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_prio_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_workload_slice;
 252         enum wl_type_t saved_workload;
 253         enum wl_prio_t saved_serving_prio;
 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_prio_t serving_prio;
 284         enum wl_type_t serving_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 *service_tree_for(struct cfq_group *cfqg,
 357                                             enum wl_prio_t prio,
 358                                             enum wl_type_t type)
 359 {
 360         if (!cfqg)
 361                 return NULL;
 362
 363         if (prio == IDLE_WORKLOAD)
 364                 return &cfqg->service_tree_idle;
 365
 366         return &cfqg->service_trees[prio][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_prio_t cfqq_prio(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_prio_t wl,
 755                                         struct cfq_data *cfqd,
 756                                         struct cfq_group *cfqg)
 757 {
 758         if (wl == IDLE_WORKLOAD)
 759                 return cfqg->service_tree_idle.count;
 760
 761         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
 762                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
 763                 + cfqg->service_trees[wl][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_workload_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_workload_slice = cfqd->workload_expires
1305                                                 - jiffies;
1306                 cfqg->saved_workload = cfqd->serving_type;
1307                 cfqg->saved_serving_prio = cfqd->serving_prio;
1308         } else
1309                 cfqg->saved_workload_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 *service_tree;
1616         int left;
1617         int new_cfqq = 1;
1618
1619         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1620                                                 cfqq_type(cfqq));
1621         if (cfq_class_idle(cfqq)) {
1622                 rb_key = CFQ_IDLE_DELAY;
1623                 parent = rb_last(&service_tree->rb);
1624                 if (parent && parent != &cfqq->rb_node) {
1625                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1626                         rb_key += __cfqq->rb_key;
1627                 } else
1628                         rb_key += jiffies;
1629         } else if (!add_front) {
1630                 /*
1631                  * Get our rb key offset. Subtract any residual slice
1632                  * value carried from last service. A negative resid
1633                  * count indicates slice overrun, and this should position
1634                  * the next service time further away in the tree.
1635                  */
1636                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1637                 rb_key -= cfqq->slice_resid;
1638                 cfqq->slice_resid = 0;
1639         } else {
1640                 rb_key = -HZ;
1641                 __cfqq = cfq_rb_first(service_tree);
1642                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1643         }
1644
1645         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1646                 new_cfqq = 0;
1647                 /*
1648                  * same position, nothing more to do
1649                  */
1650                 if (rb_key == cfqq->rb_key &&
1651                     cfqq->service_tree == service_tree)
1652                         return;
1653
1654                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1655                 cfqq->service_tree = NULL;
1656         }
1657
1658         left = 1;
1659         parent = NULL;
1660         cfqq->service_tree = service_tree;
1661         p = &service_tree->rb.rb_node;
1662         while (*p) {
1663                 struct rb_node **n;
1664
1665                 parent = *p;
1666                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1667
1668                 /*
1669                  * sort by key, that represents service time.
1670                  */
1671                 if (time_before(rb_key, __cfqq->rb_key))
1672                         n = &(*p)->rb_left;
1673                 else {
1674                         n = &(*p)->rb_right;
1675                         left = 0;
1676                 }
1677
1678                 p = n;
1679         }
1680
1681         if (left)
1682                 service_tree->left = &cfqq->rb_node;
1683
1684         cfqq->rb_key = rb_key;
1685         rb_link_node(&cfqq->rb_node, parent, p);
1686         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1687         service_tree->count++;
1688         if (add_front || !new_cfqq)
1689                 return;
1690         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1691 }
1692
1693 static struct cfq_queue *
1694 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1695                      sector_t sector, struct rb_node **ret_parent,
1696                      struct rb_node ***rb_link)
1697 {
1698         struct rb_node **p, *parent;
1699         struct cfq_queue *cfqq = NULL;
1700
1701         parent = NULL;
1702         p = &root->rb_node;
1703         while (*p) {
1704                 struct rb_node **n;
1705
1706                 parent = *p;
1707                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1708
1709                 /*
1710                  * Sort strictly based on sector.  Smallest to the left,
1711                  * largest to the right.
1712                  */
1713                 if (sector > blk_rq_pos(cfqq->next_rq))
1714                         n = &(*p)->rb_right;
1715                 else if (sector < blk_rq_pos(cfqq->next_rq))
1716                         n = &(*p)->rb_left;
1717                 else
1718                         break;
1719                 p = n;
1720                 cfqq = NULL;
1721         }
1722
1723         *ret_parent = parent;
1724         if (rb_link)
1725                 *rb_link = p;
1726         return cfqq;
1727 }
1728
1729 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1730 {
1731         struct rb_node **p, *parent;
1732         struct cfq_queue *__cfqq;
1733
1734         if (cfqq->p_root) {
1735                 rb_erase(&cfqq->p_node, cfqq->p_root);
1736                 cfqq->p_root = NULL;
1737         }
1738
1739         if (cfq_class_idle(cfqq))
1740                 return;
1741         if (!cfqq->next_rq)
1742                 return;
1743
1744         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1745         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1746                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1747         if (!__cfqq) {
1748                 rb_link_node(&cfqq->p_node, parent, p);
1749                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1750         } else
1751                 cfqq->p_root = NULL;
1752 }
1753
1754 /*
1755  * Update cfqq's position in the service tree.
1756  */
1757 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1758 {
1759         /*
1760          * Resorting requires the cfqq to be on the RR list already.
1761          */
1762         if (cfq_cfqq_on_rr(cfqq)) {
1763                 cfq_service_tree_add(cfqd, cfqq, 0);
1764                 cfq_prio_tree_add(cfqd, cfqq);
1765         }
1766 }
1767
1768 /*
1769  * add to busy list of queues for service, trying to be fair in ordering
1770  * the pending list according to last request service
1771  */
1772 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1773 {
1774         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1775         BUG_ON(cfq_cfqq_on_rr(cfqq));
1776         cfq_mark_cfqq_on_rr(cfqq);
1777         cfqd->busy_queues++;
1778         if (cfq_cfqq_sync(cfqq))
1779                 cfqd->busy_sync_queues++;
1780
1781         cfq_resort_rr_list(cfqd, cfqq);
1782 }
1783
1784 /*
1785  * Called when the cfqq no longer has requests pending, remove it from
1786  * the service tree.
1787  */
1788 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1789 {
1790         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1791         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1792         cfq_clear_cfqq_on_rr(cfqq);
1793
1794         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1795                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1796                 cfqq->service_tree = NULL;
1797         }
1798         if (cfqq->p_root) {
1799                 rb_erase(&cfqq->p_node, cfqq->p_root);
1800                 cfqq->p_root = NULL;
1801         }
1802
1803         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1804         BUG_ON(!cfqd->busy_queues);
1805         cfqd->busy_queues--;
1806         if (cfq_cfqq_sync(cfqq))
1807                 cfqd->busy_sync_queues--;
1808 }
1809
1810 /*
1811  * rb tree support functions
1812  */
1813 static void cfq_del_rq_rb(struct request *rq)
1814 {
1815         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1816         const int sync = rq_is_sync(rq);
1817
1818         BUG_ON(!cfqq->queued[sync]);
1819         cfqq->queued[sync]--;
1820
1821         elv_rb_del(&cfqq->sort_list, rq);
1822
1823         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1824                 /*
1825                  * Queue will be deleted from service tree when we actually
1826                  * expire it later. Right now just remove it from prio tree
1827                  * as it is empty.
