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 /*
  21  * tunables
  22  */
  23 /* max queue in one round of service */
  24 static const int cfq_quantum = 8;
  25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  26 /* maximum backwards seek, in KiB */
  27 static const int cfq_back_max = 16 * 1024;
  28 /* penalty of a backwards seek */
  29 static const int cfq_back_penalty = 2;
  30 static const int cfq_slice_sync = HZ / 10;
  31 static int cfq_slice_async = HZ / 25;
  32 static const int cfq_slice_async_rq = 2;
  33 static int cfq_slice_idle = HZ / 125;
  34 static int cfq_group_idle = HZ / 125;
  35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
  36 static const int cfq_hist_divisor = 4;
  37
  38 /*
  39  * offset from end of service tree
  40  */
  41 #define CFQ_IDLE_DELAY          (HZ / 5)
  42
  43 /*
  44  * below this threshold, we consider thinktime immediate
  45  */
  46 #define CFQ_MIN_TT              (2)
  47
  48 #define CFQ_SLICE_SCALE         (5)
  49 #define CFQ_HW_QUEUE_MIN        (5)
  50 #define CFQ_SERVICE_SHIFT       12
  51
  52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  56
  57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  60
  61 static struct kmem_cache *cfq_pool;
  62
  63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  69
  70 struct cfq_ttime {
  71         unsigned long last_end_request;
  72
  73         unsigned long ttime_total;
  74         unsigned long ttime_samples;
  75         unsigned long ttime_mean;
  76 };
  77
  78 /*
  79  * Most of our rbtree usage is for sorting with min extraction, so
  80  * if we cache the leftmost node we don't have to walk down the tree
  81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  82  * move this into the elevator for the rq sorting as well.
  83  */
  84 struct cfq_rb_root {
  85         struct rb_root rb;
  86         struct rb_node *left;
  87         unsigned count;
  88         unsigned total_weight;
  89         u64 min_vdisktime;
  90         struct cfq_ttime ttime;
  91 };
  92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  93                         .ttime = {.last_end_request = jiffies,},}
  94
  95 /*
  96  * Per process-grouping structure
  97  */
  98 struct cfq_queue {
  99         /* reference count */
 100         int ref;
 101         /* various state flags, see below */
 102         unsigned int flags;
 103         /* parent cfq_data */
 104         struct cfq_data *cfqd;
 105         /* service_tree member */
 106         struct rb_node rb_node;
 107         /* service_tree key */
 108         unsigned long rb_key;
 109         /* prio tree member */
 110         struct rb_node p_node;
 111         /* prio tree root we belong to, if any */
 112         struct rb_root *p_root;
 113         /* sorted list of pending requests */
 114         struct rb_root sort_list;
 115         /* if fifo isn't expired, next request to serve */
 116         struct request *next_rq;
 117         /* requests queued in sort_list */
 118         int queued[2];
 119         /* currently allocated requests */
 120         int allocated[2];
 121         /* fifo list of requests in sort_list */
 122         struct list_head fifo;
 123
 124         /* time when queue got scheduled in to dispatch first request. */
 125         unsigned long dispatch_start;
 126         unsigned int allocated_slice;
 127         unsigned int slice_dispatch;
 128         /* time when first request from queue completed and slice started. */
 129         unsigned long slice_start;
 130         unsigned long slice_end;
 131         long slice_resid;
 132
 133         /* pending priority requests */
 134         int prio_pending;
 135         /* number of requests that are on the dispatch list or inside driver */
 136         int dispatched;
 137
 138         /* io prio of this group */
 139         unsigned short ioprio, org_ioprio;
 140         unsigned short ioprio_class;
 141
 142         pid_t pid;
 143
 144         u32 seek_history;
 145         sector_t last_request_pos;
 146
 147         struct cfq_rb_root *service_tree;
 148         struct cfq_queue *new_cfqq;
 149         struct cfq_group *cfqg;
 150         /* Number of sectors dispatched from queue in single dispatch round */
 151         unsigned long nr_sectors;
 152 };
 153
 154 /*
 155  * First index in the service_trees.
 156  * IDLE is handled separately, so it has negative index
 157  */
 158 enum wl_class_t {
 159         BE_WORKLOAD = 0,
 160         RT_WORKLOAD = 1,
 161         IDLE_WORKLOAD = 2,
 162         CFQ_PRIO_NR,
 163 };
 164
 165 /*
 166  * Second index in the service_trees.
 167  */
 168 enum wl_type_t {
 169         ASYNC_WORKLOAD = 0,
 170         SYNC_NOIDLE_WORKLOAD = 1,
 171         SYNC_WORKLOAD = 2
 172 };
 173
 174 struct cfqg_stats {
 175 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 176         /* total bytes transferred */
 177         struct blkg_rwstat              service_bytes;
 178         /* total IOs serviced, post merge */
 179         struct blkg_rwstat              serviced;
 180         /* number of ios merged */
 181         struct blkg_rwstat              merged;
 182         /* total time spent on device in ns, may not be accurate w/ queueing */
 183         struct blkg_rwstat              service_time;
 184         /* total time spent waiting in scheduler queue in ns */
 185         struct blkg_rwstat              wait_time;
 186         /* number of IOs queued up */
 187         struct blkg_rwstat              queued;
 188         /* total sectors transferred */
 189         struct blkg_stat                sectors;
 190         /* total disk time and nr sectors dispatched by this group */
 191         struct blkg_stat                time;
 192 #ifdef CONFIG_DEBUG_BLK_CGROUP
 193         /* time not charged to this cgroup */
 194         struct blkg_stat                unaccounted_time;
 195         /* sum of number of ios queued across all samples */
 196         struct blkg_stat                avg_queue_size_sum;
 197         /* count of samples taken for average */
 198         struct blkg_stat                avg_queue_size_samples;
 199         /* how many times this group has been removed from service tree */
 200         struct blkg_stat                dequeue;
 201         /* total time spent waiting for it to be assigned a timeslice. */
 202         struct blkg_stat                group_wait_time;
 203         /* time spent idling for this blkcg_gq */
 204         struct blkg_stat                idle_time;
 205         /* total time with empty current active q with other requests queued */
 206         struct blkg_stat                empty_time;
 207         /* fields after this shouldn't be cleared on stat reset */
 208         uint64_t                        start_group_wait_time;
 209         uint64_t                        start_idle_time;
 210         uint64_t                        start_empty_time;
 211         uint16_t                        flags;
 212 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 213 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 214 };
 215
 216 /* This is per cgroup per device grouping structure */
 217 struct cfq_group {
 218         /* must be the first member */
 219         struct blkg_policy_data pd;
 220
 221         /* group service_tree member */
 222         struct rb_node rb_node;
 223
 224         /* group service_tree key */
 225         u64 vdisktime;
 226         unsigned int weight;
 227         unsigned int new_weight;
 228         unsigned int dev_weight;
 229
 230         /* number of cfqq currently on this group */
 231         int nr_cfqq;
 232
 233         /*
 234          * Per group busy queues average. Useful for workload slice calc. We
 235          * create the array for each prio class but at run time it is used
 236          * only for RT and BE class and slot for IDLE class remains unused.
 237          * This is primarily done to avoid confusion and a gcc warning.
 238          */
 239         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 240         /*
 241          * rr lists of queues with requests. We maintain service trees for
 242          * RT and BE classes. These trees are subdivided in subclasses
 243          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 244          * class there is no subclassification and all the cfq queues go on
 245          * a single tree service_tree_idle.
 246          * Counts are embedded in the cfq_rb_root
 247          */
 248         struct cfq_rb_root service_trees[2][3];
 249         struct cfq_rb_root service_tree_idle;
 250
 251         unsigned long saved_wl_slice;
 252         enum wl_type_t saved_wl_type;
 253         enum wl_class_t saved_wl_class;
 254
 255         /* number of requests that are on the dispatch list or inside driver */
 256         int dispatched;
 257         struct cfq_ttime ttime;
 258         struct cfqg_stats stats;
 259 };
 260
 261 struct cfq_io_cq {
 262         struct io_cq            icq;            /* must be the first member */
 263         struct cfq_queue        *cfqq[2];
 264         struct cfq_ttime        ttime;
 265         int                     ioprio;         /* the current ioprio */
 266 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 267         uint64_t                blkcg_id;       /* the current blkcg ID */
 268 #endif
 269 };
 270
 271 /*
 272  * Per block device queue structure
 273  */
 274 struct cfq_data {
 275         struct request_queue *queue;
 276         /* Root service tree for cfq_groups */
 277         struct cfq_rb_root grp_service_tree;
 278         struct cfq_group *root_group;
 279
 280         /*
 281          * The priority currently being served
 282          */
 283         enum wl_class_t serving_wl_class;
 284         enum wl_type_t serving_wl_type;
 285         unsigned long workload_expires;
 286         struct cfq_group *serving_group;
 287
 288         /*
 289          * Each priority tree is sorted by next_request position.  These
 290          * trees are used when determining if two or more queues are
 291          * interleaving requests (see cfq_close_cooperator).
