = offsetof(struct cfq_group, stats.sectors),
2032                 .seq_show = cfqg_print_stat_recursive,
2033         },
2034         {
2035                 .name = "io_service_bytes_recursive",
2036                 .private = offsetof(struct cfq_group, stats.service_bytes),
2037                 .seq_show = cfqg_print_rwstat_recursive,
2038         },
2039         {
2040                 .name = "io_serviced_recursive",
2041                 .private = offsetof(struct cfq_group, stats.serviced),
2042                 .seq_show = cfqg_print_rwstat_recursive,
2043         },
2044         {
2045                 .name = "io_service_time_recursive",
2046                 .private = offsetof(struct cfq_group, stats.service_time),
2047                 .seq_show = cfqg_print_rwstat_recursive,
2048         },
2049         {
2050                 .name = "io_wait_time_recursive",
2051                 .private = offsetof(struct cfq_group, stats.wait_time),
2052                 .seq_show = cfqg_print_rwstat_recursive,
2053         },
2054         {
2055                 .name = "io_merged_recursive",
2056                 .private = offsetof(struct cfq_group, stats.merged),
2057                 .seq_show = cfqg_print_rwstat_recursive,
2058         },
2059         {
2060                 .name = "io_queued_recursive",
2061                 .private = offsetof(struct cfq_group, stats.queued),
2062                 .seq_show = cfqg_print_rwstat_recursive,
2063         },
2064 #ifdef CONFIG_DEBUG_BLK_CGROUP
2065         {
2066                 .name = "avg_queue_size",
2067                 .seq_show = cfqg_print_avg_queue_size,
2068         },
2069         {
2070                 .name = "group_wait_time",
2071                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2072                 .seq_show = cfqg_print_stat,
2073         },
2074         {
2075                 .name = "idle_time",
2076                 .private = offsetof(struct cfq_group, stats.idle_time),
2077                 .seq_show = cfqg_print_stat,
2078         },
2079         {
2080                 .name = "empty_time",
2081                 .private = offsetof(struct cfq_group, stats.empty_time),
2082                 .seq_show = cfqg_print_stat,
2083         },
2084         {
2085                 .name = "dequeue",
2086                 .private = offsetof(struct cfq_group, stats.dequeue),
2087                 .seq_show = cfqg_print_stat,
2088         },
2089         {
2090                 .name = "unaccounted_time",
2091                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2092                 .seq_show = cfqg_print_stat,
2093         },
2094 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2095         { }     /* terminate */
2096 };
2097 #else /* GROUP_IOSCHED */
2098 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2099                                          struct blkcg *blkcg)
2100 {
2101         return cfqd->root_group;
2102 }
2103
2104 static inline void
2105 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2106         cfqq->cfqg = cfqg;
2107 }
2108
2109 #endif /* GROUP_IOSCHED */
2110
2111 /*
2112  * The cfqd->service_trees holds all pending cfq_queue's that have
2113  * requests waiting to be processed. It is sorted in the order that
2114  * we will service the queues.
2115  */
2116 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2117                                  bool add_front)
2118 {
2119         struct rb_node **p, *parent;
2120         struct cfq_queue *__cfqq;
2121         unsigned long rb_key;
2122         struct cfq_rb_root *st;
2123         int left;
2124         int new_cfqq = 1;
2125
2126         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2127         if (cfq_class_idle(cfqq)) {
2128                 rb_key = CFQ_IDLE_DELAY;
2129                 parent = rb_last(&st->rb);
2130                 if (parent && parent != &cfqq->rb_node) {
2131                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2132                         rb_key += __cfqq->rb_key;
2133                 } else
2134                         rb_key += jiffies;
2135         } else if (!add_front) {
2136                 /*
2137                  * Get our rb key offset. Subtract any residual slice
2138                  * value carried from last service. A negative resid
2139                  * count indicates slice overrun, and this should position
2140                  * the next service time further away in the tree.
2141                  */
2142                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2143                 rb_key -= cfqq->slice_resid;
2144                 cfqq->slice_resid = 0;
2145         } else {
2146                 rb_key = -HZ;
2147                 __cfqq = cfq_rb_first(st);
2148                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2149         }
2150
2151         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2152                 new_cfqq = 0;
2153                 /*
2154                  * same position, nothing more to do
2155                  */
2156                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2157                         return;
2158
2159                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2160                 cfqq->service_tree = NULL;
2161         }
2162
2163         left = 1;
2164         parent = NULL;
2165         cfqq->service_tree = st;
2166         p = &st->rb.rb_node;
2167         while (*p) {
2168                 parent = *p;
2169                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2170
2171                 /*
2172                  * sort by key, that represents service time.
2173                  */
2174                 if (time_before(rb_key, __cfqq->rb_key))
2175                         p = &parent->rb_left;
2176                 else {
2177                         p = &parent->rb_right;
2178                         left = 0;
2179                 }
2180         }
2181
2182         if (left)
2183                 st->left = &cfqq->rb_node;
2184
2185         cfqq->rb_key = rb_key;
2186         rb_link_node(&cfqq->rb_node, parent, p);
2187         rb_insert_color(&cfqq->rb_node, &st->rb);
2188         st->count++;
2189         if (add_front || !new_cfqq)
2190                 return;
2191         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2192 }
2193
2194 static struct cfq_queue *
2195 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2196                      sector_t sector, struct rb_node **ret_parent,
2197                      struct rb_node ***rb_link)
2198 {
2199         struct rb_node **p, *parent;
2200         struct cfq_queue *cfqq = NULL;
2201
2202         parent = NULL;
2203         p = &root->rb_node;
2204         while (*p) {
2205                 struct rb_node **n;
2206
2207                 parent = *p;
2208                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2209
2210                 /*
2211                  * Sort strictly based on sector.  Smallest to the left,
2212                  * largest to the right.
2213                  */
2214                 if (sector > blk_rq_pos(cfqq->next_rq))
2215                         n = &(*p)->rb_right;
2216                 else if (sector < blk_rq_pos(cfqq->next_rq))
2217                         n = &(*p)->rb_left;
2218                 else
2219                         break;
2220                 p = n;
2221                 cfqq = NULL;
2222         }
2223
2224         *ret_parent = parent;
2225         if (rb_link)
2226                 *rb_link = p;
2227         return cfqq;
2228 }
2229
2230 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2231 {
2232         struct rb_node **p, *parent;
2233         struct cfq_queue *__cfqq;
2234
2235         if (cfqq->p_root) {
2236                 rb_erase(&cfqq->p_node, cfqq->p_root);
2237                 cfqq->p_root = NULL;
2238         }
2239
2240         if (cfq_class_idle(cfqq))
2241                 return;
2242         if (!cfqq->next_rq)
2243                 return;
2244
2245         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2246         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2247                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2248         if (!__cfqq) {
2249                 rb_link_node(&cfqq->p_node, parent, p);
2250                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2251         } else
2252                 cfqq->p_root = NULL;
2253 }
2254
2255 /*
2256  * Update cfqq's position in the service tree.
2257  */
2258 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2259 {
2260         /*
2261          * Resorting requires the cfqq to be on the RR list already.
2262          */
2263         if (cfq_cfqq_on_rr(cfqq)) {
2264                 cfq_service_tree_add(cfqd, cfqq, 0);
2265                 cfq_prio_tree_add(cfqd, cfqq);
2266         }
2267 }
2268
2269 /*
2270  * add to busy list of queues for service, trying to be fair in ordering
2271  * the pending list according to last request service
2272  */
2273 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2274 {
2275         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2276         BUG_ON(cfq_cfqq_on_rr(cfqq));
2277         cfq_mark_cfqq_on_rr(cfqq);
2278         cfqd->busy_queues++;
2279         if (cfq_cfqq_sync(cfqq))
2280                 cfqd->busy_sync_queues++;
2281
2282         cfq_resort_rr_list(cfqd, cfqq);
2283 }
2284
2285 /*
2286  * Called when the cfqq no longer has requests pending, remove it from
2287  * the service tree.
2288  */
2289 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2290 {
2291         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2292         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2293         cfq_clear_cfqq_on_rr(cfqq);
2294
2295         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2296                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2297                 cfqq->service_tree = NULL;
2298         }
2299         if (cfqq->p_root) {
2300                 rb_erase(&cfqq->p_node, cfqq->p_root);
2301                 cfqq->p_root = NULL;
2302         }
2303
2304         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2305         BUG_ON(!cfqd->busy_queues);
2306         cfqd->busy_queues--;
2307         if (cfq_cfqq_sync(cfqq))
2308                 cfqd->busy_sync_queues--;
2309 }
2310
2311 /*
2312  * rb tree support functions
2313  */
2314 static void cfq_del_rq_rb(struct request *rq)
2315 {
2316         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2317         const int sync = rq_is_sync(rq);
2318
2319         BUG_ON(!cfqq->queued[sync]);
2320         cfqq->queued[sync]--;
2321
2322         elv_rb_del(&cfqq->sort_list, rq);
2323
2324         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2325                 /*
2326                  * Queue will be deleted from service tree when we actually
2327                  * expire it later. Right now just remove it from prio tree
2328                  * as it is empty.