1828                  */
1829                 if (cfqq->p_root) {
1830                         rb_erase(&cfqq->p_node, cfqq->p_root);
1831                         cfqq->p_root = NULL;
1832                 }
1833         }
1834 }
1835
1836 static void cfq_add_rq_rb(struct request *rq)
1837 {
1838         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1839         struct cfq_data *cfqd = cfqq->cfqd;
1840         struct request *prev;
1841
1842         cfqq->queued[rq_is_sync(rq)]++;
1843
1844         elv_rb_add(&cfqq->sort_list, rq);
1845
1846         if (!cfq_cfqq_on_rr(cfqq))
1847                 cfq_add_cfqq_rr(cfqd, cfqq);
1848
1849         /*
1850          * check if this request is a better next-serve candidate
1851          */
1852         prev = cfqq->next_rq;
1853         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1854
1855         /*
1856          * adjust priority tree position, if ->next_rq changes
1857          */
1858         if (prev != cfqq->next_rq)
1859                 cfq_prio_tree_add(cfqd, cfqq);
1860
1861         BUG_ON(!cfqq->next_rq);
1862 }
1863
1864 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1865 {
1866         elv_rb_del(&cfqq->sort_list, rq);
1867         cfqq->queued[rq_is_sync(rq)]--;
1868         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1869         cfq_add_rq_rb(rq);
1870         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
1871                                  rq->cmd_flags);
1872 }
1873
1874 static struct request *
1875 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1876 {
1877         struct task_struct *tsk = current;
1878         struct cfq_io_cq *cic;
1879         struct cfq_queue *cfqq;
1880
1881         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1882         if (!cic)
1883                 return NULL;
1884
1885         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1886         if (cfqq) {
1887                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1888
1889                 return elv_rb_find(&cfqq->sort_list, sector);
1890         }
1891
1892         return NULL;
1893 }
1894
1895 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1896 {
1897         struct cfq_data *cfqd = q->elevator->elevator_data;
1898
1899         cfqd->rq_in_driver++;
1900         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1901                                                 cfqd->rq_in_driver);
1902
1903         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1904 }
1905
1906 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1907 {
1908         struct cfq_data *cfqd = q->elevator->elevator_data;
1909
1910         WARN_ON(!cfqd->rq_in_driver);
1911         cfqd->rq_in_driver--;
1912         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1913                                                 cfqd->rq_in_driver);
1914 }
1915
1916 static void cfq_remove_request(struct request *rq)
1917 {
1918         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1919
1920         if (cfqq->next_rq == rq)
1921                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1922
1923         list_del_init(&rq->queuelist);
1924         cfq_del_rq_rb(rq);
1925
1926         cfqq->cfqd->rq_queued--;
1927         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
1928         if (rq->cmd_flags & REQ_PRIO) {
1929                 WARN_ON(!cfqq->prio_pending);
1930                 cfqq->prio_pending--;
1931         }
1932 }
1933
1934 static int cfq_merge(struct request_queue *q, struct request **req,
1935                      struct bio *bio)
1936 {
1937         struct cfq_data *cfqd = q->elevator->elevator_data;
1938         struct request *__rq;
1939
1940         __rq = cfq_find_rq_fmerge(cfqd, bio);
1941         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1942                 *req = __rq;
1943                 return ELEVATOR_FRONT_MERGE;
1944         }
1945
1946         return ELEVATOR_NO_MERGE;
1947 }
1948
1949 static void cfq_merged_request(struct request_queue *q, struct request *req,
1950                                int type)
1951 {
1952         if (type == ELEVATOR_FRONT_MERGE) {
1953                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1954
1955                 cfq_reposition_rq_rb(cfqq, req);
1956         }
1957 }
1958
1959 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1960                                 struct bio *bio)
1961 {
1962         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
1963 }
1964
1965 static void
1966 cfq_merged_requests(struct request_queue *q, struct request *rq,
1967                     struct request *next)
1968 {
1969         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1970         struct cfq_data *cfqd = q->elevator->elevator_data;
1971
1972         /*
1973          * reposition in fifo if next is older than rq
1974          */
1975         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1976             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1977                 list_move(&rq->queuelist, &next->queuelist);
1978                 rq_set_fifo_time(rq, rq_fifo_time(next));
1979         }
1980
1981         if (cfqq->next_rq == next)
1982                 cfqq->next_rq = rq;
1983         cfq_remove_request(next);
1984         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
1985
1986         cfqq = RQ_CFQQ(next);
1987         /*
1988          * all requests of this queue are merged to other queues, delete it
1989          * from the service tree. If it's the active_queue,
1990          * cfq_dispatch_requests() will choose to expire it or do idle
1991          */
1992         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1993             cfqq != cfqd->active_queue)
1994                 cfq_del_cfqq_rr(cfqd, cfqq);
1995 }
1996
1997 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1998                            struct bio *bio)
1999 {
2000         struct cfq_data *cfqd = q->elevator->elevator_data;
2001         struct cfq_io_cq *cic;
2002         struct cfq_queue *cfqq;
2003
2004         /*
2005          * Disallow merge of a sync bio into an async request.
2006          */
2007         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2008                 return false;
2009
2010         /*
2011          * Lookup the cfqq that this bio will be queued with and allow
2012          * merge only if rq is queued there.
2013          */
2014         cic = cfq_cic_lookup(cfqd, current->io_context);
2015         if (!cic)
2016                 return false;
2017
2018         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2019         return cfqq == RQ_CFQQ(rq);
2020 }
2021
2022 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2023 {
2024         del_timer(&cfqd->idle_slice_timer);
2025         cfqg_stats_update_idle_time(cfqq->cfqg);
2026 }
2027
2028 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2029                                    struct cfq_queue *cfqq)
2030 {
2031         if (cfqq) {
2032                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
2033                                 cfqd->serving_prio, cfqd->serving_type);
2034                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2035                 cfqq->slice_start = 0;
2036                 cfqq->dispatch_start = jiffies;
2037                 cfqq->allocated_slice = 0;
2038                 cfqq->slice_end = 0;
2039                 cfqq->slice_dispatch = 0;
2040                 cfqq->nr_sectors = 0;
2041
2042                 cfq_clear_cfqq_wait_request(cfqq);
2043                 cfq_clear_cfqq_must_dispatch(cfqq);
2044                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2045                 cfq_clear_cfqq_fifo_expire(cfqq);
2046                 cfq_mark_cfqq_slice_new(cfqq);
2047
2048                 cfq_del_timer(cfqd, cfqq);
2049         }
2050
2051         cfqd->active_queue = cfqq;
2052 }
2053
2054 /*
2055  * current cfqq expired its slice (or was too idle), select new one
2056  */
2057 static void
2058 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2059                     bool timed_out)
2060 {
2061         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2062
2063         if (cfq_cfqq_wait_request(cfqq))
2064                 cfq_del_timer(cfqd, cfqq);
2065
2066         cfq_clear_cfqq_wait_request(cfqq);
2067         cfq_clear_cfqq_wait_busy(cfqq);
2068
2069         /*
2070          * If this cfqq is shared between multiple processes, check to
2071          * make sure that those processes are still issuing I/Os within
2072          * the mean seek distance.  If not, it may be time to break the
2073          * queues apart again.
2074          */
2075         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2076                 cfq_mark_cfqq_split_coop(cfqq);
2077
2078         /*
2079          * store what was left of this slice, if the queue idled/timed out
2080          */
2081         if (timed_out) {
2082                 if (cfq_cfqq_slice_new(cfqq))
2083                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2084                 else
2085                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2086                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2087         }
2088
2089         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2090
2091         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2092                 cfq_del_cfqq_rr(cfqd, cfqq);
2093
2094         cfq_resort_rr_list(cfqd, cfqq);
2095
2096         if (cfqq == cfqd->active_queue)
2097                 cfqd->active_queue = NULL;
2098
2099         if (cfqd->active_cic) {
2100                 put_io_context(cfqd->active_cic->icq.ioc);
2101                 cfqd->active_cic = NULL;
2102         }
2103 }
2104
2105 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2106 {
2107         struct cfq_queue *cfqq = cfqd->active_queue;
2108
2109         if (cfqq)
2110                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2111 }
2112
2113 /*
2114  * Get next queue for service. Unless we have a queue preemption,
2115  * we'll simply select the first cfqq in the service tree.
2116  */
2117 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2118 {
2119         struct cfq_rb_root *service_tree =
2120                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
2121                                         cfqd->serving_type);
2122
2123         if (!cfqd->rq_queued)
2124                 return NULL;
2125
2126         /* There is nothing to dispatch */
2127         if (!service_tree)
2128                 return NULL;
2129         if (RB_EMPTY_ROOT(&service_tree->rb))
2130                 return NULL;
2131         return cfq_rb_first(service_tree);
2132 }
2133
2134 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2135 {
2136         struct cfq_group *cfqg;
2137         struct cfq_queue *cfqq;
2138         int i, j;
2139         struct cfq_rb_root *st;
2140
2141         if (!cfqd->rq_queued)
2142                 return NULL;
2143
2144         cfqg = cfq_get_next_cfqg(cfqd);
2145         if (!cfqg)
2146                 return NULL;
2147
2148         for_each_cfqg_st(cfqg, i, j, st)
2149                 if ((cfqq = cfq_rb_first(st)) != NULL)
2150                         return cfqq;
2151         return NULL;
2152 }
2153
2154 /*
2155  * Get and set a new active queue for service.
2156  */
2157 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2158                                               struct cfq_queue *cfqq)
2159 {
2160         if (!cfqq)
2161                 cfqq = cfq_get_next_queue(cfqd);
2162
2163         __cfq_set_active_queue(cfqd, cfqq);
2164         return cfqq;
2165 }
2166
2167 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2168                                           struct request *rq)
2169 {
2170         if (blk_rq_pos(rq) >= cfqd->last_position)
2171                 return blk_rq_pos(rq) - cfqd->last_position;
2172         else
2173                 return cfqd->last_position - blk_rq_pos(rq);
2174 }
2175
2176 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2177                                struct request *rq)
2178 {
2179         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2180 }
2181
2182 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2183                                     struct cfq_queue *cur_cfqq)
2184 {
2185         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2186         struct rb_node *parent, *node;
2187         struct cfq_queue *__cfqq;
2188         sector_t sector = cfqd->last_position;
2189
2190         if (RB_EMPTY_ROOT(root))
2191                 return NULL;
2192
2193         /*
2194          * First, if we find a request starting at the end of the last
2195          * request, choose it.
2196          */
2197         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2198         if (__cfqq)
2199                 return __cfqq;
2200
2201         /*
2202          * If the exact sector wasn't found, the parent of the NULL leaf
2203          * will contain the closest sector.
2204          */
2205         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2206         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2207                 return __cfqq;
2208
2209         if (blk_rq_pos(__cfqq->next_rq) < sector)
2210                 node = rb_next(&__cfqq->p_node);
2211         else
2212                 node = rb_prev(&__cfqq->p_node);
2213         if (!node)
2214                 return NULL;
2215
2216         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2217         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2218                 return __cfqq;
2219
2220         return NULL;
2221 }
2222
2223 /*
2224  * cfqd - obvious
2225  * cur_cfqq - passed in so that we don't decide that the current queue is
2226  *            closely cooperating with itself.
2227  *
2228  * So, basically we're assuming that that cur_cfqq has dispatched at least
2229  * one request, and that cfqd->last_position reflects a position on the disk
2230  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2231  * assumption.
2232  */
2233 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2234                                               struct cfq_queue *cur_cfqq)
2235 {
2236         struct cfq_queue *cfqq;
2237
2238         if (cfq_class_idle(cur_cfqq))
2239                 return NULL;
2240         if (!cfq_cfqq_sync(cur_cfqq))
2241                 return NULL;
2242         if (CFQQ_SEEKY(cur_cfqq))
2243                 return NULL;
2244
2245         /*
2246          * Don't search priority tree if it's the only queue in the group.