 292          */
 293         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 294
 295         unsigned int busy_queues;
 296         unsigned int busy_sync_queues;
 297
 298         int rq_in_driver;
 299         int rq_in_flight[2];
 300
 301         /*
 302          * queue-depth detection
 303          */
 304         int rq_queued;
 305         int hw_tag;
 306         /*
 307          * hw_tag can be
 308          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 309          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 310          *  0 => no NCQ
 311          */
 312         int hw_tag_est_depth;
 313         unsigned int hw_tag_samples;
 314
 315         /*
 316          * idle window management
 317          */
 318         struct timer_list idle_slice_timer;
 319         struct work_struct unplug_work;
 320
 321         struct cfq_queue *active_queue;
 322         struct cfq_io_cq *active_cic;
 323
 324         /*
 325          * async queue for each priority case
 326          */
 327         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 328         struct cfq_queue *async_idle_cfqq;
 329
 330         sector_t last_position;
 331
 332         /*
 333          * tunables, see top of file
 334          */
 335         unsigned int cfq_quantum;
 336         unsigned int cfq_fifo_expire[2];
 337         unsigned int cfq_back_penalty;
 338         unsigned int cfq_back_max;
 339         unsigned int cfq_slice[2];
 340         unsigned int cfq_slice_async_rq;
 341         unsigned int cfq_slice_idle;
 342         unsigned int cfq_group_idle;
 343         unsigned int cfq_latency;
 344         unsigned int cfq_target_latency;
 345
 346         /*
 347          * Fallback dummy cfqq for extreme OOM conditions
 348          */
 349         struct cfq_queue oom_cfqq;
 350
 351         unsigned long last_delayed_sync;
 352 };
 353
 354 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 355
 356 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 357                                             enum wl_class_t class,
 358                                             enum wl_type_t type)
 359 {
 360         if (!cfqg)
 361                 return NULL;
 362
 363         if (class == IDLE_WORKLOAD)
 364                 return &cfqg->service_tree_idle;
 365
 366         return &cfqg->service_trees[class][type];
 367 }
 368
 369 enum cfqq_state_flags {
 370         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 371         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 372         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 373         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 374         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 375         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 376         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 377         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 378         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 379         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 380         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 381         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 382         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 383 };
 384
 385 #define CFQ_CFQQ_FNS(name)                                              \
 386 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 387 {                                                                       \
 388         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 389 }                                                                       \
 390 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 391 {                                                                       \
 392         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 393 }                                                                       \
 394 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 395 {                                                                       \
 396         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 397 }
 398
 399 CFQ_CFQQ_FNS(on_rr);
 400 CFQ_CFQQ_FNS(wait_request);
 401 CFQ_CFQQ_FNS(must_dispatch);
 402 CFQ_CFQQ_FNS(must_alloc_slice);
 403 CFQ_CFQQ_FNS(fifo_expire);
 404 CFQ_CFQQ_FNS(idle_window);
 405 CFQ_CFQQ_FNS(prio_changed);
 406 CFQ_CFQQ_FNS(slice_new);
 407 CFQ_CFQQ_FNS(sync);
 408 CFQ_CFQQ_FNS(coop);
 409 CFQ_CFQQ_FNS(split_coop);
 410 CFQ_CFQQ_FNS(deep);
 411 CFQ_CFQQ_FNS(wait_busy);
 412 #undef CFQ_CFQQ_FNS
 413
 414 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 415 {
 416         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 417 }
 418
 419 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 420 {
 421         return pd_to_blkg(&cfqg->pd);
 422 }
 423
 424 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 425
 426 /* cfqg stats flags */
 427 enum cfqg_stats_flags {
 428         CFQG_stats_waiting = 0,
 429         CFQG_stats_idling,
 430         CFQG_stats_empty,
 431 };
 432
 433 #define CFQG_FLAG_FNS(name)                                             \
 434 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 435 {                                                                       \
 436         stats->flags |= (1 << CFQG_stats_##name);                       \
 437 }                                                                       \
 438 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 439 {                                                                       \
 440         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 441 }                                                                       \
 442 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 443 {                                                                       \
 444         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 445 }                                                                       \
 446
 447 CFQG_FLAG_FNS(waiting)
 448 CFQG_FLAG_FNS(idling)
 449 CFQG_FLAG_FNS(empty)
 450 #undef CFQG_FLAG_FNS
 451
 452 /* This should be called with the queue_lock held. */
 453 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 454 {
 455         unsigned long long now;
 456
 457         if (!cfqg_stats_waiting(stats))
 458                 return;
 459
 460         now = sched_clock();
 461         if (time_after64(now, stats->start_group_wait_time))
 462                 blkg_stat_add(&stats->group_wait_time,
 463                               now - stats->start_group_wait_time);
 464         cfqg_stats_clear_waiting(stats);
 465 }
 466
 467 /* This should be called with the queue_lock held. */
 468 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 469                                                  struct cfq_group *curr_cfqg)
 470 {
 471         struct cfqg_stats *stats = &cfqg->stats;
 472
 473         if (cfqg_stats_waiting(stats))
 474                 return;
 475         if (cfqg == curr_cfqg)
 476                 return;
 477         stats->start_group_wait_time = sched_clock();
 478         cfqg_stats_mark_waiting(stats);
 479 }
 480
 481 /* This should be called with the queue_lock held. */
 482 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 483 {
 484         unsigned long long now;
 485
 486         if (!cfqg_stats_empty(stats))
 487                 return;
 488
 489         now = sched_clock();
 490         if (time_after64(now, stats->start_empty_time))
 491                 blkg_stat_add(&stats->empty_time,
 492                               now - stats->start_empty_time);
 493         cfqg_stats_clear_empty(stats);
 494 }
 495
 496 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 497 {
 498         blkg_stat_add(&cfqg->stats.dequeue, 1);
 499 }
 500
 501 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 502 {
 503         struct cfqg_stats *stats = &cfqg->stats;
 504
 505         if (blkg_rwstat_sum(&stats->queued))
 506                 return;
 507
 508         /*
 509          * group is already marked empty. This can happen if cfqq got new
 510          * request in parent group and moved to this group while being added
 511          * to service tree. Just ignore the event and move on.
 512          */
 513         if (cfqg_stats_empty(stats))
 514                 return;
 515
 516         stats->start_empty_time = sched_clock();
 517         cfqg_stats_mark_empty(stats);
 518 }
 519
 520 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 521 {
 522         struct cfqg_stats *stats = &cfqg->stats;
 523
 524         if (cfqg_stats_idling(stats)) {
 525                 unsigned long long now = sched_clock();
 526
 527                 if (time_after64(now, stats->start_idle_time))
 528                         blkg_stat_add(&stats->idle_time,
 529                                       now - stats->start_idle_time);
 530                 cfqg_stats_clear_idling(stats);
 531         }
 532 }
 533
 534 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 535 {
 536         struct cfqg_stats *stats = &cfqg->stats;
 537
 538         BUG_ON(cfqg_stats_idling(stats));
 539
 540         stats->start_idle_time = sched_clock();
 541         cfqg_stats_mark_idling(stats);
 542 }
 543
 544 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 545 {
 546         struct cfqg_stats *stats = &cfqg->stats;
 547
 548         blkg_stat_add(&stats->avg_queue_size_sum,
 549                       blkg_rwstat_sum(&stats->queued));
 550         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 551         cfqg_stats_update_group_wait_time(stats);
 552 }
 553
 554 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 555
 556 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 557 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 558 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 559 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 560 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 561 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 562 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 563
 564 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 565
 566 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 567
 568 static struct blkcg_policy blkcg_policy_cfq;
 569
 570 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 571 {
 572         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 573 }
 574
 575 static inline void cfqg_get(struct cfq_group *cfqg)
 576 {
 577         return blkg_get(cfqg_to_blkg(cfqg));
 578 }
 579
 580 static inline void cfqg_put(struct cfq_group *cfqg)
 581 {
 582         return blkg_put(cfqg_to_blkg(cfqg));
 583 }
 584
 585 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 586         char __pbuf[128];                                               \
 587                                                                         \
 588         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 589         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
 590                           cfq_cfqq_sync((cfqq)) ? 'S' : 'A',            \
 591                           __pbuf, ##args);                              \
 592 } while (0)
 593
 594 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 595         char __pbuf[128];                                               \
 596                                                                         \
 597         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 598         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 599 } while (0)
 600
 601 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 602                                             struct cfq_group *curr_cfqg, int rw)
 603 {
 604         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 605         cfqg_stats_end_empty_time(&cfqg->stats);
 606         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 607 }
 608
 609 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 610                         unsigned long time, unsigned long unaccounted_time)
 611 {
 612         blkg_stat_add(&cfqg->stats.time, time);
 613 #ifdef CONFIG_DEBUG_BLK_CGROUP
 614         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 615 #endif
 616 }
 617
 618 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 619 {
 620         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 621 }
 622
 623 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 624 {
 625         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 626 }
 627
 628 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 629                                               uint64_t bytes, int rw)
 630 {
 631         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
 632         blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
 633         blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
 634 }
 635
 636 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 637                         uint64_t start_time, uint64_t io_start_time, int rw)
 638 {
 639         struct cfqg_stats *stats = &cfqg->stats;
 640         unsigned long long now = sched_clock();
 641
 642         if (time_after64(now, io_start_time))
 643                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 644         if (time_after64(io_start_time, start_time))
 645                 blkg_rwstat_add(&stats->wait_time, rw,
 646                                 io_start_time - start_time);
 647 }
 648
 649 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
 650 {
 651         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
 652         struct cfqg_stats *stats = &cfqg->stats;
 653
 654         /* queued stats shouldn't be cleared */
 655         blkg_rwstat_reset(&stats->service_bytes);
 656         blkg_rwstat_reset(&stats->serviced);
 657         blkg_rwstat_reset(&stats->merged);
 658         blkg_rwstat_reset(&stats->service_time);
 659         blkg_rwstat_reset(&stats->wait_time);
 660         blkg_stat_reset(&stats->time);
 661 #ifdef CONFIG_DEBUG_BLK_CGROUP
 662         blkg_stat_reset(&stats->unaccounted_time);
 663         blkg_stat_reset(&stats->avg_queue_size_sum);
 664         blkg_stat_reset(&stats->avg_queue_size_samples);
 665         blkg_stat_reset(&stats->dequeue);
 666         blkg_stat_reset(&stats->group_wait_time);
 667         blkg_stat_reset(&stats->idle_time);
 668         blkg_stat_reset(&stats->empty_time);
 669 #endif
 670 }
 671
 672 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 673
 674 static inline void cfqg_get(struct cfq_group *cfqg) { }
 675 static inline void cfqg_put(struct cfq_group *cfqg) { }
 676
 677 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 678         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 679 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 680
 681 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 682                         struct cfq_group *curr_cfqg, int rw) { }
 683 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 684                         unsigned long time, unsigned long unaccounted_time) { }
 685 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 686 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 687 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 688                                               uint64_t bytes, int rw) { }
 689 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 690                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 691
 692 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 693
 694 #define cfq_log(cfqd, fmt, args...)     \
 695         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 696
 697 /* Traverses through cfq group service trees */
 698 #define for_each_cfqg_st(cfqg, i, j, st) \
 699         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 700                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 701                         : &cfqg->service_tree_idle; \
 702                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 703                         (i == IDLE_WORKLOAD && j == 0); \
 704                         j++, st = i < IDLE_WORKLOAD ? \
 705                         &cfqg->service_trees[i][j]: NULL) \
 706
 707 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 708         struct cfq_ttime *ttime, bool group_idle)
 709 {
 710         unsigned long slice;
 711         if (!sample_valid(ttime->ttime_samples))
 712                 return false;
 713         if (group_idle)
 714                 slice = cfqd->cfq_group_idle;
 715         else
 716                 slice = cfqd->cfq_slice_idle;
 717         return ttime->ttime_mean > slice;
 718 }
 719
 720 static inline bool iops_mode(struct cfq_data *cfqd)
 721 {
 722         /*
 723          * If we are not idling on queues and it is a NCQ drive, parallel
 724          * execution of requests is on and measuring time is not possible
 725          * in most of the cases until and unless we drive shallower queue
 726          * depths and that becomes a performance bottleneck. In such cases
 727          * switch to start providing fairness in terms of number of IOs.