2329                  */
2330                 if (cfqq->p_root) {
2331                         rb_erase(&cfqq->p_node, cfqq->p_root);
2332                         cfqq->p_root = NULL;
2333                 }
2334         }
2335 }
2336
2337 static void cfq_add_rq_rb(struct request *rq)
2338 {
2339         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2340         struct cfq_data *cfqd = cfqq->cfqd;
2341         struct request *prev;
2342
2343         cfqq->queued[rq_is_sync(rq)]++;
2344
2345         elv_rb_add(&cfqq->sort_list, rq);
2346
2347         if (!cfq_cfqq_on_rr(cfqq))
2348                 cfq_add_cfqq_rr(cfqd, cfqq);
2349
2350         /*
2351          * check if this request is a better next-serve candidate
2352          */
2353         prev = cfqq->next_rq;
2354         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2355
2356         /*
2357          * adjust priority tree position, if ->next_rq changes
2358          */
2359         if (prev != cfqq->next_rq)
2360                 cfq_prio_tree_add(cfqd, cfqq);
2361
2362         BUG_ON(!cfqq->next_rq);
2363 }
2364
2365 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2366 {
2367         elv_rb_del(&cfqq->sort_list, rq);
2368         cfqq->queued[rq_is_sync(rq)]--;
2369         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2370         cfq_add_rq_rb(rq);
2371         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2372                                  rq->cmd_flags);
2373 }
2374
2375 static struct request *
2376 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2377 {
2378         struct task_struct *tsk = current;
2379         struct cfq_io_cq *cic;
2380         struct cfq_queue *cfqq;
2381
2382         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2383         if (!cic)
2384                 return NULL;
2385
2386         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2387         if (cfqq)
2388                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2389
2390         return NULL;
2391 }
2392
2393 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2394 {
2395         struct cfq_data *cfqd = q->elevator->elevator_data;
2396
2397         cfqd->rq_in_driver++;
2398         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2399                                                 cfqd->rq_in_driver);
2400
2401         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2402 }
2403
2404 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2405 {
2406         struct cfq_data *cfqd = q->elevator->elevator_data;
2407
2408         WARN_ON(!cfqd->rq_in_driver);
2409         cfqd->rq_in_driver--;
2410         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2411                                                 cfqd->rq_in_driver);
2412 }
2413
2414 static void cfq_remove_request(struct request *rq)
2415 {
2416         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2417
2418         if (cfqq->next_rq == rq)
2419                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2420
2421         list_del_init(&rq->queuelist);
2422         cfq_del_rq_rb(rq);
2423
2424         cfqq->cfqd->rq_queued--;
2425         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2426         if (rq->cmd_flags & REQ_PRIO) {
2427                 WARN_ON(!cfqq->prio_pending);
2428                 cfqq->prio_pending--;
2429         }
2430 }
2431
2432 static int cfq_merge(struct request_queue *q, struct request **req,
2433                      struct bio *bio)
2434 {
2435         struct cfq_data *cfqd = q->elevator->elevator_data;
2436         struct request *__rq;
2437
2438         __rq = cfq_find_rq_fmerge(cfqd, bio);
2439         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2440                 *req = __rq;
2441                 return ELEVATOR_FRONT_MERGE;
2442         }
2443
2444         return ELEVATOR_NO_MERGE;
2445 }
2446
2447 static void cfq_merged_request(struct request_queue *q, struct request *req,
2448                                int type)
2449 {
2450         if (type == ELEVATOR_FRONT_MERGE) {
2451                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2452
2453                 cfq_reposition_rq_rb(cfqq, req);
2454         }
2455 }
2456
2457 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2458                                 struct bio *bio)
2459 {
2460         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2461 }
2462
2463 static void
2464 cfq_merged_requests(struct request_queue *q, struct request *rq,
2465                     struct request *next)
2466 {
2467         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2468         struct cfq_data *cfqd = q->elevator->elevator_data;
2469
2470         /*
2471          * reposition in fifo if next is older than rq
2472          */
2473         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2474             time_before(next->fifo_time, rq->fifo_time) &&
2475             cfqq == RQ_CFQQ(next)) {
2476                 list_move(&rq->queuelist, &next->queuelist);
2477                 rq->fifo_time = next->fifo_time;
2478         }
2479
2480         if (cfqq->next_rq == next)
2481                 cfqq->next_rq = rq;
2482         cfq_remove_request(next);
2483         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2484
2485         cfqq = RQ_CFQQ(next);
2486         /*
2487          * all requests of this queue are merged to other queues, delete it
2488          * from the service tree. If it's the active_queue,
2489          * cfq_dispatch_requests() will choose to expire it or do idle
2490          */
2491         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2492             cfqq != cfqd->active_queue)
2493                 cfq_del_cfqq_rr(cfqd, cfqq);
2494 }
2495
2496 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2497                            struct bio *bio)
2498 {
2499         struct cfq_data *cfqd = q->elevator->elevator_data;
2500         struct cfq_io_cq *cic;
2501         struct cfq_queue *cfqq;
2502
2503         /*
2504          * Disallow merge of a sync bio into an async request.
2505          */
2506         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2507                 return false;
2508
2509         /*
2510          * Lookup the cfqq that this bio will be queued with and allow
2511          * merge only if rq is queued there.
2512          */
2513         cic = cfq_cic_lookup(cfqd, current->io_context);
2514         if (!cic)
2515                 return false;
2516
2517         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2518         return cfqq == RQ_CFQQ(rq);
2519 }
2520
2521 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2522 {
2523         del_timer(&cfqd->idle_slice_timer);
2524         cfqg_stats_update_idle_time(cfqq->cfqg);
2525 }
2526
2527 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2528                                    struct cfq_queue *cfqq)
2529 {
2530         if (cfqq) {
2531                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2532                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2533                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2534                 cfqq->slice_start = 0;
2535                 cfqq->dispatch_start = jiffies;
2536                 cfqq->allocated_slice = 0;
2537                 cfqq->slice_end = 0;
2538                 cfqq->slice_dispatch = 0;
2539                 cfqq->nr_sectors = 0;
2540
2541                 cfq_clear_cfqq_wait_request(cfqq);
2542                 cfq_clear_cfqq_must_dispatch(cfqq);
2543                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2544                 cfq_clear_cfqq_fifo_expire(cfqq);
2545                 cfq_mark_cfqq_slice_new(cfqq);
2546
2547                 cfq_del_timer(cfqd, cfqq);
2548         }
2549
2550         cfqd->active_queue = cfqq;
2551 }
2552
2553 /*
2554  * current cfqq expired its slice (or was too idle), select new one
2555  */
2556 static void
2557 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2558                     bool timed_out)
2559 {
2560         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2561
2562         if (cfq_cfqq_wait_request(cfqq))
2563                 cfq_del_timer(cfqd, cfqq);
2564
2565         cfq_clear_cfqq_wait_request(cfqq);
2566         cfq_clear_cfqq_wait_busy(cfqq);
2567
2568         /*
2569          * If this cfqq is shared between multiple processes, check to
2570          * make sure that those processes are still issuing I/Os within
2571          * the mean seek distance.  If not, it may be time to break the
2572          * queues apart again.
2573          */
2574         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2575                 cfq_mark_cfqq_split_coop(cfqq);
2576
2577         /*
2578          * store what was left of this slice, if the queue idled/timed out
2579          */
2580         if (timed_out) {
2581                 if (cfq_cfqq_slice_new(cfqq))
2582                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2583                 else
2584                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2585                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2586         }
2587
2588         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2589
2590         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2591                 cfq_del_cfqq_rr(cfqd, cfqq);
2592
2593         cfq_resort_rr_list(cfqd, cfqq);
2594
2595         if (cfqq == cfqd->active_queue)
2596                 cfqd->active_queue = NULL;
2597
2598         if (cfqd->active_cic) {
2599                 put_io_context(cfqd->active_cic->icq.ioc);
2600                 cfqd->active_cic = NULL;
2601         }
2602 }
2603
2604 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2605 {
2606         struct cfq_queue *cfqq = cfqd->active_queue;
2607
2608         if (cfqq)
2609                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2610 }
2611
2612 /*
2613  * Get next queue for service. Unless we have a queue preemption,
2614  * we'll simply select the first cfqq in the service tree.
2615  */
2616 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2617 {
2618         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2619                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2620
2621         if (!cfqd->rq_queued)
2622                 return NULL;
2623
2624         /* There is nothing to dispatch */
2625         if (!st)
2626                 return NULL;
2627         if (RB_EMPTY_ROOT(&st->rb))
2628                 return NULL;
2629         return cfq_rb_first(st);
2630 }
2631
2632 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2633 {
2634         struct cfq_group *cfqg;
2635         struct cfq_queue *cfqq;
2636         int i, j;
2637         struct cfq_rb_root *st;
2638
2639         if (!cfqd->rq_queued)
2640                 return NULL;
2641
2642         cfqg = cfq_get_next_cfqg(cfqd);
2643         if (!cfqg)
2644                 return NULL;
2645
2646         for_each_cfqg_st(cfqg, i, j, st)
2647                 if ((cfqq = cfq_rb_first(st)) != NULL)
2648                         return cfqq;
2649         return NULL;
2650 }
2651
2652 /*
2653  * Get and set a new active queue for service.
2654  */
2655 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2656                                               struct cfq_queue *cfqq)
2657 {
2658         if (!cfqq)
2659                 cfqq = cfq_get_next_queue(cfqd);
2660
2661         __cfq_set_active_queue(cfqd, cfqq);
2662         return cfqq;
2663 }
2664
2665 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2666                                           struct request *rq)
2667 {
2668         if (blk_rq_pos(rq) >= cfqd->last_position)
2669                 return blk_rq_pos(rq) - cfqd->last_position;
2670         else
2671                 return cfqd->last_position - blk_rq_pos(rq);
2672 }
2673
2674 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2675                                struct request *rq)
2676 {
2677         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2678 }
2679
2680 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2681                                     struct cfq_queue *cur_cfqq)
2682 {
2683         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2684         struct rb_node *parent, *node;
2685         struct cfq_queue *__cfqq;
2686         sector_t sector = cfqd->last_position;
2687
2688         if (RB_EMPTY_ROOT(root))
2689                 return NULL;
2690
2691         /*
2692          * First, if we find a request starting at the end of the last
2693          * request, choose it.
2694          */
2695         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2696         if (__cfqq)
2697                 return __cfqq;
2698
2699         /*
2700          * If the exact sector wasn't found, the parent of the NULL leaf
2701          * will contain the closest sector.