2247          */
2248         if (cur_cfqq->cfqg->nr_cfqq == 1)
2249                 return NULL;
2250
2251         /*
2252          * We should notice if some of the queues are cooperating, eg
2253          * working closely on the same area of the disk. In that case,
2254          * we can group them together and don't waste time idling.
2255          */
2256         cfqq = cfqq_close(cfqd, cur_cfqq);
2257         if (!cfqq)
2258                 return NULL;
2259
2260         /* If new queue belongs to different cfq_group, don't choose it */
2261         if (cur_cfqq->cfqg != cfqq->cfqg)
2262                 return NULL;
2263
2264         /*
2265          * It only makes sense to merge sync queues.
2266          */
2267         if (!cfq_cfqq_sync(cfqq))
2268                 return NULL;
2269         if (CFQQ_SEEKY(cfqq))
2270                 return NULL;
2271
2272         /*
2273          * Do not merge queues of different priority classes
2274          */
2275         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2276                 return NULL;
2277
2278         return cfqq;
2279 }
2280
2281 /*
2282  * Determine whether we should enforce idle window for this queue.
2283  */
2284
2285 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2286 {
2287         enum wl_prio_t prio = cfqq_prio(cfqq);
2288         struct cfq_rb_root *service_tree = cfqq->service_tree;
2289
2290         BUG_ON(!service_tree);
2291         BUG_ON(!service_tree->count);
2292
2293         if (!cfqd->cfq_slice_idle)
2294                 return false;
2295
2296         /* We never do for idle class queues. */
2297         if (prio == IDLE_WORKLOAD)
2298                 return false;
2299
2300         /* We do for queues that were marked with idle window flag. */
2301         if (cfq_cfqq_idle_window(cfqq) &&
2302            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2303                 return true;
2304
2305         /*
2306          * Otherwise, we do only if they are the last ones
2307          * in their service tree.
2308          */
2309         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2310            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2311                 return true;
2312         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2313                         service_tree->count);
2314         return false;
2315 }
2316
2317 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2318 {
2319         struct cfq_queue *cfqq = cfqd->active_queue;
2320         struct cfq_io_cq *cic;
2321         unsigned long sl, group_idle = 0;
2322
2323         /*
2324          * SSD device without seek penalty, disable idling. But only do so
2325          * for devices that support queuing, otherwise we still have a problem
2326          * with sync vs async workloads.
2327          */
2328         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2329                 return;
2330
2331         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2332         WARN_ON(cfq_cfqq_slice_new(cfqq));
2333
2334         /*
2335          * idle is disabled, either manually or by past process history
2336          */
2337         if (!cfq_should_idle(cfqd, cfqq)) {
2338                 /* no queue idling. Check for group idling */
2339                 if (cfqd->cfq_group_idle)
2340                         group_idle = cfqd->cfq_group_idle;
2341                 else
2342                         return;
2343         }
2344
2345         /*
2346          * still active requests from this queue, don't idle
2347          */
2348         if (cfqq->dispatched)
2349                 return;
2350
2351         /*
2352          * task has exited, don't wait
2353          */
2354         cic = cfqd->active_cic;
2355         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2356                 return;
2357
2358         /*
2359          * If our average think time is larger than the remaining time
2360          * slice, then don't idle. This avoids overrunning the allotted
2361          * time slice.
2362          */
2363         if (sample_valid(cic->ttime.ttime_samples) &&
2364             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2365                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2366                              cic->ttime.ttime_mean);
2367                 return;
2368         }
2369
2370         /* There are other queues in the group, don't do group idle */
2371         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2372                 return;
2373
2374         cfq_mark_cfqq_wait_request(cfqq);
2375
2376         if (group_idle)
2377                 sl = cfqd->cfq_group_idle;
2378         else
2379                 sl = cfqd->cfq_slice_idle;
2380
2381         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2382         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2383         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2384                         group_idle ? 1 : 0);
2385 }
2386
2387 /*
2388  * Move request from internal lists to the request queue dispatch list.
2389  */
2390 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2391 {
2392         struct cfq_data *cfqd = q->elevator->elevator_data;
2393         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2394
2395         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2396
2397         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2398         cfq_remove_request(rq);
2399         cfqq->dispatched++;
2400         (RQ_CFQG(rq))->dispatched++;
2401         elv_dispatch_sort(q, rq);
2402
2403         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2404         cfqq->nr_sectors += blk_rq_sectors(rq);
2405         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2406 }
2407
2408 /*
2409  * return expired entry, or NULL to just start from scratch in rbtree
2410  */
2411 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2412 {
2413         struct request *rq = NULL;
2414
2415         if (cfq_cfqq_fifo_expire(cfqq))
2416                 return NULL;
2417
2418         cfq_mark_cfqq_fifo_expire(cfqq);
2419
2420         if (list_empty(&cfqq->fifo))
2421                 return NULL;
2422
2423         rq = rq_entry_fifo(cfqq->fifo.next);
2424         if (time_before(jiffies, rq_fifo_time(rq)))
2425                 rq = NULL;
2426
2427         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2428         return rq;
2429 }
2430
2431 static inline int
2432 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2433 {
2434         const int base_rq = cfqd->cfq_slice_async_rq;
2435
2436         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2437
2438         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2439 }
2440
2441 /*
2442  * Must be called with the queue_lock held.
2443  */
2444 static int cfqq_process_refs(struct cfq_queue *cfqq)
2445 {
2446         int process_refs, io_refs;
2447
2448         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2449         process_refs = cfqq->ref - io_refs;
2450         BUG_ON(process_refs < 0);
2451         return process_refs;
2452 }
2453
2454 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2455 {
2456         int process_refs, new_process_refs;
2457         struct cfq_queue *__cfqq;
2458
2459         /*
2460          * If there are no process references on the new_cfqq, then it is
2461          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2462          * chain may have dropped their last reference (not just their
2463          * last process reference).
2464          */
2465         if (!cfqq_process_refs(new_cfqq))
2466                 return;
2467
2468         /* Avoid a circular list and skip interim queue merges */
2469         while ((__cfqq = new_cfqq->new_cfqq)) {
2470                 if (__cfqq == cfqq)
2471                         return;
2472                 new_cfqq = __cfqq;
2473         }
2474
2475         process_refs = cfqq_process_refs(cfqq);
2476         new_process_refs = cfqq_process_refs(new_cfqq);
2477         /*
2478          * If the process for the cfqq has gone away, there is no
2479          * sense in merging the queues.
2480          */
2481         if (process_refs == 0 || new_process_refs == 0)
2482                 return;
2483
2484         /*
2485          * Merge in the direction of the lesser amount of work.
2486          */
2487         if (new_process_refs >= process_refs) {
2488                 cfqq->new_cfqq = new_cfqq;
2489                 new_cfqq->ref += process_refs;
2490         } else {
2491                 new_cfqq->new_cfqq = cfqq;
2492                 cfqq->ref += new_process_refs;
2493         }
2494 }
2495
2496 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2497                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2498 {
2499         struct cfq_queue *queue;
2500         int i;
2501         bool key_valid = false;
2502         unsigned long lowest_key = 0;
2503         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2504
2505         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2506                 /* select the one with lowest rb_key */
2507                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2508                 if (queue &&
2509                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2510                         lowest_key = queue->rb_key;
2511                         cur_best = i;
2512                         key_valid = true;
2513                 }
2514         }
2515
2516         return cur_best;
2517 }
2518
2519 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2520 {
2521         unsigned slice;
2522         unsigned count;
2523         struct cfq_rb_root *st;
2524         unsigned group_slice;
2525         enum wl_prio_t original_prio = cfqd->serving_prio;
2526
2527         /* Choose next priority. RT > BE > IDLE */
2528         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2529                 cfqd->serving_prio = RT_WORKLOAD;
2530         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2531                 cfqd->serving_prio = BE_WORKLOAD;
2532         else {
2533                 cfqd->serving_prio = IDLE_WORKLOAD;
2534                 cfqd->workload_expires = jiffies + 1;
2535                 return;
2536         }
2537
2538         if (original_prio != cfqd->serving_prio)
2539                 goto new_workload;
2540
2541         /*
2542          * For RT and BE, we have to choose also the type
2543          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2544          * expiration time
2545          */
2546         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2547         count = st->count;
2548
2549         /*
2550          * check workload expiration, and that we still have other queues ready
2551          */
2552         if (count && !time_after(jiffies, cfqd->workload_expires))
2553                 return;
2554
2555 new_workload:
2556         /* otherwise select new workload type */
2557         cfqd->serving_type =
2558                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2559         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2560         count = st->count;
2561
2562         /*
2563          * the workload slice is computed as a fraction of target latency
2564          * proportional to the number of queues in that workload, over
2565          * all the queues in the same priority class
2566          */
2567         group_slice = cfq_group_slice(cfqd, cfqg);
2568
2569         slice = group_slice * count /
2570                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2571                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2572
2573         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2574                 unsigned int tmp;
2575
2576                 /*
2577                  * Async queues are currently system wide. Just taking
2578                  * proportion of queues with-in same group will lead to higher
2579                  * async ratio system wide as generally root group is going
2580                  * to have higher weight. A more accurate thing would be to
2581                  * calculate system wide asnc/sync ratio.
