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             cfqq == RQ_CFQQ(next)) {
1978                 list_move(&rq->queuelist, &next->queuelist);
1979                 rq_set_fifo_time(rq, rq_fifo_time(next));
1980         }
1981
1982         if (cfqq->next_rq == next)
1983                 cfqq->next_rq = rq;
1984         cfq_remove_request(next);
1985         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
1986
1987         cfqq = RQ_CFQQ(next);
1988         /*
1989          * all requests of this queue are merged to other queues, delete it
1990          * from the service tree. If it's the active_queue,
1991          * cfq_dispatch_requests() will choose to expire it or do idle
1992          */
1993         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1994             cfqq != cfqd->active_queue)
1995                 cfq_del_cfqq_rr(cfqd, cfqq);
1996 }
1997
1998 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1999                            struct bio *bio)
2000 {
2001         struct cfq_data *cfqd = q->elevator->elevator_data;
2002         struct cfq_io_cq *cic;
2003         struct cfq_queue *cfqq;
2004
2005         /*
2006          * Disallow merge of a sync bio into an async request.
2007          */
2008         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2009                 return false;
2010
2011         /*
2012          * Lookup the cfqq that this bio will be queued with and allow
2013          * merge only if rq is queued there.
2014          */
2015         cic = cfq_cic_lookup(cfqd, current->io_context);
2016         if (!cic)
2017                 return false;
2018
2019         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2020         return cfqq == RQ_CFQQ(rq);
2021 }
2022
2023 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2024 {
2025         del_timer(&cfqd->idle_slice_timer);
2026         cfqg_stats_update_idle_time(cfqq->cfqg);
2027 }
2028
2029 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2030                                    struct cfq_queue *cfqq)
2031 {
2032         if (cfqq) {
2033                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
2034                                 cfqd->serving_prio, cfqd->serving_type);
2035                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2036                 cfqq->slice_start = 0;
2037                 cfqq->dispatch_start = jiffies;
2038                 cfqq->allocated_slice = 0;
2039                 cfqq->slice_end = 0;
2040                 cfqq->slice_dispatch = 0;
2041                 cfqq->nr_sectors = 0;
2042
2043                 cfq_clear_cfqq_wait_request(cfqq);
2044                 cfq_clear_cfqq_must_dispatch(cfqq);
2045                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2046                 cfq_clear_cfqq_fifo_expire(cfqq);
2047                 cfq_mark_cfqq_slice_new(cfqq);
2048
2049                 cfq_del_timer(cfqd, cfqq);
2050         }
2051
2052         cfqd->active_queue = cfqq;
2053 }
2054
2055 /*
2056  * current cfqq expired its slice (or was too idle), select new one
2057  */
2058 static void
2059 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2060                     bool timed_out)
2061 {
2062         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2063
2064         if (cfq_cfqq_wait_request(cfqq))
2065                 cfq_del_timer(cfqd, cfqq);
2066
2067         cfq_clear_cfqq_wait_request(cfqq);
2068         cfq_clear_cfqq_wait_busy(cfqq);
2069
2070         /*
2071          * If this cfqq is shared between multiple processes, check to
2072          * make sure that those processes are still issuing I/Os within
2073          * the mean seek distance.  If not, it may be time to break the
2074          * queues apart again.
2075          */
2076         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2077                 cfq_mark_cfqq_split_coop(cfqq);
2078
2079         /*
2080          * store what was left of this slice, if the queue idled/timed out
2081          */
2082         if (timed_out) {
2083                 if (cfq_cfqq_slice_new(cfqq))
2084                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2085                 else
2086                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2087                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2088         }
2089
2090         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2091
2092         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2093                 cfq_del_cfqq_rr(cfqd, cfqq);
2094
2095         cfq_resort_rr_list(cfqd, cfqq);
2096
2097         if (cfqq == cfqd->active_queue)
2098                 cfqd->active_queue = NULL;
2099
2100         if (cfqd->active_cic) {
2101                 put_io_context(cfqd->active_cic->icq.ioc);
2102                 cfqd->active_cic = NULL;
2103         }
2104 }
2105
2106 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2107 {
2108         struct cfq_queue *cfqq = cfqd->active_queue;
2109
2110         if (cfqq)
2111                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2112 }
2113
2114 /*
2115  * Get next queue for service. Unless we have a queue preemption,
2116  * we'll simply select the first cfqq in the service tree.
2117  */
2118 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2119 {
2120         struct cfq_rb_root *service_tree =
2121                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
2122                                         cfqd->serving_type);
2123
2124         if (!cfqd->rq_queued)
2125                 return NULL;
2126
2127         /* There is nothing to dispatch */
2128         if (!service_tree)
2129                 return NULL;
2130         if (RB_EMPTY_ROOT(&service_tree->rb))
2131                 return NULL;
2132         return cfq_rb_first(service_tree);
2133 }
2134
2135 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2136 {
2137         struct cfq_group *cfqg;
2138         struct cfq_queue *cfqq;
2139         int i, j;
2140         struct cfq_rb_root *st;
2141
2142         if (!cfqd->rq_queued)
2143                 return NULL;
2144
2145         cfqg = cfq_get_next_cfqg(cfqd);
2146         if (!cfqg)
2147                 return NULL;
2148
2149         for_each_cfqg_st(cfqg, i, j, st)
2150                 if ((cfqq = cfq_rb_first(st)) != NULL)
2151                         return cfqq;
2152         return NULL;
2153 }
2154
2155 /*
2156  * Get and set a new active queue for service.
2157  */
2158 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2159                                               struct cfq_queue *cfqq)
2160 {
2161         if (!cfqq)
2162                 cfqq = cfq_get_next_queue(cfqd);
2163
2164         __cfq_set_active_queue(cfqd, cfqq);
2165         return cfqq;
2166 }
2167
2168 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2169                                           struct request *rq)
2170 {
2171         if (blk_rq_pos(rq) >= cfqd->last_position)
2172                 return blk_rq_pos(rq) - cfqd->last_position;
2173         else
2174                 return cfqd->last_position - blk_rq_pos(rq);
2175 }
2176
2177 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2178                                struct request *rq)
2179 {
2180         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2181 }
2182
2183 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2184                                     struct cfq_queue *cur_cfqq)
2185 {
2186         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2187         struct rb_node *parent, *node;
2188         struct cfq_queue *__cfqq;
2189         sector_t sector = cfqd->last_position;
2190
2191         if (RB_EMPTY_ROOT(root))
2192                 return NULL;
2193
2194         /*
2195          * First, if we find a request starting at the end of the last
2196          * request, choose it.
2197          */
2198         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2199         if (__cfqq)
2200                 return __cfqq;
2201
2202         /*
2203          * If the exact sector wasn't found, the parent of the NULL leaf
2204          * will contain the closest sector.
2205          */
2206         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2207         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2208                 return __cfqq;
2209
2210         if (blk_rq_pos(__cfqq->next_rq) < sector)
2211                 node = rb_next(&__cfqq->p_node);
2212         else
2213                 node = rb_prev(&__cfqq->p_node);
2214         if (!node)
2215                 return NULL;
2216
2217         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2218         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2219                 return __cfqq;
2220
2221         return NULL;
2222 }
2223
2224 /*
2225  * cfqd - obvious
2226  * cur_cfqq - passed in so that we don't decide that the current queue is
2227  *            closely cooperating with itself.
2228  *
2229  * So, basically we're assuming that that cur_cfqq has dispatched at least
2230  * one request, and that cfqd->last_position reflects a position on the disk
2231  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2232  * assumption.