2702          */
2703         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2704         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2705                 return __cfqq;
2706
2707         if (blk_rq_pos(__cfqq->next_rq) < sector)
2708                 node = rb_next(&__cfqq->p_node);
2709         else
2710                 node = rb_prev(&__cfqq->p_node);
2711         if (!node)
2712                 return NULL;
2713
2714         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2715         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2716                 return __cfqq;
2717
2718         return NULL;
2719 }
2720
2721 /*
2722  * cfqd - obvious
2723  * cur_cfqq - passed in so that we don't decide that the current queue is
2724  *            closely cooperating with itself.
2725  *
2726  * So, basically we're assuming that that cur_cfqq has dispatched at least
2727  * one request, and that cfqd->last_position reflects a position on the disk
2728  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2729  * assumption.
2730  */
2731 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2732                                               struct cfq_queue *cur_cfqq)
2733 {
2734         struct cfq_queue *cfqq;
2735
2736         if (cfq_class_idle(cur_cfqq))
2737                 return NULL;
2738         if (!cfq_cfqq_sync(cur_cfqq))
2739                 return NULL;
2740         if (CFQQ_SEEKY(cur_cfqq))
2741                 return NULL;
2742
2743         /*
2744          * Don't search priority tree if it's the only queue in the group.
2745          */
2746         if (cur_cfqq->cfqg->nr_cfqq == 1)
2747                 return NULL;
2748
2749         /*
2750          * We should notice if some of the queues are cooperating, eg
2751          * working closely on the same area of the disk. In that case,
2752          * we can group them together and don't waste time idling.
2753          */
2754         cfqq = cfqq_close(cfqd, cur_cfqq);
2755         if (!cfqq)
2756                 return NULL;
2757
2758         /* If new queue belongs to different cfq_group, don't choose it */
2759         if (cur_cfqq->cfqg != cfqq->cfqg)
2760                 return NULL;
2761
2762         /*
2763          * It only makes sense to merge sync queues.
2764          */
2765         if (!cfq_cfqq_sync(cfqq))
2766                 return NULL;
2767         if (CFQQ_SEEKY(cfqq))
2768                 return NULL;
2769
2770         /*
2771          * Do not merge queues of different priority classes
2772          */
2773         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2774                 return NULL;
2775
2776         return cfqq;
2777 }
2778
2779 /*
2780  * Determine whether we should enforce idle window for this queue.
2781  */
2782
2783 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2784 {
2785         enum wl_class_t wl_class = cfqq_class(cfqq);
2786         struct cfq_rb_root *st = cfqq->service_tree;
2787
2788         BUG_ON(!st);
2789         BUG_ON(!st->count);
2790
2791         if (!cfqd->cfq_slice_idle)
2792                 return false;
2793
2794         /* We never do for idle class queues. */
2795         if (wl_class == IDLE_WORKLOAD)
2796                 return false;
2797
2798         /* We do for queues that were marked with idle window flag. */
2799         if (cfq_cfqq_idle_window(cfqq) &&
2800            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2801                 return true;
2802
2803         /*
2804          * Otherwise, we do only if they are the last ones
2805          * in their service tree.
2806          */
2807         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2808            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2809                 return true;
2810         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2811         return false;
2812 }
2813
2814 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2815 {
2816         struct cfq_queue *cfqq = cfqd->active_queue;
2817         struct cfq_io_cq *cic;
2818         unsigned long sl, group_idle = 0;
2819
2820         /*
2821          * SSD device without seek penalty, disable idling. But only do so
2822          * for devices that support queuing, otherwise we still have a problem
2823          * with sync vs async workloads.
2824          */
2825         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2826                 return;
2827
2828         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2829         WARN_ON(cfq_cfqq_slice_new(cfqq));
2830
2831         /*
2832          * idle is disabled, either manually or by past process history
2833          */
2834         if (!cfq_should_idle(cfqd, cfqq)) {
2835                 /* no queue idling. Check for group idling */
2836                 if (cfqd->cfq_group_idle)
2837                         group_idle = cfqd->cfq_group_idle;
2838                 else
2839                         return;
2840         }
2841
2842         /*
2843          * still active requests from this queue, don't idle
2844          */
2845         if (cfqq->dispatched)
2846                 return;
2847
2848         /*
2849          * task has exited, don't wait
2850          */
2851         cic = cfqd->active_cic;
2852         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2853                 return;
2854
2855         /*
2856          * If our average think time is larger than the remaining time
2857          * slice, then don't idle. This avoids overrunning the allotted
2858          * time slice.
2859          */
2860         if (sample_valid(cic->ttime.ttime_samples) &&
2861             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2862                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2863                              cic->ttime.ttime_mean);
2864                 return;
2865         }
2866
2867         /* There are other queues in the group, don't do group idle */
2868         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2869                 return;
2870
2871         cfq_mark_cfqq_wait_request(cfqq);
2872
2873         if (group_idle)
2874                 sl = cfqd->cfq_group_idle;
2875         else
2876                 sl = cfqd->cfq_slice_idle;
2877
2878         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2879         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2880         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2881                         group_idle ? 1 : 0);
2882 }
2883
2884 /*
2885  * Move request from internal lists to the request queue dispatch list.
2886  */
2887 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2888 {
2889         struct cfq_data *cfqd = q->elevator->elevator_data;
2890         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2891
2892         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2893
2894         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2895         cfq_remove_request(rq);
2896         cfqq->dispatched++;
2897         (RQ_CFQG(rq))->dispatched++;
2898         elv_dispatch_sort(q, rq);
2899
2900         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2901         cfqq->nr_sectors += blk_rq_sectors(rq);
2902         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2903 }
2904
2905 /*
2906  * return expired entry, or NULL to just start from scratch in rbtree
2907  */
2908 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2909 {
2910         struct request *rq = NULL;
2911
2912         if (cfq_cfqq_fifo_expire(cfqq))
2913                 return NULL;
2914
2915         cfq_mark_cfqq_fifo_expire(cfqq);
2916
2917         if (list_empty(&cfqq->fifo))
2918                 return NULL;
2919
2920         rq = rq_entry_fifo(cfqq->fifo.next);
2921         if (time_before(jiffies, rq->fifo_time))
2922                 rq = NULL;
2923
2924         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2925         return rq;
2926 }
2927
2928 static inline int
2929 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2930 {
2931         const int base_rq = cfqd->cfq_slice_async_rq;
2932
2933         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2934
2935         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2936 }
2937
2938 /*
2939  * Must be called with the queue_lock held.
2940  */
2941 static int cfqq_process_refs(struct cfq_queue *cfqq)
2942 {
2943         int process_refs, io_refs;
2944
2945         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2946         process_refs = cfqq->ref - io_refs;
2947         BUG_ON(process_refs < 0);
2948         return process_refs;
2949 }
2950
2951 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2952 {
2953         int process_refs, new_process_refs;
2954         struct cfq_queue *__cfqq;
2955
2956         /*
2957          * If there are no process references on the new_cfqq, then it is
2958          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2959          * chain may have dropped their last reference (not just their
2960          * last process reference).
2961          */
2962         if (!cfqq_process_refs(new_cfqq))
2963                 return;
2964
2965         /* Avoid a circular list and skip interim queue merges */
2966         while ((__cfqq = new_cfqq->new_cfqq)) {
2967                 if (__cfqq == cfqq)
2968                         return;
2969                 new_cfqq = __cfqq;
2970         }
2971
2972         process_refs = cfqq_process_refs(cfqq);
2973         new_process_refs = cfqq_process_refs(new_cfqq);
2974         /*
2975          * If the process for the cfqq has gone away, there is no
2976          * sense in merging the queues.
2977          */
2978         if (process_refs == 0 || new_process_refs == 0)
2979                 return;
2980
2981         /*
2982          * Merge in the direction of the lesser amount of work.
2983          */
2984         if (new_process_refs >= process_refs) {
2985                 cfqq->new_cfqq = new_cfqq;
2986                 new_cfqq->ref += process_refs;
2987         } else {
2988                 new_cfqq->new_cfqq = cfqq;
2989                 cfqq->ref += new_process_refs;
2990         }
2991 }
2992
2993 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2994                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2995 {
2996         struct cfq_queue *queue;
2997         int i;
2998         bool key_valid = false;
2999         unsigned long lowest_key = 0;
3000         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3001
3002         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3003                 /* select the one with lowest rb_key */
3004                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3005                 if (queue &&
3006                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
3007                         lowest_key = queue->rb_key;
3008                         cur_best = i;
3009                         key_valid = true;
3010                 }
3011         }
3012
3013         return cur_best;
3014 }
3015
3016 static void
3017 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3018 {
3019         unsigned slice;
3020         unsigned count;
3021         struct cfq_rb_root *st;
3022         unsigned group_slice;
3023         enum wl_class_t original_class = cfqd->serving_wl_class;
3024
3025         /* Choose next priority. RT > BE > IDLE */
3026         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3027                 cfqd->serving_wl_class = RT_WORKLOAD;
3028         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3029                 cfqd->serving_wl_class = BE_WORKLOAD;
3030         else {
3031                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3032                 cfqd->workload_expires = jiffies + 1;
3033                 return;
3034         }
3035
3036         if (original_class != cfqd->serving_wl_class)
3037                 goto new_workload;
3038
3039         /*
3040          * For RT and BE, we have to choose also the type
3041          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3042          * expiration time
3043          */
3044         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3045         count = st->count;
3046
3047         /*
3048          * check workload expiration, and that we still have other queues ready
3049          */
3050         if (count && !time_after(jiffies, cfqd->workload_expires))
3051                 return;
3052
3053 new_workload:
3054         /* otherwise select new workload type */
3055         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3056                                         cfqd->serving_wl_class);
3057         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3058         count = st->count;
3059
3060         /*
3061          * the workload slice is computed as a fraction of target latency
3062          * proportional to the number of queues in that workload, over
3063          * all the queues in the same priority class
3064          */
3065         group_slice = cfq_group_slice(cfqd, cfqg);
3066
3067         slice = group_slice * count /
3068                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3069                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3070                                         cfqg));
3071
3072         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3073                 unsigned int tmp;
3074
3075                 /*
3076                  * Async queues are currently system wide. Just taking
3077                  * proportion of queues with-in same group will lead to higher
3078                  * async ratio system wide as generally root group is going
3079                  * to have higher weight. A more accurate thing would be to
3080                  * calculate system wide asnc/sync ratio.