2582                  */
2583                 tmp = cfqd->cfq_target_latency *
2584                         cfqg_busy_async_queues(cfqd, cfqg);
2585                 tmp = tmp/cfqd->busy_queues;
2586                 slice = min_t(unsigned, slice, tmp);
2587
2588                 /* async workload slice is scaled down according to
2589                  * the sync/async slice ratio. */
2590                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2591         } else
2592                 /* sync workload slice is at least 2 * cfq_slice_idle */
2593                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2594
2595         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2596         cfq_log(cfqd, "workload slice:%d", slice);
2597         cfqd->workload_expires = jiffies + slice;
2598 }
2599
2600 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2601 {
2602         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2603         struct cfq_group *cfqg;
2604
2605         if (RB_EMPTY_ROOT(&st->rb))
2606                 return NULL;
2607         cfqg = cfq_rb_first_group(st);
2608         update_min_vdisktime(st);
2609         return cfqg;
2610 }
2611
2612 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2613 {
2614         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2615
2616         cfqd->serving_group = cfqg;
2617
2618         /* Restore the workload type data */
2619         if (cfqg->saved_workload_slice) {
2620                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2621                 cfqd->serving_type = cfqg->saved_workload;
2622                 cfqd->serving_prio = cfqg->saved_serving_prio;
2623         } else
2624                 cfqd->workload_expires = jiffies - 1;
2625
2626         choose_service_tree(cfqd, cfqg);
2627 }
2628
2629 /*
2630  * Select a queue for service. If we have a current active queue,
2631  * check whether to continue servicing it, or retrieve and set a new one.
2632  */
2633 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2634 {
2635         struct cfq_queue *cfqq, *new_cfqq = NULL;
2636
2637         cfqq = cfqd->active_queue;
2638         if (!cfqq)
2639                 goto new_queue;
2640
2641         if (!cfqd->rq_queued)
2642                 return NULL;
2643
2644         /*
2645          * We were waiting for group to get backlogged. Expire the queue
2646          */
2647         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2648                 goto expire;
2649
2650         /*
2651          * The active queue has run out of time, expire it and select new.
2652          */
2653         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2654                 /*
2655                  * If slice had not expired at the completion of last request
2656                  * we might not have turned on wait_busy flag. Don't expire
2657                  * the queue yet. Allow the group to get backlogged.
2658                  *
2659                  * The very fact that we have used the slice, that means we
2660                  * have been idling all along on this queue and it should be
2661                  * ok to wait for this request to complete.
2662                  */
2663                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2664                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2665                         cfqq = NULL;
2666                         goto keep_queue;
2667                 } else
2668                         goto check_group_idle;
2669         }
2670
2671         /*
2672          * The active queue has requests and isn't expired, allow it to
2673          * dispatch.
2674          */
2675         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2676                 goto keep_queue;
2677
2678         /*
2679          * If another queue has a request waiting within our mean seek
2680          * distance, let it run.  The expire code will check for close
2681          * cooperators and put the close queue at the front of the service
2682          * tree.  If possible, merge the expiring queue with the new cfqq.
2683          */
2684         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2685         if (new_cfqq) {
2686                 if (!cfqq->new_cfqq)
2687                         cfq_setup_merge(cfqq, new_cfqq);
2688                 goto expire;
2689         }
2690
2691         /*
2692          * No requests pending. If the active queue still has requests in
2693          * flight or is idling for a new request, allow either of these
2694          * conditions to happen (or time out) before selecting a new queue.
2695          */
2696         if (timer_pending(&cfqd->idle_slice_timer)) {
2697                 cfqq = NULL;
2698                 goto keep_queue;
2699         }
2700
2701         /*
2702          * This is a deep seek queue, but the device is much faster than
2703          * the queue can deliver, don't idle
2704          **/
2705         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2706             (cfq_cfqq_slice_new(cfqq) ||
2707             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2708                 cfq_clear_cfqq_deep(cfqq);
2709                 cfq_clear_cfqq_idle_window(cfqq);
2710         }
2711
2712         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2713                 cfqq = NULL;
2714                 goto keep_queue;
2715         }
2716
2717         /*
2718          * If group idle is enabled and there are requests dispatched from
2719          * this group, wait for requests to complete.
2720          */
2721 check_group_idle:
2722         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2723             cfqq->cfqg->dispatched &&
2724             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2725                 cfqq = NULL;
2726                 goto keep_queue;
2727         }
2728
2729 expire:
2730         cfq_slice_expired(cfqd, 0);
2731 new_queue:
2732         /*
2733          * Current queue expired. Check if we have to switch to a new
2734          * service tree
2735          */
2736         if (!new_cfqq)
2737                 cfq_choose_cfqg(cfqd);
2738
2739         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2740 keep_queue:
2741         return cfqq;
2742 }
2743
2744 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2745 {
2746         int dispatched = 0;
2747
2748         while (cfqq->next_rq) {
2749                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2750                 dispatched++;
2751         }
2752
2753         BUG_ON(!list_empty(&cfqq->fifo));
2754
2755         /* By default cfqq is not expired if it is empty. Do it explicitly */
2756         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2757         return dispatched;
2758 }
2759
2760 /*
2761  * Drain our current requests. Used for barriers and when switching
2762  * io schedulers on-the-fly.
2763  */
2764 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2765 {
2766         struct cfq_queue *cfqq;
2767         int dispatched = 0;
2768
2769         /* Expire the timeslice of the current active queue first */
2770         cfq_slice_expired(cfqd, 0);
2771         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2772                 __cfq_set_active_queue(cfqd, cfqq);
2773                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2774         }
2775
2776         BUG_ON(cfqd->busy_queues);
2777
2778         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2779         return dispatched;
2780 }
2781
2782 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2783         struct cfq_queue *cfqq)
2784 {
2785         /* the queue hasn't finished any request, can't estimate */
2786         if (cfq_cfqq_slice_new(cfqq))
2787                 return true;
2788         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2789                 cfqq->slice_end))
2790                 return true;
2791
2792         return false;
2793 }
2794
2795 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2796 {
2797         unsigned int max_dispatch;
2798
2799         /*
2800          * Drain async requests before we start sync IO
2801          */
2802         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2803                 return false;
2804
2805         /*
2806          * If this is an async queue and we have sync IO in flight, let it wait
2807          */
2808         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2809                 return false;
2810
2811         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2812         if (cfq_class_idle(cfqq))
2813                 max_dispatch = 1;
2814
2815         /*
2816          * Does this cfqq already have too much IO in flight?
2817          */
2818         if (cfqq->dispatched >= max_dispatch) {
2819                 bool promote_sync = false;
2820                 /*
2821                  * idle queue must always only have a single IO in flight
2822                  */
2823                 if (cfq_class_idle(cfqq))
2824                         return false;
2825
2826                 /*
2827                  * If there is only one sync queue
2828                  * we can ignore async queue here and give the sync
2829                  * queue no dispatch limit. The reason is a sync queue can
2830                  * preempt async queue, limiting the sync queue doesn't make
2831                  * sense. This is useful for aiostress test.
2832                  */
2833                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2834                         promote_sync = true;
2835
2836                 /*
2837                  * We have other queues, don't allow more IO from this one
2838                  */
2839                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2840                                 !promote_sync)
2841                         return false;
2842
2843                 /*
2844                  * Sole queue user, no limit
2845                  */
2846                 if (cfqd->busy_queues == 1 || promote_sync)
2847                         max_dispatch = -1;
2848                 else
2849                         /*
2850                          * Normally we start throttling cfqq when cfq_quantum/2
2851                          * requests have been dispatched. But we can drive
2852                          * deeper queue depths at the beginning of slice
2853                          * subjected to upper limit of cfq_quantum.
2854                          * */
2855                         max_dispatch = cfqd->cfq_quantum;
2856         }
2857
2858         /*
2859          * Async queues must wait a bit before being allowed dispatch.
2860          * We also ramp up the dispatch depth gradually for async IO,
2861          * based on the last sync IO we serviced
2862          */
2863         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2864                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2865                 unsigned int depth;
2866
2867                 depth = last_sync / cfqd->cfq_slice[1];
2868                 if (!depth && !cfqq->dispatched)
2869                         depth = 1;
2870                 if (depth < max_dispatch)
2871                         max_dispatch = depth;
2872         }
2873
2874         /*
2875          * If we're below the current max, allow a dispatch
2876          */
2877         return cfqq->dispatched < max_dispatch;
2878 }
2879
2880 /*
2881  * Dispatch a request from cfqq, moving them to the request queue
2882  * dispatch list.
2883  */
2884 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2885 {
2886         struct request *rq;
2887
2888         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2889
2890         if (!cfq_may_dispatch(cfqd, cfqq))
2891                 return false;
2892
2893         /*
2894          * follow expired path, else get first next available
2895          */
2896         rq = cfq_check_fifo(cfqq);
2897         if (!rq)
2898                 rq = cfqq->next_rq;
2899
2900         /*
2901          * insert request into driver dispatch list
2902          */
2903         cfq_dispatch_insert(cfqd->queue, rq);
2904
2905         if (!cfqd->active_cic) {
2906                 struct cfq_io_cq *cic = RQ_CIC(rq);
2907
2908                 atomic_long_inc(&cic->icq.ioc->refcount);
2909                 cfqd->active_cic = cic;
2910         }
2911
2912         return true;
2913 }
2914
2915 /*
2916  * Find the cfqq that we need to service and move a request from that to the
2917  * dispatch list
2918  */
2919 static int cfq_dispatch_requests(struct request_queue *q, int force)
2920 {
2921         struct cfq_data *cfqd = q->elevator->elevator_data;
2922         struct cfq_queue *cfqq;
2923
2924         if (!cfqd->busy_queues)
2925                 return 0;
2926
2927         if (unlikely(force))
2928                 return cfq_forced_dispatch(cfqd);
2929
2930         cfqq = cfq_select_queue(cfqd);
2931         if (!cfqq)
2932                 return 0;
2933
2934         /*
2935          * Dispatch a request from this cfqq, if it is allowed
2936          */
2937         if (!cfq_dispatch_request(cfqd, cfqq))
2938                 return 0;
2939
2940         cfqq->slice_dispatch++;
2941         cfq_clear_cfqq_must_dispatch(cfqq);
2942
2943         /*
2944          * expire an async queue immediately if it has used up its slice. idle
2945          * queue always expire after 1 dispatch round.