2233  */
2234 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2235                                               struct cfq_queue *cur_cfqq)
2236 {
2237         struct cfq_queue *cfqq;
2238
2239         if (cfq_class_idle(cur_cfqq))
2240                 return NULL;
2241         if (!cfq_cfqq_sync(cur_cfqq))
2242                 return NULL;
2243         if (CFQQ_SEEKY(cur_cfqq))
2244                 return NULL;
2245
2246         /*
2247          * Don't search priority tree if it's the only queue in the group.
2248          */
2249         if (cur_cfqq->cfqg->nr_cfqq == 1)
2250                 return NULL;
2251
2252         /*
2253          * We should notice if some of the queues are cooperating, eg
2254          * working closely on the same area of the disk. In that case,
2255          * we can group them together and don't waste time idling.
2256          */
2257         cfqq = cfqq_close(cfqd, cur_cfqq);
2258         if (!cfqq)
2259                 return NULL;
2260
2261         /* If new queue belongs to different cfq_group, don't choose it */
2262         if (cur_cfqq->cfqg != cfqq->cfqg)
2263                 return NULL;
2264
2265         /*
2266          * It only makes sense to merge sync queues.
2267          */
2268         if (!cfq_cfqq_sync(cfqq))
2269                 return NULL;
2270         if (CFQQ_SEEKY(cfqq))
2271                 return NULL;
2272
2273         /*
2274          * Do not merge queues of different priority classes
2275          */
2276         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2277                 return NULL;
2278
2279         return cfqq;
2280 }
2281
2282 /*
2283  * Determine whether we should enforce idle window for this queue.
2284  */
2285
2286 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2287 {
2288         enum wl_prio_t prio = cfqq_prio(cfqq);
2289         struct cfq_rb_root *service_tree = cfqq->service_tree;
2290
2291         BUG_ON(!service_tree);
2292         BUG_ON(!service_tree->count);
2293
2294         if (!cfqd->cfq_slice_idle)
2295                 return false;
2296
2297         /* We never do for idle class queues. */
2298         if (prio == IDLE_WORKLOAD)
2299                 return false;
2300
2301         /* We do for queues that were marked with idle window flag. */
2302         if (cfq_cfqq_idle_window(cfqq) &&
2303            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2304                 return true;
2305
2306         /*
2307          * Otherwise, we do only if they are the last ones
2308          * in their service tree.
2309          */
2310         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2311            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2312                 return true;
2313         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2314                         service_tree->count);
2315         return false;
2316 }
2317
2318 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2319 {
2320         struct cfq_queue *cfqq = cfqd->active_queue;
2321         struct cfq_io_cq *cic;
2322         unsigned long sl, group_idle = 0;
2323
2324         /*
2325          * SSD device without seek penalty, disable idling. But only do so
2326          * for devices that support queuing, otherwise we still have a problem
2327          * with sync vs async workloads.
2328          */
2329         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2330                 return;
2331
2332         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2333         WARN_ON(cfq_cfqq_slice_new(cfqq));
2334
2335         /*
2336          * idle is disabled, either manually or by past process history
2337          */
2338         if (!cfq_should_idle(cfqd, cfqq)) {
2339                 /* no queue idling. Check for group idling */
2340                 if (cfqd->cfq_group_idle)
2341                         group_idle = cfqd->cfq_group_idle;
2342                 else
2343                         return;
2344         }
2345
2346         /*
2347          * still active requests from this queue, don't idle
2348          */
2349         if (cfqq->dispatched)
2350                 return;
2351
2352         /*
2353          * task has exited, don't wait
2354          */
2355         cic = cfqd->active_cic;
2356         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2357                 return;
2358
2359         /*
2360          * If our average think time is larger than the remaining time
2361          * slice, then don't idle. This avoids overrunning the allotted
2362          * time slice.
2363          */
2364         if (sample_valid(cic->ttime.ttime_samples) &&
2365             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2366                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2367                              cic->ttime.ttime_mean);
2368                 return;
2369         }
2370
2371         /* There are other queues in the group, don't do group idle */
2372         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2373                 return;
2374
2375         cfq_mark_cfqq_wait_request(cfqq);
2376
2377         if (group_idle)
2378                 sl = cfqd->cfq_group_idle;
2379         else
2380                 sl = cfqd->cfq_slice_idle;
2381
2382         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2383         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2384         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2385                         group_idle ? 1 : 0);
2386 }
2387
2388 /*
2389  * Move request from internal lists to the request queue dispatch list.
2390  */
2391 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2392 {
2393         struct cfq_data *cfqd = q->elevator->elevator_data;
2394         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2395
2396         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2397
2398         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2399         cfq_remove_request(rq);
2400         cfqq->dispatched++;
2401         (RQ_CFQG(rq))->dispatched++;
2402         elv_dispatch_sort(q, rq);
2403
2404         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2405         cfqq->nr_sectors += blk_rq_sectors(rq);
2406         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2407 }
2408
2409 /*
2410  * return expired entry, or NULL to just start from scratch in rbtree
2411  */
2412 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2413 {
2414         struct request *rq = NULL;
2415
2416         if (cfq_cfqq_fifo_expire(cfqq))
2417                 return NULL;
2418
2419         cfq_mark_cfqq_fifo_expire(cfqq);
2420
2421         if (list_empty(&cfqq->fifo))
2422                 return NULL;
2423
2424         rq = rq_entry_fifo(cfqq->fifo.next);
2425         if (time_before(jiffies, rq_fifo_time(rq)))
2426                 rq = NULL;
2427
2428         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2429         return rq;
2430 }
2431
2432 static inline int
2433 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2434 {
2435         const int base_rq = cfqd->cfq_slice_async_rq;
2436
2437         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2438
2439         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2440 }
2441
2442 /*
2443  * Must be called with the queue_lock held.
2444  */
2445 static int cfqq_process_refs(struct cfq_queue *cfqq)
2446 {
2447         int process_refs, io_refs;
2448
2449         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2450         process_refs = cfqq->ref - io_refs;
2451         BUG_ON(process_refs < 0);
2452         return process_refs;
2453 }
2454
2455 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2456 {
2457         int process_refs, new_process_refs;
2458         struct cfq_queue *__cfqq;
2459
2460         /*
2461          * If there are no process references on the new_cfqq, then it is
2462          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2463          * chain may have dropped their last reference (not just their
2464          * last process reference).
2465          */
2466         if (!cfqq_process_refs(new_cfqq))
2467                 return;
2468
2469         /* Avoid a circular list and skip interim queue merges */
2470         while ((__cfqq = new_cfqq->new_cfqq)) {
2471                 if (__cfqq == cfqq)
2472                         return;
2473                 new_cfqq = __cfqq;
2474         }
2475
2476         process_refs = cfqq_process_refs(cfqq);
2477         new_process_refs = cfqq_process_refs(new_cfqq);
2478         /*
2479          * If the process for the cfqq has gone away, there is no
2480          * sense in merging the queues.
2481          */
2482         if (process_refs == 0 || new_process_refs == 0)
2483                 return;
2484
2485         /*
2486          * Merge in the direction of the lesser amount of work.