3081                  */
3082                 tmp = cfqd->cfq_target_latency *
3083                         cfqg_busy_async_queues(cfqd, cfqg);
3084                 tmp = tmp/cfqd->busy_queues;
3085                 slice = min_t(unsigned, slice, tmp);
3086
3087                 /* async workload slice is scaled down according to
3088                  * the sync/async slice ratio. */
3089                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3090         } else
3091                 /* sync workload slice is at least 2 * cfq_slice_idle */
3092                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3093
3094         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3095         cfq_log(cfqd, "workload slice:%d", slice);
3096         cfqd->workload_expires = jiffies + slice;
3097 }
3098
3099 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3100 {
3101         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3102         struct cfq_group *cfqg;
3103
3104         if (RB_EMPTY_ROOT(&st->rb))
3105                 return NULL;
3106         cfqg = cfq_rb_first_group(st);
3107         update_min_vdisktime(st);
3108         return cfqg;
3109 }
3110
3111 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3112 {
3113         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3114
3115         cfqd->serving_group = cfqg;
3116
3117         /* Restore the workload type data */
3118         if (cfqg->saved_wl_slice) {
3119                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3120                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3121                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3122         } else
3123                 cfqd->workload_expires = jiffies - 1;
3124
3125         choose_wl_class_and_type(cfqd, cfqg);
3126 }
3127
3128 /*
3129  * Select a queue for service. If we have a current active queue,
3130  * check whether to continue servicing it, or retrieve and set a new one.
3131  */
3132 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3133 {
3134         struct cfq_queue *cfqq, *new_cfqq = NULL;
3135
3136         cfqq = cfqd->active_queue;
3137         if (!cfqq)
3138                 goto new_queue;
3139
3140         if (!cfqd->rq_queued)
3141                 return NULL;
3142
3143         /*
3144          * We were waiting for group to get backlogged. Expire the queue
3145          */
3146         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3147                 goto expire;
3148
3149         /*
3150          * The active queue has run out of time, expire it and select new.
3151          */
3152         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3153                 /*
3154                  * If slice had not expired at the completion of last request
3155                  * we might not have turned on wait_busy flag. Don't expire
3156                  * the queue yet. Allow the group to get backlogged.
3157                  *
3158                  * The very fact that we have used the slice, that means we
3159                  * have been idling all along on this queue and it should be
3160                  * ok to wait for this request to complete.
3161                  */
3162                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3163                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3164                         cfqq = NULL;
3165                         goto keep_queue;
3166                 } else
3167                         goto check_group_idle;
3168         }
3169
3170         /*
3171          * The active queue has requests and isn't expired, allow it to
3172          * dispatch.
3173          */
3174         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3175                 goto keep_queue;
3176
3177         /*
3178          * If another queue has a request waiting within our mean seek
3179          * distance, let it run.  The expire code will check for close
3180          * cooperators and put the close queue at the front of the service
3181          * tree.  If possible, merge the expiring queue with the new cfqq.
3182          */
3183         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3184         if (new_cfqq) {
3185                 if (!cfqq->new_cfqq)
3186                         cfq_setup_merge(cfqq, new_cfqq);
3187                 goto expire;
3188         }
3189
3190         /*
3191          * No requests pending. If the active queue still has requests in
3192          * flight or is idling for a new request, allow either of these
3193          * conditions to happen (or time out) before selecting a new queue.
3194          */
3195         if (timer_pending(&cfqd->idle_slice_timer)) {
3196                 cfqq = NULL;
3197                 goto keep_queue;
3198         }
3199
3200         /*
3201          * This is a deep seek queue, but the device is much faster than
3202          * the queue can deliver, don't idle
3203          **/
3204         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3205             (cfq_cfqq_slice_new(cfqq) ||
3206             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3207                 cfq_clear_cfqq_deep(cfqq);
3208                 cfq_clear_cfqq_idle_window(cfqq);
3209         }
3210
3211         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3212                 cfqq = NULL;
3213                 goto keep_queue;
3214         }
3215
3216         /*
3217          * If group idle is enabled and there are requests dispatched from
3218          * this group, wait for requests to complete.
3219          */
3220 check_group_idle:
3221         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3222             cfqq->cfqg->dispatched &&
3223             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3224                 cfqq = NULL;
3225                 goto keep_queue;
3226         }
3227
3228 expire:
3229         cfq_slice_expired(cfqd, 0);
3230 new_queue:
3231         /*
3232          * Current queue expired. Check if we have to switch to a new
3233          * service tree
3234          */
3235         if (!new_cfqq)
3236                 cfq_choose_cfqg(cfqd);
3237
3238         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3239 keep_queue:
3240         return cfqq;
3241 }
3242
3243 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3244 {
3245         int dispatched = 0;
3246
3247         while (cfqq->next_rq) {
3248                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3249                 dispatched++;
3250         }
3251
3252         BUG_ON(!list_empty(&cfqq->fifo));
3253
3254         /* By default cfqq is not expired if it is empty. Do it explicitly */
3255         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3256         return dispatched;
3257 }
3258
3259 /*
3260  * Drain our current requests. Used for barriers and when switching
3261  * io schedulers on-the-fly.
3262  */
3263 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3264 {
3265         struct cfq_queue *cfqq;
3266         int dispatched = 0;
3267
3268         /* Expire the timeslice of the current active queue first */
3269         cfq_slice_expired(cfqd, 0);
3270         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3271                 __cfq_set_active_queue(cfqd, cfqq);
3272                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3273         }
3274
3275         BUG_ON(cfqd->busy_queues);
3276
3277         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3278         return dispatched;
3279 }
3280
3281 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3282         struct cfq_queue *cfqq)
3283 {
3284         /* the queue hasn't finished any request, can't estimate */
3285         if (cfq_cfqq_slice_new(cfqq))
3286                 return true;
3287         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3288                 cfqq->slice_end))
3289                 return true;
3290
3291         return false;
3292 }
3293
3294 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3295 {
3296         unsigned int max_dispatch;
3297
3298         /*
3299          * Drain async requests before we start sync IO
3300          */
3301         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3302                 return false;
3303
3304         /*
3305          * If this is an async queue and we have sync IO in flight, let it wait
3306          */
3307         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3308                 return false;
3309
3310         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3311         if (cfq_class_idle(cfqq))
3312                 max_dispatch = 1;
3313
3314         /*
3315          * Does this cfqq already have too much IO in flight?
3316          */
3317         if (cfqq->dispatched >= max_dispatch) {
3318                 bool promote_sync = false;
3319                 /*
3320                  * idle queue must always only have a single IO in flight
3321                  */
3322                 if (cfq_class_idle(cfqq))
3323                         return false;
3324
3325                 /*
3326                  * If there is only one sync queue
3327                  * we can ignore async queue here and give the sync
3328                  * queue no dispatch limit. The reason is a sync queue can
3329                  * preempt async queue, limiting the sync queue doesn't make
3330                  * sense. This is useful for aiostress test.
3331                  */
3332                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3333                         promote_sync = true;
3334
3335                 /*
3336                  * We have other queues, don't allow more IO from this one
3337                  */
3338                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3339                                 !promote_sync)
3340                         return false;
3341
3342                 /*
3343                  * Sole queue user, no limit
3344                  */
3345                 if (cfqd->busy_queues == 1 || promote_sync)
3346                         max_dispatch = -1;
3347                 else
3348                         /*
3349                          * Normally we start throttling cfqq when cfq_quantum/2
3350                          * requests have been dispatched. But we can drive
3351                          * deeper queue depths at the beginning of slice
3352                          * subjected to upper limit of cfq_quantum.
3353                          * */
3354                         max_dispatch = cfqd->cfq_quantum;
3355         }
3356
3357         /*
3358          * Async queues must wait a bit before being allowed dispatch.
3359          * We also ramp up the dispatch depth gradually for async IO,
3360          * based on the last sync IO we serviced
3361          */
3362         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3363                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3364                 unsigned int depth;
3365
3366                 depth = last_sync / cfqd->cfq_slice[1];
3367                 if (!depth && !cfqq->dispatched)
3368                         depth = 1;
3369                 if (depth < max_dispatch)
3370                         max_dispatch = depth;
3371         }
3372
3373         /*
3374          * If we're below the current max, allow a dispatch
3375          */
3376         return cfqq->dispatched < max_dispatch;
3377 }
3378
3379 /*
3380  * Dispatch a request from cfqq, moving them to the request queue
3381  * dispatch list.