2946          */
2947         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2948             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2949             cfq_class_idle(cfqq))) {
2950                 cfqq->slice_end = jiffies + 1;
2951                 cfq_slice_expired(cfqd, 0);
2952         }
2953
2954         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2955         return 1;
2956 }
2957
2958 /*
2959  * task holds one reference to the queue, dropped when task exits. each rq
2960  * in-flight on this queue also holds a reference, dropped when rq is freed.
2961  *
2962  * Each cfq queue took a reference on the parent group. Drop it now.
2963  * queue lock must be held here.
2964  */
2965 static void cfq_put_queue(struct cfq_queue *cfqq)
2966 {
2967         struct cfq_data *cfqd = cfqq->cfqd;
2968         struct cfq_group *cfqg;
2969
2970         BUG_ON(cfqq->ref <= 0);
2971
2972         cfqq->ref--;
2973         if (cfqq->ref)
2974                 return;
2975
2976         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2977         BUG_ON(rb_first(&cfqq->sort_list));
2978         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2979         cfqg = cfqq->cfqg;
2980
2981         if (unlikely(cfqd->active_queue == cfqq)) {
2982                 __cfq_slice_expired(cfqd, cfqq, 0);
2983                 cfq_schedule_dispatch(cfqd);
2984         }
2985
2986         BUG_ON(cfq_cfqq_on_rr(cfqq));
2987         kmem_cache_free(cfq_pool, cfqq);
2988         cfqg_put(cfqg);
2989 }
2990
2991 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2992 {
2993         struct cfq_queue *__cfqq, *next;
2994
2995         /*
2996          * If this queue was scheduled to merge with another queue, be
2997          * sure to drop the reference taken on that queue (and others in
2998          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2999          */
3000         __cfqq = cfqq->new_cfqq;
3001         while (__cfqq) {
3002                 if (__cfqq == cfqq) {
3003                         WARN(1, "cfqq->new_cfqq loop detected\n");
3004                         break;
3005                 }
3006                 next = __cfqq->new_cfqq;
3007                 cfq_put_queue(__cfqq);
3008                 __cfqq = next;
3009         }
3010 }
3011
3012 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3013 {
3014         if (unlikely(cfqq == cfqd->active_queue)) {
3015                 __cfq_slice_expired(cfqd, cfqq, 0);
3016                 cfq_schedule_dispatch(cfqd);
3017         }
3018
3019         cfq_put_cooperator(cfqq);
3020
3021         cfq_put_queue(cfqq);
3022 }
3023
3024 static void cfq_init_icq(struct io_cq *icq)
3025 {
3026         struct cfq_io_cq *cic = icq_to_cic(icq);
3027
3028         cic->ttime.last_end_request = jiffies;
3029 }
3030
3031 static void cfq_exit_icq(struct io_cq *icq)
3032 {
3033         struct cfq_io_cq *cic = icq_to_cic(icq);
3034         struct cfq_data *cfqd = cic_to_cfqd(cic);
3035
3036         if (cic->cfqq[BLK_RW_ASYNC]) {
3037                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3038                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3039         }
3040
3041         if (cic->cfqq[BLK_RW_SYNC]) {
3042                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3043                 cic->cfqq[BLK_RW_SYNC] = NULL;
3044         }
3045 }
3046
3047 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3048 {
3049         struct task_struct *tsk = current;
3050         int ioprio_class;
3051
3052         if (!cfq_cfqq_prio_changed(cfqq))
3053                 return;
3054
3055         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3056         switch (ioprio_class) {
3057         default:
3058                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3059         case IOPRIO_CLASS_NONE:
3060                 /*
3061                  * no prio set, inherit CPU scheduling settings
3062                  */
3063                 cfqq->ioprio = task_nice_ioprio(tsk);
3064                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3065                 break;
3066         case IOPRIO_CLASS_RT:
3067                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3068                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3069                 break;
3070         case IOPRIO_CLASS_BE:
3071                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3072                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3073                 break;
3074         case IOPRIO_CLASS_IDLE:
3075                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3076                 cfqq->ioprio = 7;
3077                 cfq_clear_cfqq_idle_window(cfqq);
3078                 break;
3079         }
3080
3081         /*
3082          * keep track of original prio settings in case we have to temporarily
3083          * elevate the priority of this queue
3084          */
3085         cfqq->org_ioprio = cfqq->ioprio;
3086         cfq_clear_cfqq_prio_changed(cfqq);
3087 }
3088
3089 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3090 {
3091         int ioprio = cic->icq.ioc->ioprio;
3092         struct cfq_data *cfqd = cic_to_cfqd(cic);
3093         struct cfq_queue *cfqq;
3094
3095         /*
3096          * Check whether ioprio has changed.  The condition may trigger
3097          * spuriously on a newly created cic but there's no harm.
3098          */
3099         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3100                 return;
3101
3102         cfqq = cic->cfqq[BLK_RW_ASYNC];
3103         if (cfqq) {
3104                 struct cfq_queue *new_cfqq;
3105                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3106                                          GFP_ATOMIC);
3107                 if (new_cfqq) {
3108                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3109                         cfq_put_queue(cfqq);
3110                 }
3111         }
3112
3113         cfqq = cic->cfqq[BLK_RW_SYNC];
3114         if (cfqq)
3115                 cfq_mark_cfqq_prio_changed(cfqq);
3116
3117         cic->ioprio = ioprio;
3118 }
3119
3120 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3121                           pid_t pid, bool is_sync)
3122 {
3123         RB_CLEAR_NODE(&cfqq->rb_node);
3124         RB_CLEAR_NODE(&cfqq->p_node);
3125         INIT_LIST_HEAD(&cfqq->fifo);
3126
3127         cfqq->ref = 0;
3128         cfqq->cfqd = cfqd;
3129
3130         cfq_mark_cfqq_prio_changed(cfqq);
3131
3132         if (is_sync) {
3133                 if (!cfq_class_idle(cfqq))
3134                         cfq_mark_cfqq_idle_window(cfqq);
3135                 cfq_mark_cfqq_sync(cfqq);
3136         }
3137         cfqq->pid = pid;
3138 }
3139
3140 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3141 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3142 {
3143         struct cfq_data *cfqd = cic_to_cfqd(cic);
3144         struct cfq_queue *sync_cfqq;
3145         uint64_t id;
3146
3147         rcu_read_lock();
3148         id = bio_blkcg(bio)->id;
3149         rcu_read_unlock();
3150
3151         /*
3152          * Check whether blkcg has changed.  The condition may trigger
3153          * spuriously on a newly created cic but there's no harm.
3154          */
3155         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3156                 return;
3157
3158         sync_cfqq = cic_to_cfqq(cic, 1);
3159         if (sync_cfqq) {
3160                 /*
3161                  * Drop reference to sync queue. A new sync queue will be
3162                  * assigned in new group upon arrival of a fresh request.
3163                  */
3164                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3165                 cic_set_cfqq(cic, NULL, 1);
3166                 cfq_put_queue(sync_cfqq);
3167         }
3168
3169         cic->blkcg_id = id;
3170 }
3171 #else
3172 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3173 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3174
3175 static struct cfq_queue *
3176 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3177                      struct bio *bio, gfp_t gfp_mask)
3178 {
3179         struct blkcg *blkcg;
3180         struct cfq_queue *cfqq, *new_cfqq = NULL;
3181         struct cfq_group *cfqg;
3182
3183 retry:
3184         rcu_read_lock();
3185
3186         blkcg = bio_blkcg(bio);
3187         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3188         cfqq = cic_to_cfqq(cic, is_sync);
3189
3190         /*
3191          * Always try a new alloc if we fell back to the OOM cfqq
3192          * originally, since it should just be a temporary situation.