2487          */
2488         if (new_process_refs >= process_refs) {
2489                 cfqq->new_cfqq = new_cfqq;
2490                 new_cfqq->ref += process_refs;
2491         } else {
2492                 new_cfqq->new_cfqq = cfqq;
2493                 cfqq->ref += new_process_refs;
2494         }
2495 }
2496
2497 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2498                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2499 {
2500         struct cfq_queue *queue;
2501         int i;
2502         bool key_valid = false;
2503         unsigned long lowest_key = 0;
2504         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2505
2506         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2507                 /* select the one with lowest rb_key */
2508                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2509                 if (queue &&
2510                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2511                         lowest_key = queue->rb_key;
2512                         cur_best = i;
2513                         key_valid = true;
2514                 }
2515         }
2516
2517         return cur_best;
2518 }
2519
2520 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2521 {
2522         unsigned slice;
2523         unsigned count;
2524         struct cfq_rb_root *st;
2525         unsigned group_slice;
2526         enum wl_prio_t original_prio = cfqd->serving_prio;
2527
2528         /* Choose next priority. RT > BE > IDLE */
2529         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2530                 cfqd->serving_prio = RT_WORKLOAD;
2531         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2532                 cfqd->serving_prio = BE_WORKLOAD;
2533         else {
2534                 cfqd->serving_prio = IDLE_WORKLOAD;
2535                 cfqd->workload_expires = jiffies + 1;
2536                 return;
2537         }
2538
2539         if (original_prio != cfqd->serving_prio)
2540                 goto new_workload;
2541
2542         /*
2543          * For RT and BE, we have to choose also the type
2544          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2545          * expiration time
2546          */
2547         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2548         count = st->count;
2549
2550         /*
2551          * check workload expiration, and that we still have other queues ready
2552          */
2553         if (count && !time_after(jiffies, cfqd->workload_expires))
2554                 return;
2555
2556 new_workload:
2557         /* otherwise select new workload type */
2558         cfqd->serving_type =
2559                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2560         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2561         count = st->count;
2562
2563         /*
2564          * the workload slice is computed as a fraction of target latency
2565          * proportional to the number of queues in that workload, over
2566          * all the queues in the same priority class
2567          */
2568         group_slice = cfq_group_slice(cfqd, cfqg);
2569
2570         slice = group_slice * count /
2571                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2572                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2573
2574         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2575                 unsigned int tmp;
2576
2577                 /*
2578                  * Async queues are currently system wide. Just taking
2579                  * proportion of queues with-in same group will lead to higher
2580                  * async ratio system wide as generally root group is going
2581                  * to have higher weight. A more accurate thing would be to
2582                  * calculate system wide asnc/sync ratio.
2583                  */
2584                 tmp = cfqd->cfq_target_latency *
2585                         cfqg_busy_async_queues(cfqd, cfqg);
2586                 tmp = tmp/cfqd->busy_queues;
2587                 slice = min_t(unsigned, slice, tmp);
2588
2589                 /* async workload slice is scaled down according to
2590                  * the sync/async slice ratio. */
2591                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2592         } else
2593                 /* sync workload slice is at least 2 * cfq_slice_idle */
2594                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2595
2596         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2597         cfq_log(cfqd, "workload slice:%d", slice);
2598         cfqd->workload_expires = jiffies + slice;
2599 }
2600
2601 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2602 {
2603         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2604         struct cfq_group *cfqg;
2605
2606         if (RB_EMPTY_ROOT(&st->rb))
2607                 return NULL;
2608         cfqg = cfq_rb_first_group(st);
2609         update_min_vdisktime(st);
2610         return cfqg;
2611 }
2612
2613 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2614 {
2615         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2616
2617         cfqd->serving_group = cfqg;
2618
2619         /* Restore the workload type data */
2620         if (cfqg->saved_workload_slice) {
2621                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2622                 cfqd->serving_type = cfqg->saved_workload;
2623                 cfqd->serving_prio = cfqg->saved_serving_prio;
2624         } else
2625                 cfqd->workload_expires = jiffies - 1;
2626
2627         choose_service_tree(cfqd, cfqg);
2628 }
2629
2630 /*
2631  * Select a queue for service. If we have a current active queue,
2632  * check whether to continue servicing it, or retrieve and set a new one.
2633  */
2634 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2635 {
2636         struct cfq_queue *cfqq, *new_cfqq = NULL;
2637
2638         cfqq = cfqd->active_queue;
2639         if (!cfqq)
2640                 goto new_queue;
2641
2642         if (!cfqd->rq_queued)
2643                 return NULL;
2644
2645         /*
2646          * We were waiting for group to get backlogged. Expire the queue
2647          */
2648         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2649                 goto expire;
2650
2651         /*
2652          * The active queue has run out of time, expire it and select new.
2653          */
2654         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2655                 /*
2656                  * If slice had not expired at the completion of last request
2657                  * we might not have turned on wait_busy flag. Don't expire
2658                  * the queue yet. Allow the group to get backlogged.
2659                  *
2660                  * The very fact that we have used the slice, that means we
2661                  * have been idling all along on this queue and it should be
2662                  * ok to wait for this request to complete.
2663                  */
2664                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2665                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2666                         cfqq = NULL;
2667                         goto keep_queue;
2668                 } else
2669                         goto check_group_idle;
2670         }
2671
2672         /*
2673          * The active queue has requests and isn't expired, allow it to
2674          * dispatch.
2675          */
2676         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2677                 goto keep_queue;
2678
2679         /*
2680          * If another queue has a request waiting within our mean seek
2681          * distance, let it run.  The expire code will check for close
2682          * cooperators and put the close queue at the front of the service
2683          * tree.  If possible, merge the expiring queue with the new cfqq.
2684          */
2685         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2686         if (new_cfqq) {
2687                 if (!cfqq->new_cfqq)
2688                         cfq_setup_merge(cfqq, new_cfqq);
2689                 goto expire;
2690         }
2691
2692         /*
2693          * No requests pending. If the active queue still has requests in
2694          * flight or is idling for a new request, allow either of these
2695          * conditions to happen (or time out) before selecting a new queue.
2696          */
2697         if (timer_pending(&cfqd->idle_slice_timer)) {
2698                 cfqq = NULL;
2699                 goto keep_queue;
2700         }
2701
2702         /*
2703          * This is a deep seek queue, but the device is much faster than
2704          * the queue can deliver, don't idle
2705          **/
2706         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2707             (cfq_cfqq_slice_new(cfqq) ||
2708             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2709                 cfq_clear_cfqq_deep(cfqq);
2710                 cfq_clear_cfqq_idle_window(cfqq);
2711         }
2712
2713         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2714                 cfqq = NULL;
2715                 goto keep_queue;
2716         }
2717
2718         /*
2719          * If group idle is enabled and there are requests dispatched from
2720          * this group, wait for requests to complete.
2721          */
2722 check_group_idle:
2723         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2724             cfqq->cfqg->dispatched &&
2725             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2726                 cfqq = NULL;
2727                 goto keep_queue;
2728         }
2729
2730 expire:
2731         cfq_slice_expired(cfqd, 0);
2732 new_queue:
2733         /*
2734          * Current queue expired. Check if we have to switch to a new
2735          * service tree
2736          */
2737         if (!new_cfqq)
2738                 cfq_choose_cfqg(cfqd);
2739
2740         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2741 keep_queue:
2742         return cfqq;
2743 }
2744
2745 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2746 {
2747         int dispatched = 0;
2748
2749         while (cfqq->next_rq) {
2750                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2751                 dispatched++;
2752         }
2753
2754         BUG_ON(!list_empty(&cfqq->fifo));
2755
2756         /* By default cfqq is not expired if it is empty. Do it explicitly */
2757         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2758         return dispatched;
2759 }
2760
2761 /*
2762  * Drain our current requests. Used for barriers and when switching
2763  * io schedulers on-the-fly.