3382  */
3383 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3384 {
3385         struct request *rq;
3386
3387         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3388
3389         if (!cfq_may_dispatch(cfqd, cfqq))
3390                 return false;
3391
3392         /*
3393          * follow expired path, else get first next available
3394          */
3395         rq = cfq_check_fifo(cfqq);
3396         if (!rq)
3397                 rq = cfqq->next_rq;
3398
3399         /*
3400          * insert request into driver dispatch list
3401          */
3402         cfq_dispatch_insert(cfqd->queue, rq);
3403
3404         if (!cfqd->active_cic) {
3405                 struct cfq_io_cq *cic = RQ_CIC(rq);
3406
3407                 atomic_long_inc(&cic->icq.ioc->refcount);
3408                 cfqd->active_cic = cic;
3409         }
3410
3411         return true;
3412 }
3413
3414 /*
3415  * Find the cfqq that we need to service and move a request from that to the
3416  * dispatch list
3417  */
3418 static int cfq_dispatch_requests(struct request_queue *q, int force)
3419 {
3420         struct cfq_data *cfqd = q->elevator->elevator_data;
3421         struct cfq_queue *cfqq;
3422
3423         if (!cfqd->busy_queues)
3424                 return 0;
3425
3426         if (unlikely(force))
3427                 return cfq_forced_dispatch(cfqd);
3428
3429         cfqq = cfq_select_queue(cfqd);
3430         if (!cfqq)
3431                 return 0;
3432
3433         /*
3434          * Dispatch a request from this cfqq, if it is allowed
3435          */
3436         if (!cfq_dispatch_request(cfqd, cfqq))
3437                 return 0;
3438
3439         cfqq->slice_dispatch++;
3440         cfq_clear_cfqq_must_dispatch(cfqq);
3441
3442         /*
3443          * expire an async queue immediately if it has used up its slice. idle
3444          * queue always expire after 1 dispatch round.
3445          */
3446         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3447             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3448             cfq_class_idle(cfqq))) {
3449                 cfqq->slice_end = jiffies + 1;
3450                 cfq_slice_expired(cfqd, 0);
3451         }
3452
3453         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3454         return 1;
3455 }
3456
3457 /*
3458  * task holds one reference to the queue, dropped when task exits. each rq
3459  * in-flight on this queue also holds a reference, dropped when rq is freed.
3460  *
3461  * Each cfq queue took a reference on the parent group. Drop it now.
3462  * queue lock must be held here.
3463  */
3464 static void cfq_put_queue(struct cfq_queue *cfqq)
3465 {
3466         struct cfq_data *cfqd = cfqq->cfqd;
3467         struct cfq_group *cfqg;
3468
3469         BUG_ON(cfqq->ref <= 0);
3470
3471         cfqq->ref--;
3472         if (cfqq->ref)
3473                 return;
3474
3475         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3476         BUG_ON(rb_first(&cfqq->sort_list));
3477         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3478         cfqg = cfqq->cfqg;
3479
3480         if (unlikely(cfqd->active_queue == cfqq)) {
3481                 __cfq_slice_expired(cfqd, cfqq, 0);
3482                 cfq_schedule_dispatch(cfqd);
3483         }
3484
3485         BUG_ON(cfq_cfqq_on_rr(cfqq));
3486         kmem_cache_free(cfq_pool, cfqq);
3487         cfqg_put(cfqg);
3488 }
3489
3490 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3491 {
3492         struct cfq_queue *__cfqq, *next;
3493
3494         /*
3495          * If this queue was scheduled to merge with another queue, be
3496          * sure to drop the reference taken on that queue (and others in
3497          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3498          */
3499         __cfqq = cfqq->new_cfqq;
3500         while (__cfqq) {
3501                 if (__cfqq == cfqq) {
3502                         WARN(1, "cfqq->new_cfqq loop detected\n");
3503                         break;
3504                 }
3505                 next = __cfqq->new_cfqq;
3506                 cfq_put_queue(__cfqq);
3507                 __cfqq = next;
3508         }
3509 }
3510
3511 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3512 {
3513         if (unlikely(cfqq == cfqd->active_queue)) {
3514                 __cfq_slice_expired(cfqd, cfqq, 0);
3515                 cfq_schedule_dispatch(cfqd);
3516         }
3517
3518         cfq_put_cooperator(cfqq);
3519
3520         cfq_put_queue(cfqq);
3521 }
3522
3523 static void cfq_init_icq(struct io_cq *icq)
3524 {
3525         struct cfq_io_cq *cic = icq_to_cic(icq);
3526
3527         cic->ttime.last_end_request = jiffies;
3528 }
3529
3530 static void cfq_exit_icq(struct io_cq *icq)
3531 {
3532         struct cfq_io_cq *cic = icq_to_cic(icq);
3533         struct cfq_data *cfqd = cic_to_cfqd(cic);
3534
3535         if (cic_to_cfqq(cic, false)) {
3536                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3537                 cic_set_cfqq(cic, NULL, false);
3538         }
3539
3540         if (cic_to_cfqq(cic, true)) {
3541                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3542                 cic_set_cfqq(cic, NULL, true);
3543         }
3544 }
3545
3546 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3547 {
3548         struct task_struct *tsk = current;
3549         int ioprio_class;
3550
3551         if (!cfq_cfqq_prio_changed(cfqq))
3552                 return;
3553
3554         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3555         switch (ioprio_class) {
3556         default:
3557                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3558         case IOPRIO_CLASS_NONE:
3559                 /*
3560                  * no prio set, inherit CPU scheduling settings
3561                  */
3562                 cfqq->ioprio = task_nice_ioprio(tsk);
3563                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3564                 break;
3565         case IOPRIO_CLASS_RT:
3566                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3567                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3568                 break;
3569         case IOPRIO_CLASS_BE:
3570                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3571                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3572                 break;
3573         case IOPRIO_CLASS_IDLE:
3574                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3575                 cfqq->ioprio = 7;
3576                 cfq_clear_cfqq_idle_window(cfqq);
3577                 break;
3578         }
3579
3580         /*
3581          * keep track of original prio settings in case we have to temporarily
3582          * elevate the priority of this queue
3583          */
3584         cfqq->org_ioprio = cfqq->ioprio;
3585         cfq_clear_cfqq_prio_changed(cfqq);
3586 }
3587
3588 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3589 {
3590         int ioprio = cic->icq.ioc->ioprio;
3591         struct cfq_data *cfqd = cic_to_cfqd(cic);
3592         struct cfq_queue *cfqq;
3593
3594         /*
3595          * Check whether ioprio has changed.  The condition may trigger
3596          * spuriously on a newly created cic but there's no harm.
3597          */
3598         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3599                 return;
3600
3601         cfqq = cic_to_cfqq(cic, false);
3602         if (cfqq) {
3603                 cfq_put_queue(cfqq);
3604                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3605                 cic_set_cfqq(cic, cfqq, false);
3606         }
3607
3608         cfqq = cic_to_cfqq(cic, true);
3609         if (cfqq)
3610                 cfq_mark_cfqq_prio_changed(cfqq);
3611
3612         cic->ioprio = ioprio;
3613 }
3614
3615 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3616                           pid_t pid, bool is_sync)
3617 {
3618         RB_CLEAR_NODE(&cfqq->rb_node);
3619         RB_CLEAR_NODE(&cfqq->p_node);
3620         INIT_LIST_HEAD(&cfqq->fifo);
3621
3622         cfqq->ref = 0;
3623         cfqq->cfqd = cfqd;
3624
3625         cfq_mark_cfqq_prio_changed(cfqq);
3626
3627         if (is_sync) {
3628                 if (!cfq_class_idle(cfqq))
3629                         cfq_mark_cfqq_idle_window(cfqq);
3630                 cfq_mark_cfqq_sync(cfqq);
3631         }
3632         cfqq->pid = pid;
3633 }
3634
3635 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3636 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3637 {
3638         struct cfq_data *cfqd = cic_to_cfqd(cic);
3639         struct cfq_queue *cfqq;
3640         uint64_t serial_nr;
3641
3642         rcu_read_lock();
3643         serial_nr = bio_blkcg(bio)->css.serial_nr;
3644         rcu_read_unlock();
3645
3646         /*
3647          * Check whether blkcg has changed.  The condition may trigger
3648          * spuriously on a newly created cic but there's no harm.
3649          */
3650         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3651                 return;
3652
3653         /*
3654          * Drop reference to queues.  New queues will be assigned in new
3655          * group upon arrival of fresh requests.