3193          */
3194         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3195                 cfqq = NULL;
3196                 if (new_cfqq) {
3197                         cfqq = new_cfqq;
3198                         new_cfqq = NULL;
3199                 } else if (gfp_mask & __GFP_WAIT) {
3200                         rcu_read_unlock();
3201                         spin_unlock_irq(cfqd->queue->queue_lock);
3202                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3203                                         gfp_mask | __GFP_ZERO,
3204                                         cfqd->queue->node);
3205                         spin_lock_irq(cfqd->queue->queue_lock);
3206                         if (new_cfqq)
3207                                 goto retry;
3208                 } else {
3209                         cfqq = kmem_cache_alloc_node(cfq_pool,
3210                                         gfp_mask | __GFP_ZERO,
3211                                         cfqd->queue->node);
3212                 }
3213
3214                 if (cfqq) {
3215                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3216                         cfq_init_prio_data(cfqq, cic);
3217                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3218                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3219                 } else
3220                         cfqq = &cfqd->oom_cfqq;
3221         }
3222
3223         if (new_cfqq)
3224                 kmem_cache_free(cfq_pool, new_cfqq);
3225
3226         rcu_read_unlock();
3227         return cfqq;
3228 }
3229
3230 static struct cfq_queue **
3231 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3232 {
3233         switch (ioprio_class) {
3234         case IOPRIO_CLASS_RT:
3235                 return &cfqd->async_cfqq[0][ioprio];
3236         case IOPRIO_CLASS_NONE:
3237                 ioprio = IOPRIO_NORM;
3238                 /* fall through */
3239         case IOPRIO_CLASS_BE:
3240                 return &cfqd->async_cfqq[1][ioprio];
3241         case IOPRIO_CLASS_IDLE:
3242                 return &cfqd->async_idle_cfqq;
3243         default:
3244                 BUG();
3245         }
3246 }
3247
3248 static struct cfq_queue *
3249 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3250               struct bio *bio, gfp_t gfp_mask)
3251 {
3252         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3253         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3254         struct cfq_queue **async_cfqq = NULL;
3255         struct cfq_queue *cfqq = NULL;
3256
3257         if (!is_sync) {
3258                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3259                 cfqq = *async_cfqq;
3260         }
3261
3262         if (!cfqq)
3263                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3264
3265         /*
3266          * pin the queue now that it's allocated, scheduler exit will prune it
3267          */
3268         if (!is_sync && !(*async_cfqq)) {
3269                 cfqq->ref++;
3270                 *async_cfqq = cfqq;
3271         }
3272
3273         cfqq->ref++;
3274         return cfqq;
3275 }
3276
3277 static void
3278 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3279 {
3280         unsigned long elapsed = jiffies - ttime->last_end_request;
3281         elapsed = min(elapsed, 2UL * slice_idle);
3282
3283         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3284         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3285         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3286 }
3287
3288 static void
3289 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3290                         struct cfq_io_cq *cic)
3291 {
3292         if (cfq_cfqq_sync(cfqq)) {
3293                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3294                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3295                         cfqd->cfq_slice_idle);
3296         }
3297 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3298         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3299 #endif
3300 }
3301
3302 static void
3303 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3304                        struct request *rq)
3305 {
3306         sector_t sdist = 0;
3307         sector_t n_sec = blk_rq_sectors(rq);
3308         if (cfqq->last_request_pos) {
3309                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3310                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3311                 else
3312                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3313         }
3314
3315         cfqq->seek_history <<= 1;
3316         if (blk_queue_nonrot(cfqd->queue))
3317                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3318         else
3319                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3320 }
3321
3322 /*
3323  * Disable idle window if the process thinks too long or seeks so much that
3324  * it doesn't matter
3325  */
3326 static void
3327 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3328                        struct cfq_io_cq *cic)
3329 {
3330         int old_idle, enable_idle;
3331
3332         /*
3333          * Don't idle for async or idle io prio class
3334          */
3335         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3336                 return;
3337
3338         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3339
3340         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3341                 cfq_mark_cfqq_deep(cfqq);
3342
3343         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3344                 enable_idle = 0;
3345         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3346                  !cfqd->cfq_slice_idle ||
3347                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3348                 enable_idle = 0;
3349         else if (sample_valid(cic->ttime.ttime_samples)) {
3350                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3351                         enable_idle = 0;
3352                 else
3353                         enable_idle = 1;
3354         }
3355
3356         if (old_idle != enable_idle) {
3357                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3358                 if (enable_idle)
3359                         cfq_mark_cfqq_idle_window(cfqq);
3360                 else
3361                         cfq_clear_cfqq_idle_window(cfqq);
3362         }
3363 }
3364
3365 /*
3366  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3367  * no or if we aren't sure, a 1 will cause a preempt.
3368  */
3369 static bool
3370 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3371                    struct request *rq)
3372 {
3373         struct cfq_queue *cfqq;
3374
3375         cfqq = cfqd->active_queue;
3376         if (!cfqq)
3377                 return false;
3378
3379         if (cfq_class_idle(new_cfqq))
3380                 return false;
3381
3382         if (cfq_class_idle(cfqq))
3383                 return true;
3384
3385         /*
3386          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3387          */
3388         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3389                 return false;
3390
3391         /*
3392          * if the new request is sync, but the currently running queue is
3393          * not, let the sync request have priority.
3394          */
3395         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3396                 return true;
3397
3398         if (new_cfqq->cfqg != cfqq->cfqg)
3399                 return false;
3400
3401         if (cfq_slice_used(cfqq))
3402                 return true;
3403
3404         /* Allow preemption only if we are idling on sync-noidle tree */
3405         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3406             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3407             new_cfqq->service_tree->count == 2 &&
3408             RB_EMPTY_ROOT(&cfqq->sort_list))
3409                 return true;
3410
3411         /*
3412          * So both queues are sync. Let the new request get disk time if
3413          * it's a metadata request and the current queue is doing regular IO.
3414          */
3415         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3416                 return true;
3417
3418         /*
3419          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3420          */
3421         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3422                 return true;
3423
3424         /* An idle queue should not be idle now for some reason */
3425         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3426                 return true;
3427
3428         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3429                 return false;
3430
3431         /*
3432          * if this request is as-good as one we would expect from the
3433          * current cfqq, let it preempt
3434          */
3435         if (cfq_rq_close(cfqd, cfqq, rq))
3436                 return true;
3437
3438         return false;
3439 }
3440
3441 /*
3442  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3443  * let it have half of its nominal slice.
3444  */
3445 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3446 {
3447         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3448
3449         cfq_log_cfqq(cfqd, cfqq, "preempt");
3450         cfq_slice_expired(cfqd, 1);
3451
3452         /*
3453          * workload type is changed, don't save slice, otherwise preempt
3454          * doesn't happen
3455          */
3456         if (old_type != cfqq_type(cfqq))
3457                 cfqq->cfqg->saved_workload_slice = 0;
3458
3459         /*
3460          * Put the new queue at the front of the of the current list,
3461          * so we know that it will be selected next.
3462          */
3463         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3464
3465         cfq_service_tree_add(cfqd, cfqq, 1);
3466
3467         cfqq->slice_end = 0;
3468         cfq_mark_cfqq_slice_new(cfqq);
3469 }
3470
3471 /*
3472  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3473  * something we should do about it
3474  */
3475 static void
3476 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3477                 struct request *rq)
3478 {
3479         struct cfq_io_cq *cic = RQ_CIC(rq);
3480
3481         cfqd->rq_queued++;
3482         if (rq->cmd_flags & REQ_PRIO)
3483                 cfqq->prio_pending++;
3484
3485         cfq_update_io_thinktime(cfqd, cfqq, cic);
3486         cfq_update_io_seektime(cfqd, cfqq, rq);
3487         cfq_update_idle_window(cfqd, cfqq, cic);
3488
3489         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3490
3491         if (cfqq == cfqd->active_queue) {
3492                 /*
3493                  * Remember that we saw a request from this process, but
3494                  * don't start queuing just yet. Otherwise we risk seeing lots
3495                  * of tiny requests, because we disrupt the normal plugging
3496                  * and merging. If the request is already larger than a single
3497                  * page, let it rip immediately. For that case we assume that
3498                  * merging is already done. Ditto for a busy system that
3499                  * has other work pending, don't risk delaying until the
3500                  * idle timer unplug to continue working.
3501                  */
3502                 if (cfq_cfqq_wait_request(cfqq)) {
3503                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3504                             cfqd->busy_queues > 1) {
3505                                 cfq_del_timer(cfqd, cfqq);
3506                                 cfq_clear_cfqq_wait_request(cfqq);
3507                                 __blk_run_queue(cfqd->queue);
3508                         } else {
3509                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3510                                 cfq_mark_cfqq_must_dispatch(cfqq);
3511                         }
3512                 }
3513         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3514                 /*
3515                  * not the active queue - expire current slice if it is
3516                  * idle and has expired it's mean thinktime or this new queue
3517                  * has some old slice time left and is of higher priority or
3518                  * this new queue is RT and the current one is BE
3519                  */
3520                 cfq_preempt_queue(cfqd, cfqq);
3521                 __blk_run_queue(cfqd->queue);
3522         }
3523 }
3524
3525 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3526 {
3527         struct cfq_data *cfqd = q->elevator->elevator_data;
3528         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3529
3530         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3531         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3532
3533         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3534         list_add_tail(&rq->queuelist, &cfqq->fifo);
3535         cfq_add_rq_rb(rq);
3536         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3537                                  rq->cmd_flags);
3538         cfq_rq_enqueued(cfqd, cfqq, rq);
3539 }
3540
3541 /*
3542  * Update hw_tag based on peak queue depth over 50 samples under
3543  * sufficient load.
3544  */
3545 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3546 {
3547         struct cfq_queue *cfqq = cfqd->active_queue;
3548
3549         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3550                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3551
3552         if (cfqd->hw_tag == 1)
3553                 return;
3554
3555         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3556             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3557                 return;
3558
3559         /*
3560          * If active queue hasn't enough requests and can idle, cfq might not
3561          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3562          * case
3563          */
3564         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3565             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3566             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3567                 return;
3568
3569         if (cfqd->hw_tag_samples++ < 50)
3570                 return;
3571
3572         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3573                 cfqd->hw_tag = 1;
3574         else
3575                 cfqd->hw_tag = 0;
3576 }
3577
3578 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3579 {
3580         struct cfq_io_cq *cic = cfqd->active_cic;
3581
3582         /* If the queue already has requests, don't wait */
3583         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3584                 return false;
3585
3586         /* If there are other queues in the group, don't wait */
3587         if (cfqq->cfqg->nr_cfqq > 1)
3588                 return false;
3589
3590         /* the only queue in the group, but think time is big */
3591         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3592                 return false;
3593
3594         if (cfq_slice_used(cfqq))
3595                 return true;
3596
3597         /* if slice left is less than think time, wait busy */
3598         if (cic && sample_valid(cic->ttime.ttime_samples)
3599             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3600                 return true;
3601
3602         /*
3603          * If think times is less than a jiffy than ttime_mean=0 and above
3604          * will not be true. It might happen that slice has not expired yet
3605          * but will expire soon (4-5 ns) during select_queue(). To cover the
3606          * case where think time is less than a jiffy, mark the queue wait
3607          * busy if only 1 jiffy is left in the slice.