2764  */
2765 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2766 {
2767         struct cfq_queue *cfqq;
2768         int dispatched = 0;
2769
2770         /* Expire the timeslice of the current active queue first */
2771         cfq_slice_expired(cfqd, 0);
2772         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2773                 __cfq_set_active_queue(cfqd, cfqq);
2774                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2775         }
2776
2777         BUG_ON(cfqd->busy_queues);
2778
2779         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2780         return dispatched;
2781 }
2782
2783 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2784         struct cfq_queue *cfqq)
2785 {
2786         /* the queue hasn't finished any request, can't estimate */
2787         if (cfq_cfqq_slice_new(cfqq))
2788                 return true;
2789         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2790                 cfqq->slice_end))
2791                 return true;
2792
2793         return false;
2794 }
2795
2796 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2797 {
2798         unsigned int max_dispatch;
2799
2800         /*
2801          * Drain async requests before we start sync IO
2802          */
2803         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2804                 return false;
2805
2806         /*
2807          * If this is an async queue and we have sync IO in flight, let it wait
2808          */
2809         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2810                 return false;
2811
2812         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2813         if (cfq_class_idle(cfqq))
2814                 max_dispatch = 1;
2815
2816         /*
2817          * Does this cfqq already have too much IO in flight?
2818          */
2819         if (cfqq->dispatched >= max_dispatch) {
2820                 bool promote_sync = false;
2821                 /*
2822                  * idle queue must always only have a single IO in flight
2823                  */
2824                 if (cfq_class_idle(cfqq))
2825                         return false;
2826
2827                 /*
2828                  * If there is only one sync queue
2829                  * we can ignore async queue here and give the sync
2830                  * queue no dispatch limit. The reason is a sync queue can
2831                  * preempt async queue, limiting the sync queue doesn't make
2832                  * sense. This is useful for aiostress test.
2833                  */
2834                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2835                         promote_sync = true;
2836
2837                 /*
2838                  * We have other queues, don't allow more IO from this one
2839                  */
2840                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2841                                 !promote_sync)
2842                         return false;
2843
2844                 /*
2845                  * Sole queue user, no limit
2846                  */
2847                 if (cfqd->busy_queues == 1 || promote_sync)
2848                         max_dispatch = -1;
2849                 else
2850                         /*
2851                          * Normally we start throttling cfqq when cfq_quantum/2
2852                          * requests have been dispatched. But we can drive
2853                          * deeper queue depths at the beginning of slice
2854                          * subjected to upper limit of cfq_quantum.
2855                          * */
2856                         max_dispatch = cfqd->cfq_quantum;
2857         }
2858
2859         /*
2860          * Async queues must wait a bit before being allowed dispatch.
2861          * We also ramp up the dispatch depth gradually for async IO,
2862          * based on the last sync IO we serviced
2863          */
2864         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2865                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2866                 unsigned int depth;
2867
2868                 depth = last_sync / cfqd->cfq_slice[1];
2869                 if (!depth && !cfqq->dispatched)
2870                         depth = 1;
2871                 if (depth < max_dispatch)
2872                         max_dispatch = depth;
2873         }
2874
2875         /*
2876          * If we're below the current max, allow a dispatch
2877          */
2878         return cfqq->dispatched < max_dispatch;
2879 }
2880
2881 /*
2882  * Dispatch a request from cfqq, moving them to the request queue
2883  * dispatch list.
2884  */
2885 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2886 {
2887         struct request *rq;
2888
2889         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2890
2891         if (!cfq_may_dispatch(cfqd, cfqq))
2892                 return false;
2893
2894         /*
2895          * follow expired path, else get first next available
2896          */
2897         rq = cfq_check_fifo(cfqq);
2898         if (!rq)
2899                 rq = cfqq->next_rq;
2900
2901         /*
2902          * insert request into driver dispatch list
2903          */
2904         cfq_dispatch_insert(cfqd->queue, rq);
2905
2906         if (!cfqd->active_cic) {
2907                 struct cfq_io_cq *cic = RQ_CIC(rq);
2908
2909                 atomic_long_inc(&cic->icq.ioc->refcount);
2910                 cfqd->active_cic = cic;
2911         }
2912
2913         return true;
2914 }
2915
2916 /*
2917  * Find the cfqq that we need to service and move a request from that to the
2918  * dispatch list
2919  */
2920 static int cfq_dispatch_requests(struct request_queue *q, int force)
2921 {
2922         struct cfq_data *cfqd = q->elevator->elevator_data;
2923         struct cfq_queue *cfqq;
2924
2925         if (!cfqd->busy_queues)
2926                 return 0;
2927
2928         if (unlikely(force))
2929                 return cfq_forced_dispatch(cfqd);
2930
2931         cfqq = cfq_select_queue(cfqd);
2932         if (!cfqq)
2933                 return 0;
2934
2935         /*
2936          * Dispatch a request from this cfqq, if it is allowed
2937          */
2938         if (!cfq_dispatch_request(cfqd, cfqq))
2939                 return 0;
2940
2941         cfqq->slice_dispatch++;
2942         cfq_clear_cfqq_must_dispatch(cfqq);
2943
2944         /*
2945          * expire an async queue immediately if it has used up its slice. idle
2946          * queue always expire after 1 dispatch round.
2947          */
2948         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2949             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2950             cfq_class_idle(cfqq))) {
2951                 cfqq->slice_end = jiffies + 1;
2952                 cfq_slice_expired(cfqd, 0);
2953         }
2954
2955         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2956         return 1;
2957 }
2958
2959 /*
2960  * task holds one reference to the queue, dropped when task exits. each rq
2961  * in-flight on this queue also holds a reference, dropped when rq is freed.
2962  *
2963  * Each cfq queue took a reference on the parent group. Drop it now.
2964  * queue lock must be held here.