3656          */
3657         cfqq = cic_to_cfqq(cic, false);
3658         if (cfqq) {
3659                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3660                 cic_set_cfqq(cic, NULL, false);
3661                 cfq_put_queue(cfqq);
3662         }
3663
3664         cfqq = cic_to_cfqq(cic, true);
3665         if (cfqq) {
3666                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3667                 cic_set_cfqq(cic, NULL, true);
3668                 cfq_put_queue(cfqq);
3669         }
3670
3671         cic->blkcg_serial_nr = serial_nr;
3672 }
3673 #else
3674 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3675 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3676
3677 static struct cfq_queue **
3678 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3679 {
3680         switch (ioprio_class) {
3681         case IOPRIO_CLASS_RT:
3682                 return &cfqg->async_cfqq[0][ioprio];
3683         case IOPRIO_CLASS_NONE:
3684                 ioprio = IOPRIO_NORM;
3685                 /* fall through */
3686         case IOPRIO_CLASS_BE:
3687                 return &cfqg->async_cfqq[1][ioprio];
3688         case IOPRIO_CLASS_IDLE:
3689                 return &cfqg->async_idle_cfqq;
3690         default:
3691                 BUG();
3692         }
3693 }
3694
3695 static struct cfq_queue *
3696 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3697               struct bio *bio)
3698 {
3699         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3700         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3701         struct cfq_queue **async_cfqq = NULL;
3702         struct cfq_queue *cfqq;
3703         struct cfq_group *cfqg;
3704
3705         rcu_read_lock();
3706         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3707         if (!cfqg) {
3708                 cfqq = &cfqd->oom_cfqq;
3709                 goto out;
3710         }
3711
3712         if (!is_sync) {
3713                 if (!ioprio_valid(cic->ioprio)) {
3714                         struct task_struct *tsk = current;
3715                         ioprio = task_nice_ioprio(tsk);
3716                         ioprio_class = task_nice_ioclass(tsk);
3717                 }
3718                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3719                 cfqq = *async_cfqq;
3720                 if (cfqq)
3721                         goto out;
3722         }
3723
3724         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3725                                      cfqd->queue->node);
3726         if (!cfqq) {
3727                 cfqq = &cfqd->oom_cfqq;
3728                 goto out;
3729         }
3730
3731         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3732         cfq_init_prio_data(cfqq, cic);
3733         cfq_link_cfqq_cfqg(cfqq, cfqg);
3734         cfq_log_cfqq(cfqd, cfqq, "alloced");
3735
3736         if (async_cfqq) {
3737                 /* a new async queue is created, pin and remember */
3738                 cfqq->ref++;
3739                 *async_cfqq = cfqq;
3740         }
3741 out:
3742         cfqq->ref++;
3743         rcu_read_unlock();
3744         return cfqq;
3745 }
3746
3747 static void
3748 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3749 {
3750         unsigned long elapsed = jiffies - ttime->last_end_request;
3751         elapsed = min(elapsed, 2UL * slice_idle);
3752
3753         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3754         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3755         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3756 }
3757
3758 static void
3759 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3760                         struct cfq_io_cq *cic)
3761 {
3762         if (cfq_cfqq_sync(cfqq)) {
3763                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3764                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3765                         cfqd->cfq_slice_idle);
3766         }
3767 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3768         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3769 #endif
3770 }
3771
3772 static void
3773 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3774                        struct request *rq)
3775 {
3776         sector_t sdist = 0;
3777         sector_t n_sec = blk_rq_sectors(rq);
3778         if (cfqq->last_request_pos) {
3779                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3780                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3781                 else
3782                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3783         }
3784
3785         cfqq->seek_history <<= 1;
3786         if (blk_queue_nonrot(cfqd->queue))
3787                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3788         else
3789                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3790 }
3791
3792 /*
3793  * Disable idle window if the process thinks too long or seeks so much that
3794  * it doesn't matter
3795  */
3796 static void
3797 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3798                        struct cfq_io_cq *cic)
3799 {
3800         int old_idle, enable_idle;
3801
3802         /*
3803          * Don't idle for async or idle io prio class
3804          */
3805         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3806                 return;
3807
3808         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3809
3810         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3811                 cfq_mark_cfqq_deep(cfqq);
3812
3813         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3814                 enable_idle = 0;
3815         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3816                  !cfqd->cfq_slice_idle ||
3817                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3818                 enable_idle = 0;
3819         else if (sample_valid(cic->ttime.ttime_samples)) {
3820                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3821                         enable_idle = 0;
3822                 else
3823                         enable_idle = 1;
3824         }
3825
3826         if (old_idle != enable_idle) {
3827                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3828                 if (enable_idle)
3829                         cfq_mark_cfqq_idle_window(cfqq);
3830                 else
3831                         cfq_clear_cfqq_idle_window(cfqq);
3832         }
3833 }
3834
3835 /*
3836  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3837  * no or if we aren't sure, a 1 will cause a preempt.
3838  */
3839 static bool
3840 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3841                    struct request *rq)
3842 {
3843         struct cfq_queue *cfqq;
3844
3845         cfqq = cfqd->active_queue;
3846         if (!cfqq)
3847                 return false;
3848
3849         if (cfq_class_idle(new_cfqq))
3850                 return false;
3851
3852         if (cfq_class_idle(cfqq))
3853                 return true;
3854
3855         /*
3856          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3857          */
3858         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3859                 return false;
3860
3861         /*
3862          * if the new request is sync, but the currently running queue is
3863          * not, let the sync request have priority.
3864          */
3865         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3866                 return true;
3867
3868         if (new_cfqq->cfqg != cfqq->cfqg)
3869                 return false;
3870
3871         if (cfq_slice_used(cfqq))
3872                 return true;
3873
3874         /* Allow preemption only if we are idling on sync-noidle tree */
3875         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3876             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3877             new_cfqq->service_tree->count == 2 &&
3878             RB_EMPTY_ROOT(&cfqq->sort_list))
3879                 return true;
3880
3881         /*
3882          * So both queues are sync. Let the new request get disk time if
3883          * it's a metadata request and the current queue is doing regular IO.
3884          */
3885         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3886                 return true;
3887
3888         /*
3889          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3890          */
3891         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3892                 return true;
3893
3894         /* An idle queue should not be idle now for some reason */
3895         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3896                 return true;
3897
3898         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3899                 return false;
3900
3901         /*
3902          * if this request is as-good as one we would expect from the
3903          * current cfqq, let it preempt
3904          */
3905         if (cfq_rq_close(cfqd, cfqq, rq))
3906                 return true;
3907
3908         return false;
3909 }
3910
3911 /*
3912  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3913  * let it have half of its nominal slice.
3914  */
3915 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3916 {
3917         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3918
3919         cfq_log_cfqq(cfqd, cfqq, "preempt");
3920         cfq_slice_expired(cfqd, 1);
3921
3922         /*
3923          * workload type is changed, don't save slice, otherwise preempt
3924          * doesn't happen
3925          */
3926         if (old_type != cfqq_type(cfqq))
3927                 cfqq->cfqg->saved_wl_slice = 0;
3928
3929         /*
3930          * Put the new queue at the front of the of the current list,
3931          * so we know that it will be selected next.
3932          */
3933         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3934
3935         cfq_service_tree_add(cfqd, cfqq, 1);
3936
3937         cfqq->slice_end = 0;
3938         cfq_mark_cfqq_slice_new(cfqq);
3939 }
3940
3941 /*
3942  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3943  * something we should do about it
3944  */
3945 static void
3946 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3947                 struct request *rq)
3948 {
3949         struct cfq_io_cq *cic = RQ_CIC(rq);
3950
3951         cfqd->rq_queued++;
3952         if (rq->cmd_flags & REQ_PRIO)
3953                 cfqq->prio_pending++;
3954
3955         cfq_update_io_thinktime(cfqd, cfqq, cic);
3956         cfq_update_io_seektime(cfqd, cfqq, rq);
3957         cfq_update_idle_window(cfqd, cfqq, cic);
3958
3959         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3960
3961         if (cfqq == cfqd->active_queue) {
3962                 /*
3963                  * Remember that we saw a request from this process, but
3964                  * don't start queuing just yet. Otherwise we risk seeing lots
3965                  * of tiny requests, because we disrupt the normal plugging
3966                  * and merging. If the request is already larger than a single
3967                  * page, let it rip immediately. For that case we assume that
3968                  * merging is already done. Ditto for a busy system that
3969                  * has other work pending, don't risk delaying until the
3970                  * idle timer unplug to continue working.
3971                  */
3972                 if (cfq_cfqq_wait_request(cfqq)) {
3973                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3974                             cfqd->busy_queues > 1) {
3975                                 cfq_del_timer(cfqd, cfqq);
3976                                 cfq_clear_cfqq_wait_request(cfqq);
3977                                 __blk_run_queue(cfqd->queue);
3978                         } else {
3979                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3980                                 cfq_mark_cfqq_must_dispatch(cfqq);
3981                         }
3982                 }
3983         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3984                 /*
3985                  * not the active queue - expire current slice if it is
3986                  * idle and has expired it's mean thinktime or this new queue
3987                  * has some old slice time left and is of higher priority or
3988                  * this new queue is RT and the current one is BE
3989                  */
3990                 cfq_preempt_queue(cfqd, cfqq);
3991                 __blk_run_queue(cfqd->queue);
3992         }
3993 }
3994
3995 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3996 {
3997         struct cfq_data *cfqd = q->elevator->elevator_data;
3998         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3999
4000         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4001         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4002
4003         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4004         list_add_tail(&rq->queuelist, &cfqq->fifo);
4005         cfq_add_rq_rb(rq);
4006         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4007                                  rq->cmd_flags);
4008         cfq_rq_enqueued(cfqd, cfqq, rq);
4009 }
4010
4011 /*
4012  * Update hw_tag based on peak queue depth over 50 samples under
4013  * sufficient load.
4014  */
4015 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4016 {
4017         struct cfq_queue *cfqq = cfqd->active_queue;
4018
4019         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4020                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4021
4022         if (cfqd->hw_tag == 1)
4023                 return;
4024
4025         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4026             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4027                 return;
4028
4029         /*
4030          * If active queue hasn't enough requests and can idle, cfq might not
4031          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4032          * case
4033          */
4034         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4035             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4036             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4037                 return;
4038
4039         if (cfqd->hw_tag_samples++ < 50)
4040                 return;
4041
4042         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4043                 cfqd->hw_tag = 1;
4044         else
4045                 cfqd->hw_tag = 0;
4046 }
4047
4048 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4049 {
4050         struct cfq_io_cq *cic = cfqd->active_cic;
4051
4052         /* If the queue already has requests, don't wait */
4053         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4054                 return false;
4055
4056         /* If there are other queues in the group, don't wait */
4057         if (cfqq->cfqg->nr_cfqq > 1)
4058                 return false;
4059
4060         /* the only queue in the group, but think time is big */
4061         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4062                 return false;
4063
4064         if (cfq_slice_used(cfqq))
4065                 return true;
4066
4067         /* if slice left is less than think time, wait busy */
4068         if (cic && sample_valid(cic->ttime.ttime_samples)
4069             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4070                 return true;
4071
4072         /*
4073          * If think times is less than a jiffy than ttime_mean=0 and above
4074          * will not be true. It might happen that slice has not expired yet
4075          * but will expire soon (4-5 ns) during select_queue(). To cover the
4076          * case where think time is less than a jiffy, mark the queue wait
4077          * busy if only 1 jiffy is left in the slice.