 728          */
 729         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 730                 return true;
 731         else
 732                 return false;
 733 }
 734
 735 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 736 {
 737         if (cfq_class_idle(cfqq))
 738                 return IDLE_WORKLOAD;
 739         if (cfq_class_rt(cfqq))
 740                 return RT_WORKLOAD;
 741         return BE_WORKLOAD;
 742 }
 743
 744
 745 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 746 {
 747         if (!cfq_cfqq_sync(cfqq))
 748                 return ASYNC_WORKLOAD;
 749         if (!cfq_cfqq_idle_window(cfqq))
 750                 return SYNC_NOIDLE_WORKLOAD;
 751         return SYNC_WORKLOAD;
 752 }
 753
 754 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 755                                         struct cfq_data *cfqd,
 756                                         struct cfq_group *cfqg)
 757 {
 758         if (wl_class == IDLE_WORKLOAD)
 759                 return cfqg->service_tree_idle.count;
 760
 761         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 762                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 763                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 764 }
 765
 766 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 767                                         struct cfq_group *cfqg)
 768 {
 769         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 770                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 771 }
 772
 773 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 774 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 775                                        struct cfq_io_cq *cic, struct bio *bio,
 776                                        gfp_t gfp_mask);
 777
 778 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 779 {
 780         /* cic->icq is the first member, %NULL will convert to %NULL */
 781         return container_of(icq, struct cfq_io_cq, icq);
 782 }
 783
 784 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 785                                                struct io_context *ioc)
 786 {
 787         if (ioc)
 788                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 789         return NULL;
 790 }
 791
 792 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 793 {
 794         return cic->cfqq[is_sync];
 795 }
 796
 797 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 798                                 bool is_sync)
 799 {
 800         cic->cfqq[is_sync] = cfqq;
 801 }
 802
 803 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 804 {
 805         return cic->icq.q->elevator->elevator_data;
 806 }
 807
 808 /*
 809  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 810  * set (in which case it could also be direct WRITE).
 811  */
 812 static inline bool cfq_bio_sync(struct bio *bio)
 813 {
 814         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 815 }
 816
 817 /*
 818  * scheduler run of queue, if there are requests pending and no one in the
 819  * driver that will restart queueing
 820  */
 821 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 822 {
 823         if (cfqd->busy_queues) {
 824                 cfq_log(cfqd, "schedule dispatch");
 825                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 826         }
 827 }
 828
 829 /*
 830  * Scale schedule slice based on io priority. Use the sync time slice only
 831  * if a queue is marked sync and has sync io queued. A sync queue with async
 832  * io only, should not get full sync slice length.
 833  */
 834 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 835                                  unsigned short prio)
 836 {
 837         const int base_slice = cfqd->cfq_slice[sync];
 838
 839         WARN_ON(prio >= IOPRIO_BE_NR);
 840
 841         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 842 }
 843
 844 static inline int
 845 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 846 {
 847         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 848 }
 849
 850 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
 851 {
 852         u64 d = delta << CFQ_SERVICE_SHIFT;
 853
 854         d = d * CFQ_WEIGHT_DEFAULT;
 855         do_div(d, cfqg->weight);
 856         return d;
 857 }
 858
 859 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 860 {
 861         s64 delta = (s64)(vdisktime - min_vdisktime);
 862         if (delta > 0)
 863                 min_vdisktime = vdisktime;
 864
 865         return min_vdisktime;
 866 }
 867
 868 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 869 {
 870         s64 delta = (s64)(vdisktime - min_vdisktime);
 871         if (delta < 0)
 872                 min_vdisktime = vdisktime;
 873
 874         return min_vdisktime;
 875 }
 876
 877 static void update_min_vdisktime(struct cfq_rb_root *st)
 878 {
 879         struct cfq_group *cfqg;
 880
 881         if (st->left) {
 882                 cfqg = rb_entry_cfqg(st->left);
 883                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 884                                                   cfqg->vdisktime);
 885         }
 886 }
 887
 888 /*
 889  * get averaged number of queues of RT/BE priority.
 890  * average is updated, with a formula that gives more weight to higher numbers,
 891  * to quickly follows sudden increases and decrease slowly
 892  */
 893
 894 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
 895                                         struct cfq_group *cfqg, bool rt)
 896 {
 897         unsigned min_q, max_q;
 898         unsigned mult  = cfq_hist_divisor - 1;
 899         unsigned round = cfq_hist_divisor / 2;
 900         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
 901
 902         min_q = min(cfqg->busy_queues_avg[rt], busy);
 903         max_q = max(cfqg->busy_queues_avg[rt], busy);
 904         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
 905                 cfq_hist_divisor;
 906         return cfqg->busy_queues_avg[rt];
 907 }
 908
 909 static inline unsigned
 910 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
 911 {
 912         struct cfq_rb_root *st = &cfqd->grp_service_tree;
 913
 914         return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
 915 }
 916
 917 static inline unsigned
 918 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 919 {
 920         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
 921         if (cfqd->cfq_latency) {
 922                 /*
 923                  * interested queues (we consider only the ones with the same
 924                  * priority class in the cfq group)
 925                  */
 926                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
 927                                                 cfq_class_rt(cfqq));
 928                 unsigned sync_slice = cfqd->cfq_slice[1];
 929                 unsigned expect_latency = sync_slice * iq;
 930                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
 931
 932                 if (expect_latency > group_slice) {
 933                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
 934                         /* scale low_slice according to IO priority
 935                          * and sync vs async */
 936                         unsigned low_slice =
 937                                 min(slice, base_low_slice * slice / sync_slice);
 938                         /* the adapted slice value is scaled to fit all iqs
 939                          * into the target latency */
 940                         slice = max(slice * group_slice / expect_latency,
 941                                     low_slice);
 942                 }
 943         }
 944         return slice;
 945 }
 946
 947 static inline void
 948 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 949 {
 950         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
 951
 952         cfqq->slice_start = jiffies;
 953         cfqq->slice_end = jiffies + slice;
 954         cfqq->allocated_slice = slice;
 955         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 956 }
 957
 958 /*
 959  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 960  * isn't valid until the first request from the dispatch is activated
 961  * and the slice time set.
 962  */
 963 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
 964 {
 965         if (cfq_cfqq_slice_new(cfqq))
 966                 return false;
 967         if (time_before(jiffies, cfqq->slice_end))
 968                 return false;
 969
 970         return true;
 971 }
 972
 973 /*
 974  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 975  * We choose the request that is closest to the head right now. Distance
 976  * behind the head is penalized and only allowed to a certain extent.
 977  */
 978 static struct request *
 979 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
 980 {
 981         sector_t s1, s2, d1 = 0, d2 = 0;
 982         unsigned long back_max;
 983 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 984 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 985         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 986
 987         if (rq1 == NULL || rq1 == rq2)
 988                 return rq2;
 989         if (rq2 == NULL)
 990                 return rq1;
 991
 992         if (rq_is_sync(rq1) != rq_is_sync(rq2))
 993                 return rq_is_sync(rq1) ? rq1 : rq2;
 994
 995         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
 996                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
 997
 998         s1 = blk_rq_pos(rq1);
 999         s2 = blk_rq_pos(rq2);
1000
1001         /*
1002          * by definition, 1KiB is 2 sectors
1003          */
1004         back_max = cfqd->cfq_back_max * 2;
1005
1006         /*
1007          * Strict one way elevator _except_ in the case where we allow
1008          * short backward seeks which are biased as twice the cost of a
1009          * similar forward seek.
1010          */
1011         if (s1 >= last)
1012                 d1 = s1 - last;
1013         else if (s1 + back_max >= last)
1014                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1015         else
1016                 wrap |= CFQ_RQ1_WRAP;
1017
1018         if (s2 >= last)
1019                 d2 = s2 - last;
1020         else if (s2 + back_max >= last)
1021                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1022         else
1023                 wrap |= CFQ_RQ2_WRAP;
1024
1025         /* Found required data */
1026
1027         /*
1028          * By doing switch() on the bit mask "wrap" we avoid having to
1029          * check two variables for all permutations: --> faster!
1030          */
1031         switch (wrap) {
1032         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1033                 if (d1 < d2)
1034                         return rq1;
1035                 else if (d2 < d1)
1036                         return rq2;
1037                 else {
1038                         if (s1 >= s2)
1039                                 return rq1;
1040                         else
1041                                 return rq2;
1042                 }
1043
1044         case CFQ_RQ2_WRAP:
1045                 return rq1;
1046         case CFQ_RQ1_WRAP:
1047                 return rq2;
1048         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1049         default:
1050                 /*
1051                  * Since both rqs are wrapped,
1052                  * start with the one that's further behind head
1053                  * (--> only *one* back seek required),
1054                  * since back seek takes more time than forward.
1055                  */
1056                 if (s1 <= s2)
1057                         return rq1;
1058                 else
1059                         return rq2;
1060         }
1061 }
1062
1063 /*
1064  * The below is leftmost cache rbtree addon
1065  */
1066 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1067 {
1068         /* Service tree is empty */
1069         if (!root->count)
1070                 return NULL;
1071
1072         if (!root->left)
1073                 root->left = rb_first(&root->rb);
1074
1075         if (root->left)
1076                 return rb_entry(root->left, struct cfq_queue, rb_node);
1077
1078         return NULL;
1079 }
1080
1081 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1082 {
1083         if (!root->left)
1084                 root->left = rb_first(&root->rb);
1085
1086         if (root->left)
1087                 return rb_entry_cfqg(root->left);
1088
1089         return NULL;
1090 }
1091
1092 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1093 {
1094         rb_erase(n, root);
1095         RB_CLEAR_NODE(n);
1096 }
1097
1098 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1099 {
1100         if (root->left == n)
1101                 root->left = NULL;
1102         rb_erase_init(n, &root->rb);
1103         --root->count;
1104 }
1105
1106 /*
1107  * would be nice to take fifo expire time into account as well
1108  */
1109 static struct request *
1110 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1111                   struct request *last)
1112 {
1113         struct rb_node *rbnext = rb_next(&last->rb_node);
1114         struct rb_node *rbprev = rb_prev(&last->rb_node);
1115         struct request *next = NULL, *prev = NULL;
1116
1117         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1118
1119         if (rbprev)
1120                 prev = rb_entry_rq(rbprev);
1121
1122         if (rbnext)
1123                 next = rb_entry_rq(rbnext);
1124         else {
1125                 rbnext = rb_first(&cfqq->sort_list);
1126                 if (rbnext && rbnext != &last->rb_node)
1127                         next = rb_entry_rq(rbnext);
1128         }
1129
1130         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1131 }
1132
1133 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1134                                       struct cfq_queue *cfqq)
1135 {
1136         /*
1137          * just an approximation, should be ok.