2965  */
2966 static void cfq_put_queue(struct cfq_queue *cfqq)
2967 {
2968         struct cfq_data *cfqd = cfqq->cfqd;
2969         struct cfq_group *cfqg;
2970
2971         BUG_ON(cfqq->ref <= 0);
2972
2973         cfqq->ref--;
2974         if (cfqq->ref)
2975                 return;
2976
2977         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2978         BUG_ON(rb_first(&cfqq->sort_list));
2979         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2980         cfqg = cfqq->cfqg;
2981
2982         if (unlikely(cfqd->active_queue == cfqq)) {
2983                 __cfq_slice_expired(cfqd, cfqq, 0);
2984                 cfq_schedule_dispatch(cfqd);
2985         }
2986
2987         BUG_ON(cfq_cfqq_on_rr(cfqq));
2988         kmem_cache_free(cfq_pool, cfqq);
2989         cfqg_put(cfqg);
2990 }
2991
2992 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2993 {
2994         struct cfq_queue *__cfqq, *next;
2995
2996         /*
2997          * If this queue was scheduled to merge with another queue, be
2998          * sure to drop the reference taken on that queue (and others in
2999          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3000          */
3001         __cfqq = cfqq->new_cfqq;
3002         while (__cfqq) {
3003                 if (__cfqq == cfqq) {
3004                         WARN(1, "cfqq->new_cfqq loop detected\n");
3005                         break;
3006                 }
3007                 next = __cfqq->new_cfqq;
3008                 cfq_put_queue(__cfqq);
3009                 __cfqq = next;
3010         }
3011 }
3012
3013 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3014 {
3015         if (unlikely(cfqq == cfqd->active_queue)) {
3016                 __cfq_slice_expired(cfqd, cfqq, 0);
3017                 cfq_schedule_dispatch(cfqd);
3018         }
3019
3020         cfq_put_cooperator(cfqq);
3021
3022         cfq_put_queue(cfqq);
3023 }
3024
3025 static void cfq_init_icq(struct io_cq *icq)
3026 {
3027         struct cfq_io_cq *cic = icq_to_cic(icq);
3028
3029         cic->ttime.last_end_request = jiffies;
3030 }
3031
3032 static void cfq_exit_icq(struct io_cq *icq)
3033 {
3034         struct cfq_io_cq *cic = icq_to_cic(icq);
3035         struct cfq_data *cfqd = cic_to_cfqd(cic);
3036
3037         if (cic->cfqq[BLK_RW_ASYNC]) {
3038                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3039                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3040         }
3041
3042         if (cic->cfqq[BLK_RW_SYNC]) {
3043                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3044                 cic->cfqq[BLK_RW_SYNC] = NULL;
3045         }
3046 }
3047
3048 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3049 {
3050         struct task_struct *tsk = current;
3051         int ioprio_class;
3052
3053         if (!cfq_cfqq_prio_changed(cfqq))
3054                 return;
3055
3056         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3057         switch (ioprio_class) {
3058         default:
3059                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3060         case IOPRIO_CLASS_NONE:
3061                 /*
3062                  * no prio set, inherit CPU scheduling settings
3063                  */
3064                 cfqq->ioprio = task_nice_ioprio(tsk);
3065                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3066                 break;
3067         case IOPRIO_CLASS_RT:
3068                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3069                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3070                 break;
3071         case IOPRIO_CLASS_BE:
3072                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3073                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3074                 break;
3075         case IOPRIO_CLASS_IDLE:
3076                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3077                 cfqq->ioprio = 7;
3078                 cfq_clear_cfqq_idle_window(cfqq);
3079                 break;
3080         }
3081
3082         /*
3083          * keep track of original prio settings in case we have to temporarily
3084          * elevate the priority of this queue
3085          */
3086         cfqq->org_ioprio = cfqq->ioprio;
3087         cfq_clear_cfqq_prio_changed(cfqq);
3088 }
3089
3090 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3091 {
3092         int ioprio = cic->icq.ioc->ioprio;
3093         struct cfq_data *cfqd = cic_to_cfqd(cic);
3094         struct cfq_queue *cfqq;
3095
3096         /*
3097          * Check whether ioprio has changed.  The condition may trigger
3098          * spuriously on a newly created cic but there's no harm.
3099          */
3100         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3101                 return;
3102
3103         cfqq = cic->cfqq[BLK_RW_ASYNC];
3104         if (cfqq) {
3105                 struct cfq_queue *new_cfqq;
3106                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3107                                          GFP_ATOMIC);
3108                 if (new_cfqq) {
3109                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3110                         cfq_put_queue(cfqq);
3111                 }
3112         }
3113
3114         cfqq = cic->cfqq[BLK_RW_SYNC];
3115         if (cfqq)
3116                 cfq_mark_cfqq_prio_changed(cfqq);
3117
3118         cic->ioprio = ioprio;
3119 }
3120
3121 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3122                           pid_t pid, bool is_sync)
3123 {
3124         RB_CLEAR_NODE(&cfqq->rb_node);
3125         RB_CLEAR_NODE(&cfqq->p_node);
3126         INIT_LIST_HEAD(&cfqq->fifo);
3127
3128         cfqq->ref = 0;
3129         cfqq->cfqd = cfqd;
3130
3131         cfq_mark_cfqq_prio_changed(cfqq);
3132
3133         if (is_sync) {
3134                 if (!cfq_class_idle(cfqq))
3135                         cfq_mark_cfqq_idle_window(cfqq);
3136                 cfq_mark_cfqq_sync(cfqq);
3137         }
3138         cfqq->pid = pid;
3139 }
3140
3141 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3142 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3143 {
3144         struct cfq_data *cfqd = cic_to_cfqd(cic);
3145         struct cfq_queue *sync_cfqq;
3146         uint64_t id;
3147
3148         rcu_read_lock();
3149         id = bio_blkcg(bio)->id;
3150         rcu_read_unlock();
3151
3152         /*
3153          * Check whether blkcg has changed.  The condition may trigger
3154          * spuriously on a newly created cic but there's no harm.
3155          */
3156         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3157                 return;
3158
3159         sync_cfqq = cic_to_cfqq(cic, 1);
3160         if (sync_cfqq) {
3161                 /*
3162                  * Drop reference to sync queue. A new sync queue will be
3163                  * assigned in new group upon arrival of a fresh request.
3164                  */
3165                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3166                 cic_set_cfqq(cic, NULL, 1);
3167                 cfq_put_queue(sync_cfqq);
3168         }
3169
3170         cic->blkcg_id = id;
3171 }
3172 #else
3173 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3174 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3175
3176 static struct cfq_queue *
3177 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3178                      struct bio *bio, gfp_t gfp_mask)
3179 {
3180         struct blkcg *blkcg;
3181         struct cfq_queue *cfqq, *new_cfqq = NULL;
3182         struct cfq_group *cfqg;
3183
3184 retry:
3185         rcu_read_lock();
3186
3187         blkcg = bio_blkcg(bio);
3188         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3189         cfqq = cic_to_cfqq(cic, is_sync);
3190
3191         /*
3192          * Always try a new alloc if we fell back to the OOM cfqq
3193          * originally, since it should just be a temporary situation.