4078          */
4079         if (cfqq->slice_end - jiffies == 1)
4080                 return true;
4081
4082         return false;
4083 }
4084
4085 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4086 {
4087         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4088         struct cfq_data *cfqd = cfqq->cfqd;
4089         const int sync = rq_is_sync(rq);
4090         unsigned long now;
4091
4092         now = jiffies;
4093         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4094                      !!(rq->cmd_flags & REQ_NOIDLE));
4095
4096         cfq_update_hw_tag(cfqd);
4097
4098         WARN_ON(!cfqd->rq_in_driver);
4099         WARN_ON(!cfqq->dispatched);
4100         cfqd->rq_in_driver--;
4101         cfqq->dispatched--;
4102         (RQ_CFQG(rq))->dispatched--;
4103         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4104                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4105
4106         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4107
4108         if (sync) {
4109                 struct cfq_rb_root *st;
4110
4111                 RQ_CIC(rq)->ttime.last_end_request = now;
4112
4113                 if (cfq_cfqq_on_rr(cfqq))
4114                         st = cfqq->service_tree;
4115                 else
4116                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4117                                         cfqq_type(cfqq));
4118
4119                 st->ttime.last_end_request = now;
4120                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4121                         cfqd->last_delayed_sync = now;
4122         }
4123
4124 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4125         cfqq->cfqg->ttime.last_end_request = now;
4126 #endif
4127
4128         /*
4129          * If this is the active queue, check if it needs to be expired,
4130          * or if we want to idle in case it has no pending requests.
4131          */
4132         if (cfqd->active_queue == cfqq) {
4133                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4134
4135                 if (cfq_cfqq_slice_new(cfqq)) {
4136                         cfq_set_prio_slice(cfqd, cfqq);
4137                         cfq_clear_cfqq_slice_new(cfqq);
4138                 }
4139
4140                 /*
4141                  * Should we wait for next request to come in before we expire
4142                  * the queue.
4143                  */
4144                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4145                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4146                         if (!cfqd->cfq_slice_idle)
4147                                 extend_sl = cfqd->cfq_group_idle;
4148                         cfqq->slice_end = jiffies + extend_sl;
4149                         cfq_mark_cfqq_wait_busy(cfqq);
4150                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4151                 }
4152
4153                 /*
4154                  * Idling is not enabled on:
4155                  * - expired queues
4156                  * - idle-priority queues
4157                  * - async queues
4158                  * - queues with still some requests queued
4159                  * - when there is a close cooperator
4160                  */
4161                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4162                         cfq_slice_expired(cfqd, 1);
4163                 else if (sync && cfqq_empty &&
4164                          !cfq_close_cooperator(cfqd, cfqq)) {
4165                         cfq_arm_slice_timer(cfqd);
4166                 }
4167         }
4168
4169         if (!cfqd->rq_in_driver)
4170                 cfq_schedule_dispatch(cfqd);
4171 }
4172
4173 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4174 {
4175         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4176                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4177                 return ELV_MQUEUE_MUST;
4178         }
4179
4180         return ELV_MQUEUE_MAY;
4181 }
4182
4183 static int cfq_may_queue(struct request_queue *q, int rw)
4184 {
4185         struct cfq_data *cfqd = q->elevator->elevator_data;
4186         struct task_struct *tsk = current;
4187         struct cfq_io_cq *cic;
4188         struct cfq_queue *cfqq;
4189
4190         /*
4191          * don't force setup of a queue from here, as a call to may_queue
4192          * does not necessarily imply that a request actually will be queued.
4193          * so just lookup a possibly existing queue, or return 'may queue'
4194          * if that fails
4195          */
4196         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4197         if (!cic)
4198                 return ELV_MQUEUE_MAY;
4199
4200         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4201         if (cfqq) {
4202                 cfq_init_prio_data(cfqq, cic);
4203
4204                 return __cfq_may_queue(cfqq);
4205         }
4206
4207         return ELV_MQUEUE_MAY;
4208 }
4209
4210 /*
4211  * queue lock held here
4212  */
4213 static void cfq_put_request(struct request *rq)
4214 {
4215         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4216
4217         if (cfqq) {
4218                 const int rw = rq_data_dir(rq);
4219
4220                 BUG_ON(!cfqq->allocated[rw]);
4221                 cfqq->allocated[rw]--;
4222
4223                 /* Put down rq reference on cfqg */
4224                 cfqg_put(RQ_CFQG(rq));
4225                 rq->elv.priv[0] = NULL;
4226                 rq->elv.priv[1] = NULL;
4227
4228                 cfq_put_queue(cfqq);
4229         }
4230 }
4231
4232 static struct cfq_queue *
4233 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4234                 struct cfq_queue *cfqq)
4235 {
4236         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4237         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4238         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4239         cfq_put_queue(cfqq);
4240         return cic_to_cfqq(cic, 1);
4241 }
4242
4243 /*
4244  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4245  * was the last process referring to said cfqq.
4246  */
4247 static struct cfq_queue *
4248 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4249 {
4250         if (cfqq_process_refs(cfqq) == 1) {
4251                 cfqq->pid = current->pid;
4252                 cfq_clear_cfqq_coop(cfqq);
4253                 cfq_clear_cfqq_split_coop(cfqq);
4254                 return cfqq;
4255         }
4256
4257         cic_set_cfqq(cic, NULL, 1);
4258
4259         cfq_put_cooperator(cfqq);
4260
4261         cfq_put_queue(cfqq);
4262         return NULL;
4263 }
4264 /*
4265  * Allocate cfq data structures associated with this request.
4266  */
4267 static int
4268 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4269                 gfp_t gfp_mask)
4270 {
4271         struct cfq_data *cfqd = q->elevator->elevator_data;
4272         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4273         const int rw = rq_data_dir(rq);
4274         const bool is_sync = rq_is_sync(rq);
4275         struct cfq_queue *cfqq;
4276
4277         spin_lock_irq(q->queue_lock);
4278
4279         check_ioprio_changed(cic, bio);
4280         check_blkcg_changed(cic, bio);
4281 new_queue:
4282         cfqq = cic_to_cfqq(cic, is_sync);
4283         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4284                 if (cfqq)
4285                         cfq_put_queue(cfqq);
4286                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4287                 cic_set_cfqq(cic, cfqq, is_sync);
4288         } else {
4289                 /*
4290                  * If the queue was seeky for too long, break it apart.
4291                  */
4292                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4293                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4294                         cfqq = split_cfqq(cic, cfqq);
4295                         if (!cfqq)
4296                                 goto new_queue;
4297                 }
4298
4299                 /*
4300                  * Check to see if this queue is scheduled to merge with
4301                  * another, closely cooperating queue.  The merging of
4302                  * queues happens here as it must be done in process context.
4303                  * The reference on new_cfqq was taken in merge_cfqqs.
4304                  */
4305                 if (cfqq->new_cfqq)
4306                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4307         }
4308
4309         cfqq->allocated[rw]++;
4310
4311         cfqq->ref++;
4312         cfqg_get(cfqq->cfqg);
4313         rq->elv.priv[0] = cfqq;
4314         rq->elv.priv[1] = cfqq->cfqg;
4315         spin_unlock_irq(q->queue_lock);
4316         return 0;
4317 }
4318
4319 static void cfq_kick_queue(struct work_struct *work)
4320 {
4321         struct cfq_data *cfqd =
4322                 container_of(work, struct cfq_data, unplug_work);
4323         struct request_queue *q = cfqd->queue;
4324
4325         spin_lock_irq(q->queue_lock);
4326         __blk_run_queue(cfqd->queue);
4327         spin_unlock_irq(q->queue_lock);
4328 }
4329
4330 /*
4331  * Timer running if the active_queue is currently idling inside its time slice
4332  */
4333 static void cfq_idle_slice_timer(unsigned long data)
4334 {
4335         struct cfq_data *cfqd = (struct cfq_data *) data;
4336         struct cfq_queue *cfqq;
4337         unsigned long flags;
4338         int timed_out = 1;
4339
4340         cfq_log(cfqd, "idle timer fired");
4341
4342         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4343
4344         cfqq = cfqd->active_queue;
4345         if (cfqq) {
4346                 timed_out = 0;
4347
4348                 /*
4349                  * We saw a request before the queue expired, let it through
4350                  */
4351                 if (cfq_cfqq_must_dispatch(cfqq))
4352                         goto out_kick;
4353
4354                 /*
4355                  * expired
4356                  */
4357                 if (cfq_slice_used(cfqq))
4358                         goto expire;
4359
4360                 /*
4361                  * only expire and reinvoke request handler, if there are
4362                  * other queues with pending requests
4363                  */
4364                 if (!cfqd->busy_queues)
4365                         goto out_cont;
4366
4367                 /*
4368                  * not expired and it has a request pending, let it dispatch
4369                  */
4370                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4371                         goto out_kick;
4372
4373                 /*
4374                  * Queue depth flag is reset only when the idle didn't succeed
4375                  */
4376                 cfq_clear_cfqq_deep(cfqq);
4377         }
4378 expire:
4379         cfq_slice_expired(cfqd, timed_out);
4380 out_kick:
4381         cfq_schedule_dispatch(cfqd);
4382 out_cont:
4383         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4384 }
4385
4386 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4387 {
4388         del_timer_sync(&cfqd->idle_slice_timer);
4389         cancel_work_sync(&cfqd->unplug_work);
4390 }
4391
4392 static void cfq_exit_queue(struct elevator_queue *e)
4393 {
4394         struct cfq_data *cfqd = e->elevator_data;
4395         struct request_queue *q = cfqd->queue;
4396
4397         cfq_shutdown_timer_wq(cfqd);
4398
4399         spin_lock_irq(q->queue_lock);
4400
4401         if (cfqd->active_queue)
4402                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4403
4404         spin_unlock_irq(q->queue_lock);
4405
4406         cfq_shutdown_timer_wq(cfqd);
4407
4408 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4409         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4410 #else
4411         kfree(cfqd->root_group);
4412 #endif
4413         kfree(cfqd);
4414 }
4415
4416 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4417 {
4418         struct cfq_data *cfqd;
4419         struct blkcg_gq *blkg __maybe_unused;
4420         int i, ret;
4421         struct elevator_queue *eq;
4422
4423         eq = elevator_alloc(q, e);
4424         if (!eq)
4425                 return -ENOMEM;
4426
4427         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4428         if (!cfqd) {
4429                 kobject_put(&eq->kobj);
4430                 return -ENOMEM;
4431         }
4432         eq->elevator_data = cfqd;
4433
4434         cfqd->queue = q;
4435         spin_lock_irq(q->queue_lock);
4436         q->elevator = eq;
4437         spin_unlock_irq(q->queue_lock);
4438
4439         /* Init root service tree */
4440         cfqd->grp_service_tree = CFQ_RB_ROOT;
4441
4442         /* Init root group and prefer root group over other groups by default */
4443 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4444         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4445         if (ret)
4446                 goto out_free;
4447
4448         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4449 #else
4450         ret = -ENOMEM;
4451         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4452                                         GFP_KERNEL, cfqd->queue->node);
4453         if (!cfqd->root_group)
4454                 goto out_free;
4455
4456         cfq_init_cfqg_base(cfqd->root_group);
4457 #endif
4458         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4459         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4460
4461         /*
4462          * Not strictly needed (since RB_ROOT just clears the node and we
4463          * zeroed cfqd on alloc), but better be safe in case someone decides
4464          * to add magic to the rb code
4465          */
4466         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4467                 cfqd->prio_trees[i] = RB_ROOT;
4468
4469         /*
4470          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4471          * Grab a permanent reference to it, so that the normal code flow
4472          * will not attempt to free it.  oom_cfqq is linked to root_group
4473          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4474          * the reference from linking right away.