1138          */
1139         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1140                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1141 }
1142
1143 static inline s64
1144 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1145 {
1146         return cfqg->vdisktime - st->min_vdisktime;
1147 }
1148
1149 static void
1150 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1151 {
1152         struct rb_node **node = &st->rb.rb_node;
1153         struct rb_node *parent = NULL;
1154         struct cfq_group *__cfqg;
1155         s64 key = cfqg_key(st, cfqg);
1156         int left = 1;
1157
1158         while (*node != NULL) {
1159                 parent = *node;
1160                 __cfqg = rb_entry_cfqg(parent);
1161
1162                 if (key < cfqg_key(st, __cfqg))
1163                         node = &parent->rb_left;
1164                 else {
1165                         node = &parent->rb_right;
1166                         left = 0;
1167                 }
1168         }
1169
1170         if (left)
1171                 st->left = &cfqg->rb_node;
1172
1173         rb_link_node(&cfqg->rb_node, parent, node);
1174         rb_insert_color(&cfqg->rb_node, &st->rb);
1175 }
1176
1177 static void
1178 cfq_update_group_weight(struct cfq_group *cfqg)
1179 {
1180         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1181         if (cfqg->new_weight) {
1182                 cfqg->weight = cfqg->new_weight;
1183                 cfqg->new_weight = 0;
1184         }
1185 }
1186
1187 static void
1188 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1189 {
1190         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1191
1192         cfq_update_group_weight(cfqg);
1193         __cfq_group_service_tree_add(st, cfqg);
1194         st->total_weight += cfqg->weight;
1195 }
1196
1197 static void
1198 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1199 {
1200         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1201         struct cfq_group *__cfqg;
1202         struct rb_node *n;
1203
1204         cfqg->nr_cfqq++;
1205         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1206                 return;
1207
1208         /*
1209          * Currently put the group at the end. Later implement something
1210          * so that groups get lesser vtime based on their weights, so that
1211          * if group does not loose all if it was not continuously backlogged.
1212          */
1213         n = rb_last(&st->rb);
1214         if (n) {
1215                 __cfqg = rb_entry_cfqg(n);
1216                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1217         } else
1218                 cfqg->vdisktime = st->min_vdisktime;
1219         cfq_group_service_tree_add(st, cfqg);
1220 }
1221
1222 static void
1223 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1224 {
1225         st->total_weight -= cfqg->weight;
1226         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1227                 cfq_rb_erase(&cfqg->rb_node, st);
1228 }
1229
1230 static void
1231 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1232 {
1233         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1234
1235         BUG_ON(cfqg->nr_cfqq < 1);
1236         cfqg->nr_cfqq--;
1237
1238         /* If there are other cfq queues under this group, don't delete it */
1239         if (cfqg->nr_cfqq)
1240                 return;
1241
1242         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1243         cfq_group_service_tree_del(st, cfqg);
1244         cfqg->saved_wl_slice = 0;
1245         cfqg_stats_update_dequeue(cfqg);
1246 }
1247
1248 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1249                                                 unsigned int *unaccounted_time)
1250 {
1251         unsigned int slice_used;
1252
1253         /*
1254          * Queue got expired before even a single request completed or
1255          * got expired immediately after first request completion.
1256          */
1257         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1258                 /*
1259                  * Also charge the seek time incurred to the group, otherwise
1260                  * if there are mutiple queues in the group, each can dispatch
1261                  * a single request on seeky media and cause lots of seek time
1262                  * and group will never know it.
1263                  */
1264                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1265                                         1);
1266         } else {
1267                 slice_used = jiffies - cfqq->slice_start;
1268                 if (slice_used > cfqq->allocated_slice) {
1269                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1270                         slice_used = cfqq->allocated_slice;
1271                 }
1272                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1273                         *unaccounted_time += cfqq->slice_start -
1274                                         cfqq->dispatch_start;
1275         }
1276
1277         return slice_used;
1278 }
1279
1280 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1281                                 struct cfq_queue *cfqq)
1282 {
1283         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1284         unsigned int used_sl, charge, unaccounted_sl = 0;
1285         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1286                         - cfqg->service_tree_idle.count;
1287
1288         BUG_ON(nr_sync < 0);
1289         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1290
1291         if (iops_mode(cfqd))
1292                 charge = cfqq->slice_dispatch;
1293         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1294                 charge = cfqq->allocated_slice;
1295
1296         /* Can't update vdisktime while group is on service tree */
1297         cfq_group_service_tree_del(st, cfqg);
1298         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
1299         /* If a new weight was requested, update now, off tree */
1300         cfq_group_service_tree_add(st, cfqg);
1301
1302         /* This group is being expired. Save the context */
1303         if (time_after(cfqd->workload_expires, jiffies)) {
1304                 cfqg->saved_wl_slice = cfqd->workload_expires
1305                                                 - jiffies;
1306                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1307                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1308         } else
1309                 cfqg->saved_wl_slice = 0;
1310
1311         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1312                                         st->min_vdisktime);
1313         cfq_log_cfqq(cfqq->cfqd, cfqq,
1314                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1315                      used_sl, cfqq->slice_dispatch, charge,
1316                      iops_mode(cfqd), cfqq->nr_sectors);
1317         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1318         cfqg_stats_set_start_empty_time(cfqg);
1319 }
1320
1321 /**
1322  * cfq_init_cfqg_base - initialize base part of a cfq_group
1323  * @cfqg: cfq_group to initialize
1324  *
1325  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1326  * is enabled or not.
1327  */
1328 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1329 {
1330         struct cfq_rb_root *st;
1331         int i, j;
1332
1333         for_each_cfqg_st(cfqg, i, j, st)
1334                 *st = CFQ_RB_ROOT;
1335         RB_CLEAR_NODE(&cfqg->rb_node);
1336
1337         cfqg->ttime.last_end_request = jiffies;
1338 }
1339
1340 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1341 static void cfq_pd_init(struct blkcg_gq *blkg)
1342 {
1343         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1344
1345         cfq_init_cfqg_base(cfqg);
1346         cfqg->weight = blkg->blkcg->cfq_weight;
1347 }
1348
1349 /*
1350  * Search for the cfq group current task belongs to. request_queue lock must
1351  * be held.
1352  */
1353 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1354                                                 struct blkcg *blkcg)
1355 {
1356         struct request_queue *q = cfqd->queue;
1357         struct cfq_group *cfqg = NULL;
1358
1359         /* avoid lookup for the common case where there's no blkcg */
1360         if (blkcg == &blkcg_root) {
1361                 cfqg = cfqd->root_group;
1362         } else {
1363                 struct blkcg_gq *blkg;
1364
1365                 blkg = blkg_lookup_create(blkcg, q);
1366                 if (!IS_ERR(blkg))
1367                         cfqg = blkg_to_cfqg(blkg);
1368         }
1369
1370         return cfqg;
1371 }
1372
1373 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1374 {
1375         /* Currently, all async queues are mapped to root group */
1376         if (!cfq_cfqq_sync(cfqq))
1377                 cfqg = cfqq->cfqd->root_group;
1378
1379         cfqq->cfqg = cfqg;
1380         /* cfqq reference on cfqg */
1381         cfqg_get(cfqg);
1382 }
1383
1384 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1385                                      struct blkg_policy_data *pd, int off)
1386 {
1387         struct cfq_group *cfqg = pd_to_cfqg(pd);
1388
1389         if (!cfqg->dev_weight)
1390                 return 0;
1391         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1392 }
1393
1394 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1395                                     struct seq_file *sf)
1396 {
1397         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1398                           cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1399                           false);
1400         return 0;
1401 }
1402
1403 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1404                             struct seq_file *sf)
1405 {
1406         seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1407         return 0;
1408 }
1409
1410 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1411                                   const char *buf)
1412 {
1413         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1414         struct blkg_conf_ctx ctx;
1415         struct cfq_group *cfqg;
1416         int ret;
1417
1418         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1419         if (ret)
1420                 return ret;
1421
1422         ret = -EINVAL;
1423         cfqg = blkg_to_cfqg(ctx.blkg);
1424         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1425                 cfqg->dev_weight = ctx.v;
1426                 cfqg->new_weight = cfqg->dev_weight ?: blkcg->cfq_weight;
1427                 ret = 0;
1428         }
1429
1430         blkg_conf_finish(&ctx);
1431         return ret;
1432 }
1433
1434 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1435 {
1436         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1437         struct blkcg_gq *blkg;
1438         struct hlist_node *n;
1439
1440         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1441                 return -EINVAL;
1442
1443         spin_lock_irq(&blkcg->lock);
1444         blkcg->cfq_weight = (unsigned int)val;
1445
1446         hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1447                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1448
1449                 if (cfqg && !cfqg->dev_weight)
1450                         cfqg->new_weight = blkcg->cfq_weight;
1451         }
1452
1453         spin_unlock_irq(&blkcg->lock);
1454         return 0;
1455 }
1456
1457 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1458                            struct seq_file *sf)
1459 {
1460         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1461
1462         blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1463                           cft->private, false);
1464         return 0;
1465 }
1466
1467 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1468                              struct seq_file *sf)
1469 {
1470         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1471
1472         blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1473                           cft->private, true);
1474         return 0;
1475 }
1476
1477 #ifdef CONFIG_DEBUG_BLK_CGROUP
1478 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1479                                       struct blkg_policy_data *pd, int off)
1480 {
1481         struct cfq_group *cfqg = pd_to_cfqg(pd);
1482         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1483         u64 v = 0;
1484
1485         if (samples) {
1486                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1487                 do_div(v, samples);
1488         }
1489         __blkg_prfill_u64(sf, pd, v);
1490         return 0;
1491 }
1492
1493 /* print avg_queue_size */
1494 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1495                                      struct seq_file *sf)
1496 {
1497         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1498
1499         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1500                           &blkcg_policy_cfq, 0, false);
1501         return 0;
1502 }
1503 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1504
1505 static struct cftype cfq_blkcg_files[] = {
1506         {
1507                 .name = "weight_device",
1508                 .read_seq_string = cfqg_print_weight_device,
1509                 .write_string = cfqg_set_weight_device,
1510                 .max_write_len = 256,
1511         },
1512         {
1513                 .name = "weight",
1514                 .read_seq_string = cfq_print_weight,
1515                 .write_u64 = cfq_set_weight,
1516         },
1517         {
1518                 .name = "time",
1519                 .private = offsetof(struct cfq_group, stats.time),
1520                 .read_seq_string = cfqg_print_stat,
1521         },
1522         {
1523                 .name = "sectors",
1524                 .private = offsetof(struct cfq_group, stats.sectors),
1525                 .read_seq_string = cfqg_print_stat,
1526         },
1527         {
1528                 .name = "io_service_bytes",
1529                 .private = offsetof(struct cfq_group, stats.service_bytes),
1530                 .read_seq_string = cfqg_print_rwstat,
1531         },
1532         {
1533                 .name = "io_serviced",
1534                 .private = offsetof(struct cfq_group, stats.serviced),
1535                 .read_seq_string = cfqg_print_rwstat,
1536         },
1537         {
1538                 .name = "io_service_time",
1539                 .private = offsetof(struct cfq_group, stats.service_time),
1540                 .read_seq_string = cfqg_print_rwstat,
1541         },
1542         {
1543                 .name = "io_wait_time",
1544                 .private = offsetof(struct cfq_group, stats.wait_time),
1545                 .read_seq_string = cfqg_print_rwstat,
1546         },
1547         {
1548                 .name = "io_merged",
1549                 .private = offsetof(struct cfq_group, stats.merged),
1550                 .read_seq_string = cfqg_print_rwstat,
1551         },
1552         {
1553                 .name = "io_queued",
1554                 .private = offsetof(struct cfq_group, stats.queued),
1555                 .read_seq_string = cfqg_print_rwstat,
1556         },
1557 #ifdef CONFIG_DEBUG_BLK_CGROUP
1558         {
1559                 .name = "avg_queue_size",
1560                 .read_seq_string = cfqg_print_avg_queue_size,
1561         },
1562         {
1563                 .name = "group_wait_time",
1564                 .private = offsetof(struct cfq_group, stats.group_wait_time),
1565                 .read_seq_string = cfqg_print_stat,
1566         },
1567         {
1568                 .name = "idle_time",
1569                 .private = offsetof(struct cfq_group, stats.idle_time),
1570                 .read_seq_string = cfqg_print_stat,
1571         },
1572         {
1573                 .name = "empty_time",
1574                 .private = offsetof(struct cfq_group, stats.empty_time),
1575                 .read_seq_string = cfqg_print_stat,
1576         },
1577         {
1578                 .name = "dequeue",
1579                 .private = offsetof(struct cfq_group, stats.dequeue),
1580                 .read_seq_string = cfqg_print_stat,
1581         },
1582         {
1583                 .name = "unaccounted_time",
1584                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1585                 .read_seq_string = cfqg_print_stat,
1586         },
1587 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1588         { }     /* terminate */
1589 };
1590 #else /* GROUP_IOSCHED */
1591 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1592                                                 struct blkcg *blkcg)
1593 {
1594         return cfqd->root_group;
1595 }
1596
1597 static inline void
1598 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1599         cfqq->cfqg = cfqg;
1600 }
1601
1602 #endif /* GROUP_IOSCHED */
1603
1604 /*
1605  * The cfqd->service_trees holds all pending cfq_queue's that have
1606  * requests waiting to be processed. It is sorted in the order that
1607  * we will service the queues.