3194          */
3195         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3196                 cfqq = NULL;
3197                 if (new_cfqq) {
3198                         cfqq = new_cfqq;
3199                         new_cfqq = NULL;
3200                 } else if (gfp_mask & __GFP_WAIT) {
3201                         rcu_read_unlock();
3202                         spin_unlock_irq(cfqd->queue->queue_lock);
3203                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3204                                         gfp_mask | __GFP_ZERO,
3205                                         cfqd->queue->node);
3206                         spin_lock_irq(cfqd->queue->queue_lock);
3207                         if (new_cfqq)
3208                                 goto retry;
3209                 } else {
3210                         cfqq = kmem_cache_alloc_node(cfq_pool,
3211                                         gfp_mask | __GFP_ZERO,
3212                                         cfqd->queue->node);
3213                 }
3214
3215                 if (cfqq) {
3216                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3217                         cfq_init_prio_data(cfqq, cic);
3218                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3219                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3220                 } else
3221                         cfqq = &cfqd->oom_cfqq;
3222         }
3223
3224         if (new_cfqq)
3225                 kmem_cache_free(cfq_pool, new_cfqq);
3226
3227         rcu_read_unlock();
3228         return cfqq;
3229 }
3230
3231 static struct cfq_queue **
3232 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3233 {
3234         switch (ioprio_class) {
3235         case IOPRIO_CLASS_RT:
3236                 return &cfqd->async_cfqq[0][ioprio];
3237         case IOPRIO_CLASS_NONE:
3238                 ioprio = IOPRIO_NORM;
3239                 /* fall through */
3240         case IOPRIO_CLASS_BE:
3241                 return &cfqd->async_cfqq[1][ioprio];
3242         case IOPRIO_CLASS_IDLE:
3243                 return &cfqd->async_idle_cfqq;
3244         default:
3245                 BUG();
3246         }
3247 }
3248
3249 static struct cfq_queue *
3250 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3251               struct bio *bio, gfp_t gfp_mask)
3252 {
3253         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3254         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3255         struct cfq_queue **async_cfqq = NULL;
3256         struct cfq_queue *cfqq = NULL;
3257
3258         if (!is_sync) {
3259                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3260                 cfqq = *async_cfqq;
3261         }
3262
3263         if (!cfqq)
3264                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3265
3266         /*
3267          * pin the queue now that it's allocated, scheduler exit will prune it
3268          */
3269         if (!is_sync && !(*async_cfqq)) {
3270                 cfqq->ref++;
3271                 *async_cfqq = cfqq;
3272         }
3273
3274         cfqq->ref++;
3275         return cfqq;
3276 }
3277
3278 static void
3279 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3280 {
3281         unsigned long elapsed = jiffies - ttime->last_end_request;
3282         elapsed = min(elapsed, 2UL * slice_idle);
3283
3284         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3285         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3286         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3287 }
3288
3289 static void
3290 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3291                         struct cfq_io_cq *cic)
3292 {
3293         if (cfq_cfqq_sync(cfqq)) {
3294                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3295                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3296                         cfqd->cfq_slice_idle);
3297         }
3298 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3299         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3300 #endif
3301 }
3302
3303 static void
3304 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3305                        struct request *rq)
3306 {
3307         sector_t sdist = 0;
3308         sector_t n_sec = blk_rq_sectors(rq);
3309         if (cfqq->last_request_pos) {
3310                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3311                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3312                 else
3313                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3314         }
3315
3316         cfqq->seek_history <<= 1;
3317         if (blk_queue_nonrot(cfqd->queue))
3318                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3319         else
3320                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3321 }
3322
3323 /*
3324  * Disable idle window if the process thinks too long or seeks so much that
3325  * it doesn't matter
3326  */
3327 static void
3328 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3329                        struct cfq_io_cq *cic)
3330 {
3331         int old_idle, enable_idle;
3332
3333         /*
3334          * Don't idle for async or idle io prio class
3335          */
3336         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3337                 return;
3338
3339         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3340
3341         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3342                 cfq_mark_cfqq_deep(cfqq);
3343
3344         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3345                 enable_idle = 0;
3346         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3347                  !cfqd->cfq_slice_idle ||
3348                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3349                 enable_idle = 0;
3350         else if (sample_valid(cic->ttime.ttime_samples)) {
3351                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3352                         enable_idle = 0;
3353                 else
3354                         enable_idle = 1;
3355         }
3356
3357         if (old_idle != enable_idle) {
3358                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3359                 if (enable_idle)
3360                         cfq_mark_cfqq_idle_window(cfqq);
3361                 else
3362                         cfq_clear_cfqq_idle_window(cfqq);
3363         }
3364 }
3365
3366 /*
3367  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3368  * no or if we aren't sure, a 1 will cause a preempt.
3369  */
3370 static bool
3371 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3372                    struct request *rq)
3373 {
3374         struct cfq_queue *cfqq;
3375
3376         cfqq = cfqd->active_queue;
3377         if (!cfqq)
3378                 return false;
3379
3380         if (cfq_class_idle(new_cfqq))
3381                 return false;
3382
3383         if (cfq_class_idle(cfqq))
3384                 return true;
3385
3386         /*
3387          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3388          */
3389         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3390                 return false;
3391
3392         /*
3393          * if the new request is sync, but the currently running queue is
3394          * not, let the sync request have priority.
3395          */
3396         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3397                 return true;
3398
3399         if (new_cfqq->cfqg != cfqq->cfqg)
3400                 return false;
3401
3402         if (cfq_slice_used(cfqq))
3403                 return true;
3404
3405         /* Allow preemption only if we are idling on sync-noidle tree */
3406         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3407             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3408             new_cfqq->service_tree->count == 2 &&
3409             RB_EMPTY_ROOT(&cfqq->sort_list))
3410                 return true;
3411
3412         /*
3413          * So both queues are sync. Let the new request get disk time if
3414          * it's a metadata request and the current queue is doing regular IO.
3415          */
3416         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3417                 return true;
3418
3419         /*
3420          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3421          */
3422         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3423                 return true;
3424
3425         /* An idle queue should not be idle now for some reason */
3426         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3427                 return true;
3428
3429         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3430                 return false;
3431
3432         /*
3433          * if this request is as-good as one we would expect from the
3434          * current cfqq, let it preempt
3435          */
3436         if (cfq_rq_close(cfqd, cfqq, rq))
3437                 return true;
3438
3439         return false;
3440 }
3441
3442 /*
3443  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3444  * let it have half of its nominal slice.
3445  */
3446 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3447 {
3448         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3449
3450         cfq_log_cfqq(cfqd, cfqq, "preempt");
3451         cfq_slice_expired(cfqd, 1);
3452
3453         /*
3454          * workload type is changed, don't save slice, otherwise preempt
3455          * doesn't happen
3456          */
3457         if (old_type != cfqq_type(cfqq))
3458                 cfqq->cfqg->saved_workload_slice = 0;
3459
3460         /*
3461          * Put the new queue at the front of the of the current list,
3462          * so we know that it will be selected next.
3463          */
3464         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3465
3466         cfq_service_tree_add(cfqd, cfqq, 1);
3467
3468         cfqq->slice_end = 0;
3469         cfq_mark_cfqq_slice_new(cfqq);
3470 }
3471
3472 /*
3473  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3474  * something we should do about it
3475  */
3476 static void
3477 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3478                 struct request *rq)
3479 {
3480         struct cfq_io_cq *cic = RQ_CIC(rq);
3481
3482         cfqd->rq_queued++;
3483         if (rq->cmd_flags & REQ_PRIO)
3484                 cfqq->prio_pending++;
3485
3486         cfq_update_io_thinktime(cfqd, cfqq, cic);
3487         cfq_update_io_seektime(cfqd, cfqq, rq);
3488         cfq_update_idle_window(cfqd, cfqq, cic);
3489
3490         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3491
3492         if (cfqq == cfqd->active_queue) {
3493                 /*
3494                  * Remember that we saw a request from this process, but
3495                  * don't start queuing just yet. Otherwise we risk seeing lots
3496                  * of tiny requests, because we disrupt the normal plugging
3497                  * and merging. If the request is already larger than a single
3498                  * page, let it rip immediately. For that case we assume that
3499                  * merging is already done. Ditto for a busy system that
3500                  * has other work pending, don't risk delaying until the
3501                  * idle timer unplug to continue working.
3502                  */
3503                 if (cfq_cfqq_wait_request(cfqq)) {
3504                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3505                             cfqd->busy_queues > 1) {
3506                                 cfq_del_timer(cfqd, cfqq);
3507                                 cfq_clear_cfqq_wait_request(cfqq);
3508                                 __blk_run_queue(cfqd->queue);
3509                         } else {
3510                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3511                                 cfq_mark_cfqq_must_dispatch(cfqq);
3512                         }
3513                 }
3514         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3515                 /*
3516                  * not the active queue - expire current slice if it is
3517                  * idle and has expired it's mean thinktime or this new queue
3518                  * has some old slice time left and is of higher priority or
3519                  * this new queue is RT and the current one is BE
3520                  */
3521                 cfq_preempt_queue(cfqd, cfqq);
3522                 __blk_run_queue(cfqd->queue);
3523         }
3524 }
3525
3526 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3527 {
3528         struct cfq_data *cfqd = q->elevator->elevator_data;
3529         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3530
3531         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3532         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3533
3534         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3535         list_add_tail(&rq->queuelist, &cfqq->fifo);
3536         cfq_add_rq_rb(rq);
3537         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3538                                  rq->cmd_flags);
3539         cfq_rq_enqueued(cfqd, cfqq, rq);
3540 }
3541
3542 /*
3543  * Update hw_tag based on peak queue depth over 50 samples under
3544  * sufficient load.