4475          */
4476         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4477         cfqd->oom_cfqq.ref++;
4478
4479         spin_lock_irq(q->queue_lock);
4480         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4481         cfqg_put(cfqd->root_group);
4482         spin_unlock_irq(q->queue_lock);
4483
4484         init_timer(&cfqd->idle_slice_timer);
4485         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4486         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4487
4488         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4489
4490         cfqd->cfq_quantum = cfq_quantum;
4491         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4492         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4493         cfqd->cfq_back_max = cfq_back_max;
4494         cfqd->cfq_back_penalty = cfq_back_penalty;
4495         cfqd->cfq_slice[0] = cfq_slice_async;
4496         cfqd->cfq_slice[1] = cfq_slice_sync;
4497         cfqd->cfq_target_latency = cfq_target_latency;
4498         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4499         cfqd->cfq_slice_idle = cfq_slice_idle;
4500         cfqd->cfq_group_idle = cfq_group_idle;
4501         cfqd->cfq_latency = 1;
4502         cfqd->hw_tag = -1;
4503         /*
4504          * we optimistically start assuming sync ops weren't delayed in last
4505          * second, in order to have larger depth for async operations.
4506          */
4507         cfqd->last_delayed_sync = jiffies - HZ;
4508         return 0;
4509
4510 out_free:
4511         kfree(cfqd);
4512         kobject_put(&eq->kobj);
4513         return ret;
4514 }
4515
4516 static void cfq_registered_queue(struct request_queue *q)
4517 {
4518         struct elevator_queue *e = q->elevator;
4519         struct cfq_data *cfqd = e->elevator_data;
4520
4521         /*
4522          * Default to IOPS mode with no idling for SSDs
4523          */
4524         if (blk_queue_nonrot(q))
4525                 cfqd->cfq_slice_idle = 0;
4526 }
4527
4528 /*
4529  * sysfs parts below -->
4530  */
4531 static ssize_t
4532 cfq_var_show(unsigned int var, char *page)
4533 {
4534         return sprintf(page, "%u\n", var);
4535 }
4536
4537 static ssize_t
4538 cfq_var_store(unsigned int *var, const char *page, size_t count)
4539 {
4540         char *p = (char *) page;
4541
4542         *var = simple_strtoul(p, &p, 10);
4543         return count;
4544 }
4545
4546 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4547 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4548 {                                                                       \
4549         struct cfq_data *cfqd = e->elevator_data;                       \
4550         unsigned int __data = __VAR;                                    \
4551         if (__CONV)                                                     \
4552                 __data = jiffies_to_msecs(__data);                      \
4553         return cfq_var_show(__data, (page));                            \
4554 }
4555 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4556 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4557 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4558 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4559 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4560 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4561 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4562 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4563 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4564 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4565 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4566 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4567 #undef SHOW_FUNCTION
4568
4569 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4570 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4571 {                                                                       \
4572         struct cfq_data *cfqd = e->elevator_data;                       \
4573         unsigned int __data;                                            \
4574         int ret = cfq_var_store(&__data, (page), count);                \
4575         if (__data < (MIN))                                             \
4576                 __data = (MIN);                                         \
4577         else if (__data > (MAX))                                        \
4578                 __data = (MAX);                                         \
4579         if (__CONV)                                                     \
4580                 *(__PTR) = msecs_to_jiffies(__data);                    \
4581         else                                                            \
4582                 *(__PTR) = __data;                                      \
4583         return ret;                                                     \
4584 }
4585 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4586 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4587                 UINT_MAX, 1);
4588 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4589                 UINT_MAX, 1);
4590 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4591 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4592                 UINT_MAX, 0);
4593 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4594 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4595 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4596 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4597 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4598                 UINT_MAX, 0);
4599 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4600 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4601 #undef STORE_FUNCTION
4602
4603 #define CFQ_ATTR(name) \
4604         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4605
4606 static struct elv_fs_entry cfq_attrs[] = {
4607         CFQ_ATTR(quantum),
4608         CFQ_ATTR(fifo_expire_sync),
4609         CFQ_ATTR(fifo_expire_async),
4610         CFQ_ATTR(back_seek_max),
4611         CFQ_ATTR(back_seek_penalty),
4612         CFQ_ATTR(slice_sync),
4613         CFQ_ATTR(slice_async),
4614         CFQ_ATTR(slice_async_rq),
4615         CFQ_ATTR(slice_idle),
4616         CFQ_ATTR(group_idle),
4617         CFQ_ATTR(low_latency),
4618         CFQ_ATTR(target_latency),
4619         __ATTR_NULL
4620 };
4621
4622 static struct elevator_type iosched_cfq = {
4623         .ops = {
4624                 .elevator_merge_fn =            cfq_merge,
4625                 .elevator_merged_fn =           cfq_merged_request,
4626                 .elevator_merge_req_fn =        cfq_merged_requests,
4627                 .elevator_allow_merge_fn =      cfq_allow_merge,
4628                 .elevator_bio_merged_fn =       cfq_bio_merged,
4629                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4630                 .elevator_add_req_fn =          cfq_insert_request,
4631                 .elevator_activate_req_fn =     cfq_activate_request,
4632                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4633                 .elevator_completed_req_fn =    cfq_completed_request,
4634                 .elevator_former_req_fn =       elv_rb_former_request,
4635                 .elevator_latter_req_fn =       elv_rb_latter_request,
4636                 .elevator_init_icq_fn =         cfq_init_icq,
4637                 .elevator_exit_icq_fn =         cfq_exit_icq,
4638                 .elevator_set_req_fn =          cfq_set_request,
4639                 .elevator_put_req_fn =          cfq_put_request,
4640                 .elevator_may_queue_fn =        cfq_may_queue,
4641                 .elevator_init_fn =             cfq_init_queue,
4642                 .elevator_exit_fn =             cfq_exit_queue,
4643                 .elevator_registered_fn =       cfq_registered_queue,
4644         },
4645         .icq_size       =       sizeof(struct cfq_io_cq),
4646         .icq_align      =       __alignof__(struct cfq_io_cq),
4647         .elevator_attrs =       cfq_attrs,
4648         .elevator_name  =       "cfq",
4649         .elevator_owner =       THIS_MODULE,
4650 };
4651
4652 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4653 static struct blkcg_policy blkcg_policy_cfq = {
4654         .cftypes                = cfq_blkcg_files,
4655
4656         .cpd_alloc_fn           = cfq_cpd_alloc,
4657         .cpd_init_fn            = cfq_cpd_init,
4658         .cpd_free_fn            = cfq_cpd_free,
4659
4660         .pd_alloc_fn            = cfq_pd_alloc,
4661         .pd_init_fn             = cfq_pd_init,
4662         .pd_offline_fn          = cfq_pd_offline,
4663         .pd_free_fn             = cfq_pd_free,
4664         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4665 };
4666 #endif
4667
4668 static int __init cfq_init(void)
4669 {
4670         int ret;
4671
4672         /*
4673          * could be 0 on HZ < 1000 setups
4674          */
4675         if (!cfq_slice_async)
4676                 cfq_slice_async = 1;
4677         if (!cfq_slice_idle)
4678                 cfq_slice_idle = 1;
4679
4680 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4681         if (!cfq_group_idle)
4682                 cfq_group_idle = 1;
4683
4684         ret = blkcg_policy_register(&blkcg_policy_cfq);
4685         if (ret)
4686                 return ret;
4687 #else
4688         cfq_group_idle = 0;
4689 #endif
4690
4691         ret = -ENOMEM;
4692         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4693         if (!cfq_pool)
4694                 goto err_pol_unreg;
4695
4696         ret = elv_register(&iosched_cfq);
4697         if (ret)
4698                 goto err_free_pool;
4699
4700         return 0;
4701
4702 err_free_pool:
4703         kmem_cache_destroy(cfq_pool);
4704 err_pol_unreg:
4705 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4706         blkcg_policy_unregister(&blkcg_policy_cfq);
4707 #endif
4708         return ret;
4709 }
4710
4711 static void __exit cfq_exit(void)
4712 {
4713 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4714         blkcg_policy_unregister(&blkcg_policy_cfq);
4715 #endif
4716         elv_unregister(&iosched_cfq);
4717         kmem_cache_destroy(cfq_pool);
4718 }
4719
4720 module_init(cfq_init);
4721 module_exit(cfq_exit);
4722
4723 MODULE_AUTHOR("Jens Axboe");
4724 MODULE_LICENSE("GPL");
4725 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");