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