3608          */
3609         if (cfqq->slice_end - jiffies == 1)
3610                 return true;
3611
3612         return false;
3613 }
3614
3615 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3616 {
3617         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3618         struct cfq_data *cfqd = cfqq->cfqd;
3619         const int sync = rq_is_sync(rq);
3620         unsigned long now;
3621
3622         now = jiffies;
3623         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3624                      !!(rq->cmd_flags & REQ_NOIDLE));
3625
3626         cfq_update_hw_tag(cfqd);
3627
3628         WARN_ON(!cfqd->rq_in_driver);
3629         WARN_ON(!cfqq->dispatched);
3630         cfqd->rq_in_driver--;
3631         cfqq->dispatched--;
3632         (RQ_CFQG(rq))->dispatched--;
3633         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3634                                      rq_io_start_time_ns(rq), rq->cmd_flags);
3635
3636         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3637
3638         if (sync) {
3639                 struct cfq_rb_root *service_tree;
3640
3641                 RQ_CIC(rq)->ttime.last_end_request = now;
3642
3643                 if (cfq_cfqq_on_rr(cfqq))
3644                         service_tree = cfqq->service_tree;
3645                 else
3646                         service_tree = service_tree_for(cfqq->cfqg,
3647                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3648                 service_tree->ttime.last_end_request = now;
3649                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3650                         cfqd->last_delayed_sync = now;
3651         }
3652
3653 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3654         cfqq->cfqg->ttime.last_end_request = now;
3655 #endif
3656
3657         /*
3658          * If this is the active queue, check if it needs to be expired,
3659          * or if we want to idle in case it has no pending requests.
3660          */
3661         if (cfqd->active_queue == cfqq) {
3662                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3663
3664                 if (cfq_cfqq_slice_new(cfqq)) {
3665                         cfq_set_prio_slice(cfqd, cfqq);
3666                         cfq_clear_cfqq_slice_new(cfqq);
3667                 }
3668
3669                 /*
3670                  * Should we wait for next request to come in before we expire
3671                  * the queue.
3672                  */
3673                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3674                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3675                         if (!cfqd->cfq_slice_idle)
3676                                 extend_sl = cfqd->cfq_group_idle;
3677                         cfqq->slice_end = jiffies + extend_sl;
3678                         cfq_mark_cfqq_wait_busy(cfqq);
3679                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3680                 }
3681
3682                 /*
3683                  * Idling is not enabled on:
3684                  * - expired queues
3685                  * - idle-priority queues
3686                  * - async queues
3687                  * - queues with still some requests queued
3688                  * - when there is a close cooperator
3689                  */
3690                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3691                         cfq_slice_expired(cfqd, 1);
3692                 else if (sync && cfqq_empty &&
3693                          !cfq_close_cooperator(cfqd, cfqq)) {
3694                         cfq_arm_slice_timer(cfqd);
3695                 }
3696         }
3697
3698         if (!cfqd->rq_in_driver)
3699                 cfq_schedule_dispatch(cfqd);
3700 }
3701
3702 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3703 {
3704         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3705                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3706                 return ELV_MQUEUE_MUST;
3707         }
3708
3709         return ELV_MQUEUE_MAY;
3710 }
3711
3712 static int cfq_may_queue(struct request_queue *q, int rw)
3713 {
3714         struct cfq_data *cfqd = q->elevator->elevator_data;
3715         struct task_struct *tsk = current;
3716         struct cfq_io_cq *cic;
3717         struct cfq_queue *cfqq;
3718
3719         /*
3720          * don't force setup of a queue from here, as a call to may_queue
3721          * does not necessarily imply that a request actually will be queued.
3722          * so just lookup a possibly existing queue, or return 'may queue'
3723          * if that fails
3724          */
3725         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3726         if (!cic)
3727                 return ELV_MQUEUE_MAY;
3728
3729         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3730         if (cfqq) {
3731                 cfq_init_prio_data(cfqq, cic);
3732
3733                 return __cfq_may_queue(cfqq);
3734         }
3735
3736         return ELV_MQUEUE_MAY;
3737 }
3738
3739 /*
3740  * queue lock held here
3741  */
3742 static void cfq_put_request(struct request *rq)
3743 {
3744         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3745
3746         if (cfqq) {
3747                 const int rw = rq_data_dir(rq);
3748
3749                 BUG_ON(!cfqq->allocated[rw]);
3750                 cfqq->allocated[rw]--;
3751
3752                 /* Put down rq reference on cfqg */
3753                 cfqg_put(RQ_CFQG(rq));
3754                 rq->elv.priv[0] = NULL;
3755                 rq->elv.priv[1] = NULL;
3756
3757                 cfq_put_queue(cfqq);
3758         }
3759 }
3760
3761 static struct cfq_queue *
3762 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3763                 struct cfq_queue *cfqq)
3764 {
3765         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3766         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3767         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3768         cfq_put_queue(cfqq);
3769         return cic_to_cfqq(cic, 1);
3770 }
3771
3772 /*
3773  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3774  * was the last process referring to said cfqq.
3775  */
3776 static struct cfq_queue *
3777 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3778 {
3779         if (cfqq_process_refs(cfqq) == 1) {
3780                 cfqq->pid = current->pid;
3781                 cfq_clear_cfqq_coop(cfqq);
3782                 cfq_clear_cfqq_split_coop(cfqq);
3783                 return cfqq;
3784         }
3785
3786         cic_set_cfqq(cic, NULL, 1);
3787
3788         cfq_put_cooperator(cfqq);
3789
3790         cfq_put_queue(cfqq);
3791         return NULL;
3792 }
3793 /*
3794  * Allocate cfq data structures associated with this request.
3795  */
3796 static int
3797 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3798                 gfp_t gfp_mask)
3799 {
3800         struct cfq_data *cfqd = q->elevator->elevator_data;
3801         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3802         const int rw = rq_data_dir(rq);
3803         const bool is_sync = rq_is_sync(rq);
3804         struct cfq_queue *cfqq;
3805
3806         might_sleep_if(gfp_mask & __GFP_WAIT);
3807
3808         spin_lock_irq(q->queue_lock);
3809
3810         check_ioprio_changed(cic, bio);
3811         check_blkcg_changed(cic, bio);
3812 new_queue:
3813         cfqq = cic_to_cfqq(cic, is_sync);
3814         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3815                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
3816                 cic_set_cfqq(cic, cfqq, is_sync);
3817         } else {
3818                 /*
3819                  * If the queue was seeky for too long, break it apart.
3820                  */
3821                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3822                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3823                         cfqq = split_cfqq(cic, cfqq);
3824                         if (!cfqq)
3825                                 goto new_queue;
3826                 }
3827
3828                 /*
3829                  * Check to see if this queue is scheduled to merge with
3830                  * another, closely cooperating queue.  The merging of
3831                  * queues happens here as it must be done in process context.
3832                  * The reference on new_cfqq was taken in merge_cfqqs.
3833                  */
3834                 if (cfqq->new_cfqq)
3835                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3836         }
3837
3838         cfqq->allocated[rw]++;
3839
3840         cfqq->ref++;
3841         cfqg_get(cfqq->cfqg);
3842         rq->elv.priv[0] = cfqq;
3843         rq->elv.priv[1] = cfqq->cfqg;
3844         spin_unlock_irq(q->queue_lock);
3845         return 0;
3846 }
3847
3848 static void cfq_kick_queue(struct work_struct *work)
3849 {
3850         struct cfq_data *cfqd =
3851                 container_of(work, struct cfq_data, unplug_work);
3852         struct request_queue *q = cfqd->queue;
3853
3854         spin_lock_irq(q->queue_lock);
3855         __blk_run_queue(cfqd->queue);
3856         spin_unlock_irq(q->queue_lock);
3857 }
3858
3859 /*
3860  * Timer running if the active_queue is currently idling inside its time slice
3861  */
3862 static void cfq_idle_slice_timer(unsigned long data)
3863 {
3864         struct cfq_data *cfqd = (struct cfq_data *) data;
3865         struct cfq_queue *cfqq;
3866         unsigned long flags;
3867         int timed_out = 1;
3868
3869         cfq_log(cfqd, "idle timer fired");
3870
3871         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3872
3873         cfqq = cfqd->active_queue;
3874         if (cfqq) {
3875                 timed_out = 0;
3876
3877                 /*
3878                  * We saw a request before the queue expired, let it through
3879                  */
3880                 if (cfq_cfqq_must_dispatch(cfqq))
3881                         goto out_kick;
3882
3883                 /*
3884                  * expired
3885                  */
3886                 if (cfq_slice_used(cfqq))
3887                         goto expire;
3888
3889                 /*
3890                  * only expire and reinvoke request handler, if there are
3891                  * other queues with pending requests
3892                  */
3893                 if (!cfqd->busy_queues)
3894                         goto out_cont;
3895
3896                 /*
3897                  * not expired and it has a request pending, let it dispatch
3898                  */
3899                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3900                         goto out_kick;
3901
3902                 /*
3903                  * Queue depth flag is reset only when the idle didn't succeed
3904                  */
3905                 cfq_clear_cfqq_deep(cfqq);
3906         }
3907 expire:
3908         cfq_slice_expired(cfqd, timed_out);
3909 out_kick:
3910         cfq_schedule_dispatch(cfqd);
3911 out_cont:
3912         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3913 }
3914
3915 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3916 {
3917         del_timer_sync(&cfqd->idle_slice_timer);
3918         cancel_work_sync(&cfqd->unplug_work);
3919 }
3920
3921 static void cfq_put_async_queues(struct cfq_data *cfqd)
3922 {
3923         int i;
3924
3925         for (i = 0; i < IOPRIO_BE_NR; i++) {
3926                 if (cfqd->async_cfqq[0][i])
3927                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3928                 if (cfqd->async_cfqq[1][i])
3929                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3930         }
3931
3932         if (cfqd->async_idle_cfqq)
3933                 cfq_put_queue(cfqd->async_idle_cfqq);
3934 }
3935
3936 static void cfq_exit_queue(struct elevator_queue *e)
3937 {
3938         struct cfq_data *cfqd = e->elevator_data;
3939         struct request_queue *q = cfqd->queue;
3940
3941         cfq_shutdown_timer_wq(cfqd);
3942
3943         spin_lock_irq(q->queue_lock);
3944
3945         if (cfqd->active_queue)
3946                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3947
3948         cfq_put_async_queues(cfqd);
3949
3950         spin_unlock_irq(q->queue_lock);
3951
3952         cfq_shutdown_timer_wq(cfqd);
3953
3954 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3955         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
3956 #else
3957         kfree(cfqd->root_group);
3958 #endif
3959         kfree(cfqd);
3960 }
3961
3962 static int cfq_init_queue(struct request_queue *q)
3963 {
3964         struct cfq_data *cfqd;
3965         struct blkcg_gq *blkg __maybe_unused;
3966         int i, ret;
3967
3968         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3969         if (!cfqd)
3970                 return -ENOMEM;
3971
3972         cfqd->queue = q;
3973         q->elevator->elevator_data = cfqd;
3974
3975         /* Init root service tree */
3976         cfqd->grp_service_tree = CFQ_RB_ROOT;
3977
3978         /* Init root group and prefer root group over other groups by default */
3979 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3980         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
3981         if (ret)
3982                 goto out_free;
3983
3984         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
3985 #else
3986         ret = -ENOMEM;
3987         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3988                                         GFP_KERNEL, cfqd->queue->node);
3989         if (!cfqd->root_group)
3990                 goto out_free;
3991
3992         cfq_init_cfqg_base(cfqd->root_group);
3993 #endif
3994         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
3995
3996         /*
3997          * Not strictly needed (since RB_ROOT just clears the node and we
3998          * zeroed cfqd on alloc), but better be safe in case someone decides
3999          * to add magic to the rb code
4000          */
4001         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4002                 cfqd->prio_trees[i] = RB_ROOT;
4003
4004         /*
4005          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4006          * Grab a permanent reference to it, so that the normal code flow
4007          * will not attempt to free it.  oom_cfqq is linked to root_group
4008          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4009          * the reference from linking right away.