1608  */
1609 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1610                                  bool add_front)
1611 {
1612         struct rb_node **p, *parent;
1613         struct cfq_queue *__cfqq;
1614         unsigned long rb_key;
1615         struct cfq_rb_root *st;
1616         int left;
1617         int new_cfqq = 1;
1618
1619         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
1620         if (cfq_class_idle(cfqq)) {
1621                 rb_key = CFQ_IDLE_DELAY;
1622                 parent = rb_last(&st->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(st);
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 && cfqq->service_tree == st)
1650                         return;
1651
1652                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1653                 cfqq->service_tree = NULL;
1654         }
1655
1656         left = 1;
1657         parent = NULL;
1658         cfqq->service_tree = st;
1659         p = &st->rb.rb_node;
1660         while (*p) {
1661                 struct rb_node **n;
1662
1663                 parent = *p;
1664                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1665
1666                 /*
1667                  * sort by key, that represents service time.
1668                  */
1669                 if (time_before(rb_key, __cfqq->rb_key))
1670                         n = &(*p)->rb_left;
1671                 else {
1672                         n = &(*p)->rb_right;
1673                         left = 0;
1674                 }
1675
1676                 p = n;
1677         }
1678
1679         if (left)
1680                 st->left = &cfqq->rb_node;
1681
1682         cfqq->rb_key = rb_key;
1683         rb_link_node(&cfqq->rb_node, parent, p);
1684         rb_insert_color(&cfqq->rb_node, &st->rb);
1685         st->count++;
1686         if (add_front || !new_cfqq)
1687                 return;
1688         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1689 }
1690
1691 static struct cfq_queue *
1692 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1693                      sector_t sector, struct rb_node **ret_parent,
1694                      struct rb_node ***rb_link)
1695 {
1696         struct rb_node **p, *parent;
1697         struct cfq_queue *cfqq = NULL;
1698
1699         parent = NULL;
1700         p = &root->rb_node;
1701         while (*p) {
1702                 struct rb_node **n;
1703
1704                 parent = *p;
1705                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1706
1707                 /*
1708                  * Sort strictly based on sector.  Smallest to the left,
1709                  * largest to the right.
1710                  */
1711                 if (sector > blk_rq_pos(cfqq->next_rq))
1712                         n = &(*p)->rb_right;
1713                 else if (sector < blk_rq_pos(cfqq->next_rq))
1714                         n = &(*p)->rb_left;
1715                 else
1716                         break;
1717                 p = n;
1718                 cfqq = NULL;
1719         }
1720
1721         *ret_parent = parent;
1722         if (rb_link)
1723                 *rb_link = p;
1724         return cfqq;
1725 }
1726
1727 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1728 {
1729         struct rb_node **p, *parent;
1730         struct cfq_queue *__cfqq;
1731
1732         if (cfqq->p_root) {
1733                 rb_erase(&cfqq->p_node, cfqq->p_root);
1734                 cfqq->p_root = NULL;
1735         }
1736
1737         if (cfq_class_idle(cfqq))
1738                 return;
1739         if (!cfqq->next_rq)
1740                 return;
1741
1742         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1743         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1744                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1745         if (!__cfqq) {
1746                 rb_link_node(&cfqq->p_node, parent, p);
1747                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1748         } else
1749                 cfqq->p_root = NULL;
1750 }
1751
1752 /*
1753  * Update cfqq's position in the service tree.
1754  */
1755 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1756 {
1757         /*
1758          * Resorting requires the cfqq to be on the RR list already.
1759          */
1760         if (cfq_cfqq_on_rr(cfqq)) {
1761                 cfq_service_tree_add(cfqd, cfqq, 0);
1762                 cfq_prio_tree_add(cfqd, cfqq);
1763         }
1764 }
1765
1766 /*
1767  * add to busy list of queues for service, trying to be fair in ordering
1768  * the pending list according to last request service
1769  */
1770 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1771 {
1772         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1773         BUG_ON(cfq_cfqq_on_rr(cfqq));
1774         cfq_mark_cfqq_on_rr(cfqq);
1775         cfqd->busy_queues++;
1776         if (cfq_cfqq_sync(cfqq))
1777                 cfqd->busy_sync_queues++;
1778
1779         cfq_resort_rr_list(cfqd, cfqq);
1780 }
1781
1782 /*
1783  * Called when the cfqq no longer has requests pending, remove it from
1784  * the service tree.
1785  */
1786 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1787 {
1788         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1789         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1790         cfq_clear_cfqq_on_rr(cfqq);
1791
1792         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1793                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1794                 cfqq->service_tree = NULL;
1795         }
1796         if (cfqq->p_root) {
1797                 rb_erase(&cfqq->p_node, cfqq->p_root);
1798                 cfqq->p_root = NULL;
1799         }
1800
1801         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1802         BUG_ON(!cfqd->busy_queues);
1803         cfqd->busy_queues--;
1804         if (cfq_cfqq_sync(cfqq))
1805                 cfqd->busy_sync_queues--;
1806 }
1807
1808 /*
1809  * rb tree support functions
1810  */
1811 static void cfq_del_rq_rb(struct request *rq)
1812 {
1813         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1814         const int sync = rq_is_sync(rq);
1815
1816         BUG_ON(!cfqq->queued[sync]);
1817         cfqq->queued[sync]--;
1818
1819         elv_rb_del(&cfqq->sort_list, rq);
1820
1821         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1822                 /*
1823                  * Queue will be deleted from service tree when we actually
1824                  * expire it later. Right now just remove it from prio tree
1825                  * as it is empty.
1826                  */
1827                 if (cfqq->p_root) {
1828                         rb_erase(&cfqq->p_node, cfqq->p_root);
1829                         cfqq->p_root = NULL;
1830                 }
1831         }
1832 }
1833
1834 static void cfq_add_rq_rb(struct request *rq)
1835 {
1836         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1837         struct cfq_data *cfqd = cfqq->cfqd;
1838         struct request *prev;
1839
1840         cfqq->queued[rq_is_sync(rq)]++;
1841
1842         elv_rb_add(&cfqq->sort_list, rq);
1843
1844         if (!cfq_cfqq_on_rr(cfqq))
1845                 cfq_add_cfqq_rr(cfqd, cfqq);
1846
1847         /*
1848          * check if this request is a better next-serve candidate
1849          */
1850         prev = cfqq->next_rq;
1851         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1852
1853         /*
1854          * adjust priority tree position, if ->next_rq changes
1855          */
1856         if (prev != cfqq->next_rq)
1857                 cfq_prio_tree_add(cfqd, cfqq);
1858
1859         BUG_ON(!cfqq->next_rq);
1860 }
1861
1862 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1863 {
1864         elv_rb_del(&cfqq->sort_list, rq);
1865         cfqq->queued[rq_is_sync(rq)]--;
1866         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1867         cfq_add_rq_rb(rq);
1868         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
1869                                  rq->cmd_flags);
1870 }
1871
1872 static struct request *
1873 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1874 {
1875         struct task_struct *tsk = current;
1876         struct cfq_io_cq *cic;
1877         struct cfq_queue *cfqq;
1878
1879         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1880         if (!cic)
1881                 return NULL;
1882
1883         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1884         if (cfqq) {
1885                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1886
1887                 return elv_rb_find(&cfqq->sort_list, sector);
1888         }
1889
1890         return NULL;
1891 }
1892
1893 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1894 {
1895         struct cfq_data *cfqd = q->elevator->elevator_data;
1896
1897         cfqd->rq_in_driver++;
1898         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1899                                                 cfqd->rq_in_driver);
1900
1901         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1902 }
1903
1904 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1905 {
1906         struct cfq_data *cfqd = q->elevator->elevator_data;
1907
1908         WARN_ON(!cfqd->rq_in_driver);
1909         cfqd->rq_in_driver--;
1910         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1911                                                 cfqd->rq_in_driver);
1912 }
1913
1914 static void cfq_remove_request(struct request *rq)
1915 {
1916         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1917
1918         if (cfqq->next_rq == rq)
1919                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1920
1921         list_del_init(&rq->queuelist);
1922         cfq_del_rq_rb(rq);
1923
1924         cfqq->cfqd->rq_queued--;
1925         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1926         if (rq->cmd_flags & REQ_PRIO) {
1927                 WARN_ON(!cfqq->prio_pending);
1928                 cfqq->prio_pending--;
1929         }
1930 }
1931
1932 static int cfq_merge(struct request_queue *q, struct request **req,
1933                      struct bio *bio)
1934 {
1935         struct cfq_data *cfqd = q->elevator->elevator_data;
1936         struct request *__rq;
1937
1938         __rq = cfq_find_rq_fmerge(cfqd, bio);
1939         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1940                 *req = __rq;
1941                 return ELEVATOR_FRONT_MERGE;
1942         }
1943
1944         return ELEVATOR_NO_MERGE;
1945 }
1946
1947 static void cfq_merged_request(struct request_queue *q, struct request *req,
1948                                int type)
1949 {
1950         if (type == ELEVATOR_FRONT_MERGE) {
1951                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1952
1953                 cfq_reposition_rq_rb(cfqq, req);
1954         }
1955 }
1956
1957 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1958                                 struct bio *bio)
1959 {
1960         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
1961 }
1962
1963 static void
1964 cfq_merged_requests(struct request_queue *q, struct request *rq,
1965                     struct request *next)
1966 {
1967         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1968         struct cfq_data *cfqd = q->elevator->elevator_data;
1969
1970         /*
1971          * reposition in fifo if next is older than rq
1972          */
1973         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1974             time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
1975             cfqq == RQ_CFQQ(next)) {
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_class:%d wl_type:%d",
2032                                 cfqd->serving_wl_class, cfqd->serving_wl_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 *st = st_for(cfqd->serving_group,
2119                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2120
2121         if (!cfqd->rq_queued)
2122                 return NULL;
2123
2124         /* There is nothing to dispatch */
2125         if (!st)
2126                 return NULL;
2127         if (RB_EMPTY_ROOT(&st->rb))
2128                 return NULL;
2129         return cfq_rb_first(st);
2130 }
2131
2132 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2133 {
2134         struct cfq_group *cfqg;
2135         struct cfq_queue *cfqq;
2136         int i, j;
2137         struct cfq_rb_root *st;
2138
2139         if (!cfqd->rq_queued)
2140                 return NULL;
2141
2142         cfqg = cfq_get_next_cfqg(cfqd);
2143         if (!cfqg)
2144                 return NULL;
2145
2146         for_each_cfqg_st(cfqg, i, j, st)
2147                 if ((cfqq = cfq_rb_first(st)) != NULL)
2148                         return cfqq;
2149         return NULL;
2150 }
2151
2152 /*
2153  * Get and set a new active queue for service.