3545  */
3546 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3547 {
3548         struct cfq_queue *cfqq = cfqd->active_queue;
3549
3550         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3551                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3552
3553         if (cfqd->hw_tag == 1)
3554                 return;
3555
3556         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3557             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3558                 return;
3559
3560         /*
3561          * If active queue hasn't enough requests and can idle, cfq might not
3562          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3563          * case
3564          */
3565         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3566             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3567             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3568                 return;
3569
3570         if (cfqd->hw_tag_samples++ < 50)
3571                 return;
3572
3573         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3574                 cfqd->hw_tag = 1;
3575         else
3576                 cfqd->hw_tag = 0;
3577 }
3578
3579 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3580 {
3581         struct cfq_io_cq *cic = cfqd->active_cic;
3582
3583         /* If the queue already has requests, don't wait */
3584         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3585                 return false;
3586
3587         /* If there are other queues in the group, don't wait */
3588         if (cfqq->cfqg->nr_cfqq > 1)
3589                 return false;
3590
3591         /* the only queue in the group, but think time is big */
3592         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3593                 return false;
3594
3595         if (cfq_slice_used(cfqq))
3596                 return true;
3597
3598         /* if slice left is less than think time, wait busy */
3599         if (cic && sample_valid(cic->ttime.ttime_samples)
3600             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3601                 return true;
3602
3603         /*
3604          * If think times is less than a jiffy than ttime_mean=0 and above
3605          * will not be true. It might happen that slice has not expired yet
3606          * but will expire soon (4-5 ns) during select_queue(). To cover the
3607          * case where think time is less than a jiffy, mark the queue wait
3608          * busy if only 1 jiffy is left in the slice.
3609          */
3610         if (cfqq->slice_end - jiffies == 1)
3611                 return true;
3612
3613         return false;
3614 }
3615
3616 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3617 {
3618         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3619         struct cfq_data *cfqd = cfqq->cfqd;
3620         const int sync = rq_is_sync(rq);
3621         unsigned long now;
3622
3623         now = jiffies;
3624         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3625                      !!(rq->cmd_flags & REQ_NOIDLE));
3626
3627         cfq_update_hw_tag(cfqd);
3628
3629         WARN_ON(!cfqd->rq_in_driver);
3630         WARN_ON(!cfqq->dispatched);
3631         cfqd->rq_in_driver--;
3632         cfqq->dispatched--;
3633         (RQ_CFQG(rq))->dispatched--;
3634         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3635                                      rq_io_start_time_ns(rq), rq->cmd_flags);
3636
3637         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3638
3639         if (sync) {
3640                 struct cfq_rb_root *service_tree;
3641
3642                 RQ_CIC(rq)->ttime.last_end_request = now;
3643
3644                 if (cfq_cfqq_on_rr(cfqq))
3645                         service_tree = cfqq->service_tree;
3646                 else
3647                         service_tree = service_tree_for(cfqq->cfqg,
3648                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3649                 service_tree->ttime.last_end_request = now;
3650                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3651                         cfqd->last_delayed_sync = now;
3652         }
3653
3654 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3655         cfqq->cfqg->ttime.last_end_request = now;
3656 #endif
3657
3658         /*
3659          * If this is the active queue, check if it needs to be expired,
3660          * or if we want to idle in case it has no pending requests.
3661          */
3662         if (cfqd->active_queue == cfqq) {
3663                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3664
3665                 if (cfq_cfqq_slice_new(cfqq)) {
3666                         cfq_set_prio_slice(cfqd, cfqq);
3667                         cfq_clear_cfqq_slice_new(cfqq);
3668                 }
3669
3670                 /*
3671                  * Should we wait for next request to come in before we expire
3672                  * the queue.
3673                  */
3674                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3675                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3676                         if (!cfqd->cfq_slice_idle)
3677                                 extend_sl = cfqd->cfq_group_idle;
3678                         cfqq->slice_end = jiffies + extend_sl;
3679                         cfq_mark_cfqq_wait_busy(cfqq);
3680                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3681                 }
3682
3683                 /*
3684                  * Idling is not enabled on:
3685                  * - expired queues
3686                  * - idle-priority queues
3687                  * - async queues
3688                  * - queues with still some requests queued
3689                  * - when there is a close cooperator
3690                  */
3691                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3692                         cfq_slice_expired(cfqd, 1);
3693                 else if (sync && cfqq_empty &&
3694                          !cfq_close_cooperator(cfqd, cfqq)) {
3695                         cfq_arm_slice_timer(cfqd);
3696                 }
3697         }
3698
3699         if (!cfqd->rq_in_driver)
3700                 cfq_schedule_dispatch(cfqd);
3701 }
3702
3703 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3704 {
3705         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3706                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3707                 return ELV_MQUEUE_MUST;
3708         }
3709
3710         return ELV_MQUEUE_MAY;
3711 }
3712
3713 static int cfq_may_queue(struct request_queue *q, int rw)
3714 {
3715         struct cfq_data *cfqd = q->elevator->elevator_data;
3716         struct task_struct *tsk = current;
3717         struct cfq_io_cq *cic;
3718         struct cfq_queue *cfqq;
3719
3720         /*
3721          * don't force setup of a queue from here, as a call to may_queue
3722          * does not necessarily imply that a request actually will be queued.
3723          * so just lookup a possibly existing queue, or return 'may queue'
3724          * if that fails
3725          */
3726         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3727         if (!cic)
3728                 return ELV_MQUEUE_MAY;
3729
3730         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3731         if (cfqq) {
3732                 cfq_init_prio_data(cfqq, cic);
3733
3734                 return __cfq_may_queue(cfqq);
3735         }
3736
3737         return ELV_MQUEUE_MAY;
3738 }
3739
3740 /*
3741  * queue lock held here
3742  */
3743 static void cfq_put_request(struct request *rq)
3744 {
3745         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3746
3747         if (cfqq) {
3748                 const int rw = rq_data_dir(rq);
3749
3750                 BUG_ON(!cfqq->allocated[rw]);
3751                 cfqq->allocated[rw]--;
3752
3753                 /* Put down rq reference on cfqg */
3754                 cfqg_put(RQ_CFQG(rq));
3755                 rq->elv.priv[0] = NULL;
3756                 rq->elv.priv[1] = NULL;
3757
3758                 cfq_put_queue(cfqq);
3759         }
3760 }
3761
3762 static struct cfq_queue *
3763 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3764                 struct cfq_queue *cfqq)
3765 {
3766         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3767         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3768         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3769         cfq_put_queue(cfqq);
3770         return cic_to_cfqq(cic, 1);
3771 }
3772
3773 /*
3774  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3775  * was the last process referring to said cfqq.
3776  */
3777 static struct cfq_queue *
3778 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3779 {
3780         if (cfqq_process_refs(cfqq) == 1) {
3781                 cfqq->pid = current->pid;
3782                 cfq_clear_cfqq_coop(cfqq);
3783                 cfq_clear_cfqq_split_coop(cfqq);
3784                 return cfqq;
3785         }
3786
3787         cic_set_cfqq(cic, NULL, 1);
3788
3789         cfq_put_cooperator(cfqq);
3790
3791         cfq_put_queue(cfqq);
3792         return NULL;
3793 }
3794 /*
3795  * Allocate cfq data structures associated with this request.