4010          */
4011         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4012         cfqd->oom_cfqq.ref++;
4013
4014         spin_lock_irq(q->queue_lock);
4015         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4016         cfqg_put(cfqd->root_group);
4017         spin_unlock_irq(q->queue_lock);
4018
4019         init_timer(&cfqd->idle_slice_timer);
4020         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4021         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4022
4023         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4024
4025         cfqd->cfq_quantum = cfq_quantum;
4026         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4027         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4028         cfqd->cfq_back_max = cfq_back_max;
4029         cfqd->cfq_back_penalty = cfq_back_penalty;
4030         cfqd->cfq_slice[0] = cfq_slice_async;
4031         cfqd->cfq_slice[1] = cfq_slice_sync;
4032         cfqd->cfq_target_latency = cfq_target_latency;
4033         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4034         cfqd->cfq_slice_idle = cfq_slice_idle;
4035         cfqd->cfq_group_idle = cfq_group_idle;
4036         cfqd->cfq_latency = 1;
4037         cfqd->hw_tag = -1;
4038         /*
4039          * we optimistically start assuming sync ops weren't delayed in last
4040          * second, in order to have larger depth for async operations.
4041          */
4042         cfqd->last_delayed_sync = jiffies - HZ;
4043         return 0;
4044
4045 out_free:
4046         kfree(cfqd);
4047         return ret;
4048 }
4049
4050 /*
4051  * sysfs parts below -->
4052  */
4053 static ssize_t
4054 cfq_var_show(unsigned int var, char *page)
4055 {
4056         return sprintf(page, "%d\n", var);
4057 }
4058
4059 static ssize_t
4060 cfq_var_store(unsigned int *var, const char *page, size_t count)
4061 {
4062         char *p = (char *) page;
4063
4064         *var = simple_strtoul(p, &p, 10);
4065         return count;
4066 }
4067
4068 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4069 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4070 {                                                                       \
4071         struct cfq_data *cfqd = e->elevator_data;                       \
4072         unsigned int __data = __VAR;                                    \
4073         if (__CONV)                                                     \
4074                 __data = jiffies_to_msecs(__data);                      \
4075         return cfq_var_show(__data, (page));                            \
4076 }
4077 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4078 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4079 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4080 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4081 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4082 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4083 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4084 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4085 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4086 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4087 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4088 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4089 #undef SHOW_FUNCTION
4090
4091 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4092 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4093 {                                                                       \
4094         struct cfq_data *cfqd = e->elevator_data;                       \
4095         unsigned int __data;                                            \
4096         int ret = cfq_var_store(&__data, (page), count);                \
4097         if (__data < (MIN))                                             \
4098                 __data = (MIN);                                         \
4099         else if (__data > (MAX))                                        \
4100                 __data = (MAX);                                         \
4101         if (__CONV)                                                     \
4102                 *(__PTR) = msecs_to_jiffies(__data);                    \
4103         else                                                            \
4104                 *(__PTR) = __data;                                      \
4105         return ret;                                                     \
4106 }
4107 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4108 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4109                 UINT_MAX, 1);
4110 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4111                 UINT_MAX, 1);
4112 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4113 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4114                 UINT_MAX, 0);
4115 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4116 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4117 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4118 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4119 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4120                 UINT_MAX, 0);
4121 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4122 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4123 #undef STORE_FUNCTION
4124
4125 #define CFQ_ATTR(name) \
4126         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4127
4128 static struct elv_fs_entry cfq_attrs[] = {
4129         CFQ_ATTR(quantum),
4130         CFQ_ATTR(fifo_expire_sync),
4131         CFQ_ATTR(fifo_expire_async),
4132         CFQ_ATTR(back_seek_max),
4133         CFQ_ATTR(back_seek_penalty),
4134         CFQ_ATTR(slice_sync),
4135         CFQ_ATTR(slice_async),
4136         CFQ_ATTR(slice_async_rq),
4137         CFQ_ATTR(slice_idle),
4138         CFQ_ATTR(group_idle),
4139         CFQ_ATTR(low_latency),
4140         CFQ_ATTR(target_latency),
4141         __ATTR_NULL
4142 };
4143
4144 static struct elevator_type iosched_cfq = {
4145         .ops = {
4146                 .elevator_merge_fn =            cfq_merge,
4147                 .elevator_merged_fn =           cfq_merged_request,
4148                 .elevator_merge_req_fn =        cfq_merged_requests,
4149                 .elevator_allow_merge_fn =      cfq_allow_merge,
4150                 .elevator_bio_merged_fn =       cfq_bio_merged,
4151                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4152                 .elevator_add_req_fn =          cfq_insert_request,
4153                 .elevator_activate_req_fn =     cfq_activate_request,
4154                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4155                 .elevator_completed_req_fn =    cfq_completed_request,
4156                 .elevator_former_req_fn =       elv_rb_former_request,
4157                 .elevator_latter_req_fn =       elv_rb_latter_request,
4158                 .elevator_init_icq_fn =         cfq_init_icq,
4159                 .elevator_exit_icq_fn =         cfq_exit_icq,
4160                 .elevator_set_req_fn =          cfq_set_request,
4161                 .elevator_put_req_fn =          cfq_put_request,
4162                 .elevator_may_queue_fn =        cfq_may_queue,
4163                 .elevator_init_fn =             cfq_init_queue,
4164                 .elevator_exit_fn =             cfq_exit_queue,
4165         },
4166         .icq_size       =       sizeof(struct cfq_io_cq),
4167         .icq_align      =       __alignof__(struct cfq_io_cq),
4168         .elevator_attrs =       cfq_attrs,
4169         .elevator_name  =       "cfq",
4170         .elevator_owner =       THIS_MODULE,
4171 };
4172
4173 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4174 static struct blkcg_policy blkcg_policy_cfq = {
4175         .pd_size                = sizeof(struct cfq_group),
4176         .cftypes                = cfq_blkcg_files,
4177
4178         .pd_init_fn             = cfq_pd_init,
4179         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4180 };
4181 #endif
4182
4183 static int __init cfq_init(void)
4184 {
4185         int ret;
4186
4187         /*
4188          * could be 0 on HZ < 1000 setups
4189          */
4190         if (!cfq_slice_async)
4191                 cfq_slice_async = 1;
4192         if (!cfq_slice_idle)
4193                 cfq_slice_idle = 1;
4194
4195 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4196         if (!cfq_group_idle)
4197                 cfq_group_idle = 1;
4198
4199         ret = blkcg_policy_register(&blkcg_policy_cfq);
4200         if (ret)
4201                 return ret;
4202 #else
4203         cfq_group_idle = 0;
4204 #endif
4205
4206         ret = -ENOMEM;
4207         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4208         if (!cfq_pool)
4209                 goto err_pol_unreg;
4210
4211         ret = elv_register(&iosched_cfq);
4212         if (ret)
4213                 goto err_free_pool;
4214
4215         return 0;
4216
4217 err_free_pool:
4218         kmem_cache_destroy(cfq_pool);
4219 err_pol_unreg:
4220 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4221         blkcg_policy_unregister(&blkcg_policy_cfq);
4222 #endif
4223         return ret;
4224 }
4225
4226 static void __exit cfq_exit(void)
4227 {
4228 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4229         blkcg_policy_unregister(&blkcg_policy_cfq);
4230 #endif
4231         elv_unregister(&iosched_cfq);
4232         kmem_cache_destroy(cfq_pool);
4233 }
4234
4235 module_init(cfq_init);
4236 module_exit(cfq_exit);
4237
4238 MODULE_AUTHOR("Jens Axboe");
4239 MODULE_LICENSE("GPL");
4240 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");