2154  */
2155 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2156                                               struct cfq_queue *cfqq)
2157 {
2158         if (!cfqq)
2159                 cfqq = cfq_get_next_queue(cfqd);
2160
2161         __cfq_set_active_queue(cfqd, cfqq);
2162         return cfqq;
2163 }
2164
2165 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2166                                           struct request *rq)
2167 {
2168         if (blk_rq_pos(rq) >= cfqd->last_position)
2169                 return blk_rq_pos(rq) - cfqd->last_position;
2170         else
2171                 return cfqd->last_position - blk_rq_pos(rq);
2172 }
2173
2174 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2175                                struct request *rq)
2176 {
2177         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2178 }
2179
2180 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2181                                     struct cfq_queue *cur_cfqq)
2182 {
2183         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2184         struct rb_node *parent, *node;
2185         struct cfq_queue *__cfqq;
2186         sector_t sector = cfqd->last_position;
2187
2188         if (RB_EMPTY_ROOT(root))
2189                 return NULL;
2190
2191         /*
2192          * First, if we find a request starting at the end of the last
2193          * request, choose it.
2194          */
2195         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2196         if (__cfqq)
2197                 return __cfqq;
2198
2199         /*
2200          * If the exact sector wasn't found, the parent of the NULL leaf
2201          * will contain the closest sector.
2202          */
2203         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2204         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2205                 return __cfqq;
2206
2207         if (blk_rq_pos(__cfqq->next_rq) < sector)
2208                 node = rb_next(&__cfqq->p_node);
2209         else
2210                 node = rb_prev(&__cfqq->p_node);
2211         if (!node)
2212                 return NULL;
2213
2214         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2215         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2216                 return __cfqq;
2217
2218         return NULL;
2219 }
2220
2221 /*
2222  * cfqd - obvious
2223  * cur_cfqq - passed in so that we don't decide that the current queue is
2224  *            closely cooperating with itself.
2225  *
2226  * So, basically we're assuming that that cur_cfqq has dispatched at least
2227  * one request, and that cfqd->last_position reflects a position on the disk
2228  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2229  * assumption.
2230  */
2231 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2232                                               struct cfq_queue *cur_cfqq)
2233 {
2234         struct cfq_queue *cfqq;
2235
2236         if (cfq_class_idle(cur_cfqq))
2237                 return NULL;
2238         if (!cfq_cfqq_sync(cur_cfqq))
2239                 return NULL;
2240         if (CFQQ_SEEKY(cur_cfqq))
2241                 return NULL;
2242
2243         /*
2244          * Don't search priority tree if it's the only queue in the group.
2245          */
2246         if (cur_cfqq->cfqg->nr_cfqq == 1)
2247                 return NULL;
2248
2249         /*
2250          * We should notice if some of the queues are cooperating, eg
2251          * working closely on the same area of the disk. In that case,
2252          * we can group them together and don't waste time idling.
2253          */
2254         cfqq = cfqq_close(cfqd, cur_cfqq);
2255         if (!cfqq)
2256                 return NULL;
2257
2258         /* If new queue belongs to different cfq_group, don't choose it */
2259         if (cur_cfqq->cfqg != cfqq->cfqg)
2260                 return NULL;
2261
2262         /*
2263          * It only makes sense to merge sync queues.
2264          */
2265         if (!cfq_cfqq_sync(cfqq))
2266                 return NULL;
2267         if (CFQQ_SEEKY(cfqq))
2268                 return NULL;
2269
2270         /*
2271          * Do not merge queues of different priority classes
2272          */
2273         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2274                 return NULL;
2275
2276         return cfqq;
2277 }
2278
2279 /*
2280  * Determine whether we should enforce idle window for this queue.
2281  */
2282
2283 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2284 {
2285         enum wl_class_t wl_class = cfqq_class(cfqq);
2286         struct cfq_rb_root *st = cfqq->service_tree;
2287
2288         BUG_ON(!st);
2289         BUG_ON(!st->count);
2290
2291         if (!cfqd->cfq_slice_idle)
2292                 return false;
2293
2294         /* We never do for idle class queues. */
2295         if (wl_class == IDLE_WORKLOAD)
2296                 return false;
2297
2298         /* We do for queues that were marked with idle window flag. */
2299         if (cfq_cfqq_idle_window(cfqq) &&
2300            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2301                 return true;
2302
2303         /*
2304          * Otherwise, we do only if they are the last ones
2305          * in their service tree.
2306          */
2307         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2308            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2309                 return true;
2310         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2311         return false;
2312 }
2313
2314 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2315 {
2316         struct cfq_queue *cfqq = cfqd->active_queue;
2317         struct cfq_io_cq *cic;
2318         unsigned long sl, group_idle = 0;
2319
2320         /*
2321          * SSD device without seek penalty, disable idling. But only do so
2322          * for devices that support queuing, otherwise we still have a problem
2323          * with sync vs async workloads.
2324          */
2325         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2326                 return;
2327
2328         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2329         WARN_ON(cfq_cfqq_slice_new(cfqq));
2330
2331         /*
2332          * idle is disabled, either manually or by past process history
2333          */
2334         if (!cfq_should_idle(cfqd, cfqq)) {
2335                 /* no queue idling. Check for group idling */
2336                 if (cfqd->cfq_group_idle)
2337                         group_idle = cfqd->cfq_group_idle;
2338                 else
2339                         return;
2340         }
2341
2342         /*
2343          * still active requests from this queue, don't idle
2344          */
2345         if (cfqq->dispatched)
2346                 return;
2347
2348         /*
2349          * task has exited, don't wait
2350          */
2351         cic = cfqd->active_cic;
2352         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2353                 return;
2354
2355         /*
2356          * If our average think time is larger than the remaining time
2357          * slice, then don't idle. This avoids overrunning the allotted
2358          * time slice.
2359          */
2360         if (sample_valid(cic->ttime.ttime_samples) &&
2361             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2362                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2363                              cic->ttime.ttime_mean);
2364                 return;
2365         }
2366
2367         /* There are other queues in the group, don't do group idle */
2368         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2369                 return;
2370
2371         cfq_mark_cfqq_wait_request(cfqq);
2372
2373         if (group_idle)
2374                 sl = cfqd->cfq_group_idle;
2375         else
2376                 sl = cfqd->cfq_slice_idle;
2377
2378         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2379         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2380         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2381                         group_idle ? 1 : 0);
2382 }
2383
2384 /*
2385  * Move request from internal lists to the request queue dispatch list.
2386  */
2387 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2388 {
2389         struct cfq_data *cfqd = q->elevator->elevator_data;
2390         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2391
2392         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2393
2394         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2395         cfq_remove_request(rq);
2396         cfqq->dispatched++;
2397         (RQ_CFQG(rq))->dispatched++;
2398         elv_dispatch_sort(q, rq);
2399
2400         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2401         cfqq->nr_sectors += blk_rq_sectors(rq);
2402         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2403 }
2404
2405 /*
2406  * return expired entry, or NULL to just start from scratch in rbtree
2407  */
2408 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2409 {
2410         struct request *rq = NULL;
2411
2412         if (cfq_cfqq_fifo_expire(cfqq))
2413                 return NULL;
2414
2415         cfq_mark_cfqq_fifo_expire(cfqq);
2416
2417         if (list_empty(&cfqq->fifo))
2418                 return NULL;
2419
2420         rq = rq_entry_fifo(cfqq->fifo.next);
2421         if (time_before(jiffies, rq_fifo_time(rq)))
2422                 rq = NULL;
2423
2424         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2425         return rq;
2426 }
2427
2428 static inline int
2429 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2430 {
2431         const int base_rq = cfqd->cfq_slice_async_rq;
2432
2433         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2434
2435         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2436 }
2437
2438 /*
2439  * Must be called with the queue_lock held.
2440  */
2441 static int cfqq_process_refs(struct cfq_queue *cfqq)
2442 {
2443         int process_refs, io_refs;
2444
2445         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2446         process_refs = cfqq->ref - io_refs;
2447         BUG_ON(process_refs < 0);
2448         return process_refs;
2449 }
2450
2451 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2452 {
2453         int process_refs, new_process_refs;
2454         struct cfq_queue *__cfqq;
2455
2456         /*
2457          * If there are no process references on the new_cfqq, then it is
2458          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2459          * chain may have dropped their last reference (not just their
2460          * last process reference).
2461          */
2462         if (!cfqq_process_refs(new_cfqq))
2463                 return;
2464
2465         /* Avoid a circular list and skip interim queue merges */
2466         while ((__cfqq = new_cfqq->new_cfqq)) {
2467                 if (__cfqq == cfqq)
2468                         return;
2469                 new_cfqq = __cfqq;
2470         }
2471
2472         process_refs = cfqq_process_refs(cfqq);
2473         new_process_refs = cfqq_process_refs(new_cfqq);
2474         /*
2475          * If the process for the cfqq has gone away, there is no
2476          * sense in merging the queues.
2477          */
2478         if (process_refs == 0 || new_process_refs == 0)
2479                 return;
2480
2481         /*
2482          * Merge in the direction of the lesser amount of work.
