block/cfq-iosched.c

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