2483          */
2484         if (new_process_refs >= process_refs) {
2485                 cfqq->new_cfqq = new_cfqq;
2486                 new_cfqq->ref += process_refs;
2487         } else {
2488                 new_cfqq->new_cfqq = cfqq;
2489                 cfqq->ref += new_process_refs;
2490         }
2491 }
2492
2493 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2494                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2495 {
2496         struct cfq_queue *queue;
2497         int i;
2498         bool key_valid = false;
2499         unsigned long lowest_key = 0;
2500         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2501
2502         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2503                 /* select the one with lowest rb_key */
2504                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2505                 if (queue &&
2506                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2507                         lowest_key = queue->rb_key;
2508                         cur_best = i;
2509                         key_valid = true;
2510                 }
2511         }
2512
2513         return cur_best;
2514 }
2515
2516 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2517 {
2518         unsigned slice;
2519         unsigned count;
2520         struct cfq_rb_root *st;
2521         unsigned group_slice;
2522         enum wl_class_t original_class = cfqd->serving_wl_class;
2523
2524         /* Choose next priority. RT > BE > IDLE */
2525         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2526                 cfqd->serving_wl_class = RT_WORKLOAD;
2527         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2528                 cfqd->serving_wl_class = BE_WORKLOAD;
2529         else {
2530                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2531                 cfqd->workload_expires = jiffies + 1;
2532                 return;
2533         }
2534
2535         if (original_class != cfqd->serving_wl_class)
2536                 goto new_workload;
2537
2538         /*
2539          * For RT and BE, we have to choose also the type
2540          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2541          * expiration time
2542          */
2543         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2544         count = st->count;
2545
2546         /*
2547          * check workload expiration, and that we still have other queues ready
2548          */
2549         if (count && !time_after(jiffies, cfqd->workload_expires))
2550                 return;
2551
2552 new_workload:
2553         /* otherwise select new workload type */
2554         cfqd->serving_wl_type = cfq_choose_wl(cfqd, cfqg,
2555                                         cfqd->serving_wl_class);
2556         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2557         count = st->count;
2558
2559         /*
2560          * the workload slice is computed as a fraction of target latency
2561          * proportional to the number of queues in that workload, over
2562          * all the queues in the same priority class
2563          */
2564         group_slice = cfq_group_slice(cfqd, cfqg);
2565
2566         slice = group_slice * count /
2567                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2568                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2569                                         cfqg));
2570
2571         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2572                 unsigned int tmp;
2573
2574                 /*
2575                  * Async queues are currently system wide. Just taking
2576                  * proportion of queues with-in same group will lead to higher
2577                  * async ratio system wide as generally root group is going
2578                  * to have higher weight. A more accurate thing would be to
2579                  * calculate system wide asnc/sync ratio.
2580                  */
2581                 tmp = cfqd->cfq_target_latency *
2582                         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_wl_slice) {
2618                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
2619                 cfqd->serving_wl_type = cfqg->saved_wl_type;
2620                 cfqd->serving_wl_class = cfqg->saved_wl_class;
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_wl_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_wl_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 *st;
3638
3639                 RQ_CIC(rq)->ttime.last_end_request = now;
3640
3641                 if (cfq_cfqq_on_rr(cfqq))
3642                         st = cfqq->service_tree;
3643                 else
3644                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
3645                                         cfqq_type(cfqq));
3646
3647                 st->ttime.last_end_request = now;
3648                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3649                         cfqd->last_delayed_sync = now;
3650         }
3651
3652 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3653         cfqq->cfqg->ttime.last_end_request = now;
3654 #endif
3655
3656         /*
3657          * If this is the active queue, check if it needs to be expired,
3658          * or if we want to idle in case it has no pending requests.
3659          */
3660         if (cfqd->active_queue == cfqq) {
3661                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3662
3663                 if (cfq_cfqq_slice_new(cfqq)) {
3664                         cfq_set_prio_slice(cfqd, cfqq);
3665                         cfq_clear_cfqq_slice_new(cfqq);
3666                 }
3667
3668                 /*
3669                  * Should we wait for next request to come in before we expire
3670                  * the queue.
3671                  */
3672                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3673                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3674                         if (!cfqd->cfq_slice_idle)
3675                                 extend_sl = cfqd->cfq_group_idle;
3676                         cfqq->slice_end = jiffies + extend_sl;
3677                         cfq_mark_cfqq_wait_busy(cfqq);
3678                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3679                 }
3680
3681                 /*
3682                  * Idling is not enabled on:
3683                  * - expired queues
3684                  * - idle-priority queues
3685                  * - async queues
3686                  * - queues with still some requests queued
3687                  * - when there is a close cooperator
3688                  */
3689                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3690                         cfq_slice_expired(cfqd, 1);
3691                 else if (sync && cfqq_empty &&
3692                          !cfq_close_cooperator(cfqd, cfqq)) {
3693                         cfq_arm_slice_timer(cfqd);
3694                 }
3695         }
3696
3697         if (!cfqd->rq_in_driver)
3698                 cfq_schedule_dispatch(cfqd);
3699 }
3700
3701 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3702 {
3703         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3704                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3705                 return ELV_MQUEUE_MUST;
3706         }
3707
3708         return ELV_MQUEUE_MAY;
3709 }
3710
3711 static int cfq_may_queue(struct request_queue *q, int rw)
3712 {
3713         struct cfq_data *cfqd = q->elevator->elevator_data;
3714         struct task_struct *tsk = current;
3715         struct cfq_io_cq *cic;
3716         struct cfq_queue *cfqq;
3717
3718         /*
3719          * don't force setup of a queue from here, as a call to may_queue
3720          * does not necessarily imply that a request actually will be queued.
3721          * so just lookup a possibly existing queue, or return 'may queue'
3722          * if that fails
3723          */
3724         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3725         if (!cic)
3726                 return ELV_MQUEUE_MAY;
3727
3728         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3729         if (cfqq) {
3730                 cfq_init_prio_data(cfqq, cic);
3731
3732                 return __cfq_may_queue(cfqq);
3733         }
3734
3735         return ELV_MQUEUE_MAY;
3736 }
3737
3738 /*
3739  * queue lock held here
3740  */
3741 static void cfq_put_request(struct request *rq)
3742 {
3743         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3744
3745         if (cfqq) {
3746                 const int rw = rq_data_dir(rq);
3747
3748                 BUG_ON(!cfqq->allocated[rw]);
3749                 cfqq->allocated[rw]--;
3750
3751                 /* Put down rq reference on cfqg */
3752                 cfqg_put(RQ_CFQG(rq));
3753                 rq->elv.priv[0] = NULL;
3754                 rq->elv.priv[1] = NULL;
3755
3756                 cfq_put_queue(cfqq);
3757         }
3758 }
3759
3760 static struct cfq_queue *
3761 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3762                 struct cfq_queue *cfqq)
3763 {
3764         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3765         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3766         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3767         cfq_put_queue(cfqq);
3768         return cic_to_cfqq(cic, 1);
3769 }
3770
3771 /*
3772  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3773  * was the last process referring to said cfqq.
3774  */
3775 static struct cfq_queue *
3776 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3777 {
3778         if (cfqq_process_refs(cfqq) == 1) {
3779                 cfqq->pid = current->pid;
3780                 cfq_clear_cfqq_coop(cfqq);
3781                 cfq_clear_cfqq_split_coop(cfqq);
3782                 return cfqq;
3783         }
3784
3785         cic_set_cfqq(cic, NULL, 1);
3786
3787         cfq_put_cooperator(cfqq);
3788
3789         cfq_put_queue(cfqq);
3790         return NULL;
3791 }
3792 /*
3793  * Allocate cfq data structures associated with this request.
3794  */
3795 static int
3796 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3797                 gfp_t gfp_mask)
3798 {
3799         struct cfq_data *cfqd = q->elevator->elevator_data;
3800         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3801         const int rw = rq_data_dir(rq);
3802         const bool is_sync = rq_is_sync(rq);
3803         struct cfq_queue *cfqq;
3804
3805         might_sleep_if(gfp_mask & __GFP_WAIT);
3806
3807         spin_lock_irq(q->queue_lock);
3808
3809         check_ioprio_changed(cic, bio);
3810         check_blkcg_changed(cic, bio);
3811 new_queue:
3812         cfqq = cic_to_cfqq(cic, is_sync);
3813         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3814                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
3815                 cic_set_cfqq(cic, cfqq, is_sync);
3816         } else {
3817                 /*
3818                  * If the queue was seeky for too long, break it apart.
3819                  */
3820                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3821                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3822                         cfqq = split_cfqq(cic, cfqq);
3823                         if (!cfqq)
3824                                 goto new_queue;
3825                 }
3826
3827                 /*
3828                  * Check to see if this queue is scheduled to merge with
3829                  * another, closely cooperating queue.  The merging of
3830                  * queues happens here as it must be done in process context.
3831                  * The reference on new_cfqq was taken in merge_cfqqs.
3832                  */
3833                 if (cfqq->new_cfqq)
3834                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3835         }
3836
3837         cfqq->allocated[rw]++;
3838
3839         cfqq->ref++;
3840         cfqg_get(cfqq->cfqg);
3841         rq->elv.priv[0] = cfqq;
3842         rq->elv.priv[1] = cfqq->cfqg;
3843         spin_unlock_irq(q->queue_lock);
3844         return 0;
3845 }
3846
3847 static void cfq_kick_queue(struct work_struct *work)
3848 {
3849         struct cfq_data *cfqd =
3850                 container_of(work, struct cfq_data, unplug_work);
3851         struct request_queue *q = cfqd->queue;
3852
3853         spin_lock_irq(q->queue_lock);
3854         __blk_run_queue(cfqd->queue);
3855         spin_unlock_irq(q->queue_lock);
3856 }
3857
3858 /*
3859  * Timer running if the active_queue is currently idling inside its time slice
3860  */
3861 static void cfq_idle_slice_timer(unsigned long data)
3862 {
3863         struct cfq_data *cfqd = (struct cfq_data *) data;
3864         struct cfq_queue *cfqq;
3865         unsigned long flags;
3866         int timed_out = 1;
3867
3868         cfq_log(cfqd, "idle timer fired");
3869
3870         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3871
3872         cfqq = cfqd->active_queue;
3873         if (cfqq) {
3874                 timed_out = 0;
3875
3876                 /*
3877                  * We saw a request before the queue expired, let it through
3878                  */
3879                 if (cfq_cfqq_must_dispatch(cfqq))
3880                         goto out_kick;
3881
3882                 /*
3883                  * expired
3884                  */
3885                 if (cfq_slice_used(cfqq))
3886                         goto expire;
3887
3888                 /*
3889                  * only expire and reinvoke request handler, if there are
3890                  * other queues with pending requests
3891                  */
3892                 if (!cfqd->busy_queues)
3893                         goto out_cont;
3894
3895                 /*
3896                  * not expired and it has a request pending, let it dispatch
3897                  */
3898                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3899                         goto out_kick;
3900
3901                 /*
3902                  * Queue depth flag is reset only when the idle didn't succeed
3903                  */
3904                 cfq_clear_cfqq_deep(cfqq);
3905         }
3906 expire:
3907         cfq_slice_expired(cfqd, timed_out);
3908 out_kick:
3909         cfq_schedule_dispatch(cfqd);
3910 out_cont:
3911         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3912 }
3913
3914 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3915 {
3916         del_timer_sync(&cfqd->idle_slice_timer);
3917         cancel_work_sync(&cfqd->unplug_work);
3918 }
3919
3920 static void cfq_put_async_queues(struct cfq_data *cfqd)
3921 {
3922         int i;
3923
3924         for (i = 0; i < IOPRIO_BE_NR; i++) {
3925                 if (cfqd->async_cfqq[0][i])
3926                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3927                 if (cfqd->async_cfqq[1][i])
3928                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3929         }
3930
3931         if (cfqd->async_idle_cfqq)
3932                 cfq_put_queue(cfqd->async_idle_cfqq);
3933 }
3934
3935 static void cfq_exit_queue(struct elevator_queue *e)
3936 {
3937         struct cfq_data *cfqd = e->elevator_data;
3938         struct request_queue *q = cfqd->queue;
3939
3940         cfq_shutdown_timer_wq(cfqd);
3941
3942         spin_lock_irq(q->queue_lock);
3943
3944         if (cfqd->active_queue)
3945                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3946
3947         cfq_put_async_queues(cfqd);
3948
3949         spin_unlock_irq(q->queue_lock);
3950
3951         cfq_shutdown_timer_wq(cfqd);
3952
3953 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3954         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
3955 #else
3956         kfree(cfqd->root_group);
3957 #endif
3958         kfree(cfqd);
3959 }
3960
3961 static int cfq_init_queue(struct request_queue *q)
3962 {
3963         struct cfq_data *cfqd;
3964         struct blkcg_gq *blkg __maybe_unused;
3965         int i, ret;
3966
3967         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3968         if (!cfqd)
3969                 return -ENOMEM;
3970
3971         cfqd->queue = q;
3972         q->elevator->elevator_data = cfqd;
3973
3974         /* Init root service tree */
3975         cfqd->grp_service_tree = CFQ_RB_ROOT;
3976
3977         /* Init root group and prefer root group over other groups by default */
3978 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3979         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
3980         if (ret)
3981                 goto out_free;
3982
3983         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
3984 #else
3985         ret = -ENOMEM;
3986         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3987                                         GFP_KERNEL, cfqd->queue->node);
3988         if (!cfqd->root_group)
3989                 goto out_free;
3990
3991         cfq_init_cfqg_base(cfqd->root_group);
3992 #endif
3993         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
3994
3995         /*
3996          * Not strictly needed (since RB_ROOT just clears the node and we
3997          * zeroed cfqd on alloc), but better be safe in case someone decides
3998          * to add magic to the rb code
3999          */
4000         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4001                 cfqd->prio_trees[i] = RB_ROOT;
4002
4003         /*
4004          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4005          * Grab a permanent reference to it, so that the normal code flow
4006          * will not attempt to free it.  oom_cfqq is linked to root_group
4007          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4008          * the reference from linking right away.
4009          */
4010         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4011         cfqd->oom_cfqq.ref++;
4012
4013         spin_lock_irq(q->queue_lock);
4014         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4015         cfqg_put(cfqd->root_group);
4016         spin_unlock_irq(q->queue_lock);
4017
4018         init_timer(&cfqd->idle_slice_timer);
4019         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4020         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4021
4022         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4023
4024         cfqd->cfq_quantum = cfq_quantum;
4025         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4026         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4027         cfqd->cfq_back_max = cfq_back_max;
4028         cfqd->cfq_back_penalty = cfq_back_penalty;
4029         cfqd->cfq_slice[0] = cfq_slice_async;
4030         cfqd->cfq_slice[1] = cfq_slice_sync;
4031         cfqd->cfq_target_latency = cfq_target_latency;
4032         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4033         cfqd->cfq_slice_idle = cfq_slice_idle;
4034         cfqd->cfq_group_idle = cfq_group_idle;
4035         cfqd->cfq_latency = 1;
4036         cfqd->hw_tag = -1;
4037         /*
4038          * we optimistically start assuming sync ops weren't delayed in last
4039          * second, in order to have larger depth for async operations.
4040          */
4041         cfqd->last_delayed_sync = jiffies - HZ;
4042         return 0;
4043
4044 out_free:
4045         kfree(cfqd);
4046         return ret;
4047 }
4048
4049 /*
4050  * sysfs parts below -->
4051  */
4052 static ssize_t
4053 cfq_var_show(unsigned int var, char *page)
4054 {
4055         return sprintf(page, "%d\n", var);
4056 }
4057
4058 static ssize_t
4059 cfq_var_store(unsigned int *var, const char *page, size_t count)
4060 {
4061         char *p = (char *) page;
4062
4063         *var = simple_strtoul(p, &p, 10);
4064         return count;
4065 }
4066
4067 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4068 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4069 {                                                                       \
4070         struct cfq_data *cfqd = e->elevator_data;                       \
4071         unsigned int __data = __VAR;                                    \
4072         if (__CONV)                                                     \
4073                 __data = jiffies_to_msecs(__data);                      \
4074         return cfq_var_show(__data, (page));                            \
4075 }
4076 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4077 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4078 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4079 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4080 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4081 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4082 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4083 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4084 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4085 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4086 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4087 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4088 #undef SHOW_FUNCTION
4089
4090 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4091 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4092 {                                                                       \
4093         struct cfq_data *cfqd = e->elevator_data;                       \
4094         unsigned int __data;                                            \
4095         int ret = cfq_var_store(&__data, (page), count);                \
4096         if (__data < (MIN))                                             \
4097                 __data = (MIN);                                         \
4098         else if (__data > (MAX))                                        \
4099                 __data = (MAX);                                         \
4100         if (__CONV)                                                     \
4101                 *(__PTR) = msecs_to_jiffies(__data);                    \
4102         else                                                            \
4103                 *(__PTR) = __data;                                      \
4104         return ret;                                                     \
4105 }
4106 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4107 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4108                 UINT_MAX, 1);
4109 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4110                 UINT_MAX, 1);
4111 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4112 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4113                 UINT_MAX, 0);
4114 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4115 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4116 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4117 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4118 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4119                 UINT_MAX, 0);
4120 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4121 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4122 #undef STORE_FUNCTION
4123
4124 #define CFQ_ATTR(name) \
4125         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4126
4127 static struct elv_fs_entry cfq_attrs[] = {
4128         CFQ_ATTR(quantum),
4129         CFQ_ATTR(fifo_expire_sync),
4130         CFQ_ATTR(fifo_expire_async),
4131         CFQ_ATTR(back_seek_max),
4132         CFQ_ATTR(back_seek_penalty),
4133         CFQ_ATTR(slice_sync),
4134         CFQ_ATTR(slice_async),
4135         CFQ_ATTR(slice_async_rq),
4136         CFQ_ATTR(slice_idle),
4137         CFQ_ATTR(group_idle),
4138         CFQ_ATTR(low_latency),
4139         CFQ_ATTR(target_latency),
4140         __ATTR_NULL
4141 };
4142
4143 static struct elevator_type iosched_cfq = {
4144         .ops = {
4145                 .elevator_merge_fn =            cfq_merge,
4146                 .elevator_merged_fn =           cfq_merged_request,
4147                 .elevator_merge_req_fn =        cfq_merged_requests,
4148                 .elevator_allow_merge_fn =      cfq_allow_merge,
4149                 .elevator_bio_merged_fn =       cfq_bio_merged,
4150                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4151                 .elevator_add_req_fn =          cfq_insert_request,
4152                 .elevator_activate_req_fn =     cfq_activate_request,
4153                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4154                 .elevator_completed_req_fn =    cfq_completed_request,
4155                 .elevator_former_req_fn =       elv_rb_former_request,
4156                 .elevator_latter_req_fn =       elv_rb_latter_request,
4157                 .elevator_init_icq_fn =         cfq_init_icq,
4158                 .elevator_exit_icq_fn =         cfq_exit_icq,
4159                 .elevator_set_req_fn =          cfq_set_request,
4160                 .elevator_put_req_fn =          cfq_put_request,
4161                 .elevator_may_queue_fn =        cfq_may_queue,
4162                 .elevator_init_fn =             cfq_init_queue,
4163                 .elevator_exit_fn =             cfq_exit_queue,
4164         },
4165         .icq_size       =       sizeof(struct cfq_io_cq),
4166         .icq_align      =       __alignof__(struct cfq_io_cq),
4167         .elevator_attrs =       cfq_attrs,
4168         .elevator_name  =       "cfq",
4169         .elevator_owner =       THIS_MODULE,
4170 };
4171
4172 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4173 static struct blkcg_policy blkcg_policy_cfq = {
4174         .pd_size                = sizeof(struct cfq_group),
4175         .cftypes                = cfq_blkcg_files,
4176
4177         .pd_init_fn             = cfq_pd_init,
4178         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4179 };
4180 #endif
4181
4182 static int __init cfq_init(void)
4183 {
4184         int ret;
4185
4186         /*
4187          * could be 0 on HZ < 1000 setups
4188          */
4189         if (!cfq_slice_async)
4190                 cfq_slice_async = 1;
4191         if (!cfq_slice_idle)
4192                 cfq_slice_idle = 1;
4193
4194 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4195         if (!cfq_group_idle)
4196                 cfq_group_idle = 1;
4197
4198         ret = blkcg_policy_register(&blkcg_policy_cfq);
4199         if (ret)
4200                 return ret;
4201 #else
4202         cfq_group_idle = 0;
4203 #endif
4204
4205         ret = -ENOMEM;
4206         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4207         if (!cfq_pool)
4208                 goto err_pol_unreg;
4209
4210         ret = elv_register(&iosched_cfq);
4211         if (ret)
4212                 goto err_free_pool;
4213
4214         return 0;
4215
4216 err_free_pool:
4217         kmem_cache_destroy(cfq_pool);
4218 err_pol_unreg:
4219 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4220         blkcg_policy_unregister(&blkcg_policy_cfq);
4221 #endif
4222         return ret;
4223 }
4224
4225 static void __exit cfq_exit(void)
4226 {
4227 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4228         blkcg_policy_unregister(&blkcg_policy_cfq);
4229 #endif
4230         elv_unregister(&iosched_cfq);
4231         kmem_cache_destroy(cfq_pool);
4232 }
4233
4234 module_init(cfq_init);
4235 module_exit(cfq_exit);
4236
4237 MODULE_AUTHOR("Jens Axboe");
4238 MODULE_LICENSE("GPL");
4239 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");