3796  */
3797 static int
3798 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
3799                 gfp_t gfp_mask)
3800 {
3801         struct cfq_data *cfqd = q->elevator->elevator_data;
3802         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3803         const int rw = rq_data_dir(rq);
3804         const bool is_sync = rq_is_sync(rq);
3805         struct cfq_queue *cfqq;
3806
3807         might_sleep_if(gfp_mask & __GFP_WAIT);
3808
3809         spin_lock_irq(q->queue_lock);
3810
3811         check_ioprio_changed(cic, bio);
3812         check_blkcg_changed(cic, bio);
3813 new_queue:
3814         cfqq = cic_to_cfqq(cic, is_sync);
3815         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3816                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
3817                 cic_set_cfqq(cic, cfqq, is_sync);
3818         } else {
3819                 /*
3820                  * If the queue was seeky for too long, break it apart.
3821                  */
3822                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3823                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3824                         cfqq = split_cfqq(cic, cfqq);
3825                         if (!cfqq)
3826                                 goto new_queue;
3827                 }
3828
3829                 /*
3830                  * Check to see if this queue is scheduled to merge with
3831                  * another, closely cooperating queue.  The merging of
3832                  * queues happens here as it must be done in process context.
3833                  * The reference on new_cfqq was taken in merge_cfqqs.
3834                  */
3835                 if (cfqq->new_cfqq)
3836                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3837         }
3838
3839         cfqq->allocated[rw]++;
3840
3841         cfqq->ref++;
3842         cfqg_get(cfqq->cfqg);
3843         rq->elv.priv[0] = cfqq;
3844         rq->elv.priv[1] = cfqq->cfqg;
3845         spin_unlock_irq(q->queue_lock);
3846         return 0;
3847 }
3848
3849 static void cfq_kick_queue(struct work_struct *work)
3850 {
3851         struct cfq_data *cfqd =
3852                 container_of(work, struct cfq_data, unplug_work);
3853         struct request_queue *q = cfqd->queue;
3854
3855         spin_lock_irq(q->queue_lock);
3856         __blk_run_queue(cfqd->queue);
3857         spin_unlock_irq(q->queue_lock);
3858 }
3859
3860 /*
3861  * Timer running if the active_queue is currently idling inside its time slice
3862  */
3863 static void cfq_idle_slice_timer(unsigned long data)
3864 {
3865         struct cfq_data *cfqd = (struct cfq_data *) data;
3866         struct cfq_queue *cfqq;
3867         unsigned long flags;
3868         int timed_out = 1;
3869
3870         cfq_log(cfqd, "idle timer fired");
3871
3872         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3873
3874         cfqq = cfqd->active_queue;
3875         if (cfqq) {
3876                 timed_out = 0;
3877
3878                 /*
3879                  * We saw a request before the queue expired, let it through
3880                  */
3881                 if (cfq_cfqq_must_dispatch(cfqq))
3882                         goto out_kick;
3883
3884                 /*
3885                  * expired
3886                  */
3887                 if (cfq_slice_used(cfqq))
3888                         goto expire;
3889
3890                 /*
3891                  * only expire and reinvoke request handler, if there are
3892                  * other queues with pending requests
3893                  */
3894                 if (!cfqd->busy_queues)
3895                         goto out_cont;
3896
3897                 /*
3898                  * not expired and it has a request pending, let it dispatch
3899                  */
3900                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3901                         goto out_kick;
3902
3903                 /*
3904                  * Queue depth flag is reset only when the idle didn't succeed
3905                  */
3906                 cfq_clear_cfqq_deep(cfqq);
3907         }
3908 expire:
3909         cfq_slice_expired(cfqd, timed_out);
3910 out_kick:
3911         cfq_schedule_dispatch(cfqd);
3912 out_cont:
3913         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3914 }
3915
3916 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3917 {
3918         del_timer_sync(&cfqd->idle_slice_timer);
3919         cancel_work_sync(&cfqd->unplug_work);
3920 }
3921
3922 static void cfq_put_async_queues(struct cfq_data *cfqd)
3923 {
3924         int i;
3925
3926         for (i = 0; i < IOPRIO_BE_NR; i++) {
3927                 if (cfqd->async_cfqq[0][i])
3928                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3929                 if (cfqd->async_cfqq[1][i])
3930                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3931         }
3932
3933         if (cfqd->async_idle_cfqq)
3934                 cfq_put_queue(cfqd->async_idle_cfqq);
3935 }
3936
3937 static void cfq_exit_queue(struct elevator_queue *e)
3938 {
3939         struct cfq_data *cfqd = e->elevator_data;
3940         struct request_queue *q = cfqd->queue;
3941
3942         cfq_shutdown_timer_wq(cfqd);
3943
3944         spin_lock_irq(q->queue_lock);
3945
3946         if (cfqd->active_queue)
3947                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3948
3949         cfq_put_async_queues(cfqd);
3950
3951         spin_unlock_irq(q->queue_lock);
3952
3953         cfq_shutdown_timer_wq(cfqd);
3954
3955 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3956         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
3957 #else
3958         kfree(cfqd->root_group);
3959 #endif
3960         kfree(cfqd);
3961 }
3962
3963 static int cfq_init_queue(struct request_queue *q)
3964 {
3965         struct cfq_data *cfqd;
3966         struct blkcg_gq *blkg __maybe_unused;
3967         int i, ret;
3968
3969         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3970         if (!cfqd)
3971                 return -ENOMEM;
3972
3973         cfqd->queue = q;
3974         q->elevator->elevator_data = cfqd;
3975
3976         /* Init root service tree */
3977         cfqd->grp_service_tree = CFQ_RB_ROOT;
3978
3979         /* Init root group and prefer root group over other groups by default */
3980 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3981         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
3982         if (ret)
3983                 goto out_free;
3984
3985         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
3986 #else
3987         ret = -ENOMEM;
3988         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3989                                         GFP_KERNEL, cfqd->queue->node);
3990         if (!cfqd->root_group)
3991                 goto out_free;
3992
3993         cfq_init_cfqg_base(cfqd->root_group);
3994 #endif
3995         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
3996
3997         /*
3998          * Not strictly needed (since RB_ROOT just clears the node and we
3999          * zeroed cfqd on alloc), but better be safe in case someone decides
4000          * to add magic to the rb code
4001          */
4002         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4003                 cfqd->prio_trees[i] = RB_ROOT;
4004
4005         /*
4006          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4007          * Grab a permanent reference to it, so that the normal code flow
4008          * will not attempt to free it.  oom_cfqq is linked to root_group
4009          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4010          * the reference from linking right away.
4011          */
4012         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4013         cfqd->oom_cfqq.ref++;
4014
4015         spin_lock_irq(q->queue_lock);
4016         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4017         cfqg_put(cfqd->root_group);
4018         spin_unlock_irq(q->queue_lock);
4019
4020         init_timer(&cfqd->idle_slice_timer);
4021         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4022         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4023
4024         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4025
4026         cfqd->cfq_quantum = cfq_quantum;
4027         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4028         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4029         cfqd->cfq_back_max = cfq_back_max;
4030         cfqd->cfq_back_penalty = cfq_back_penalty;
4031         cfqd->cfq_slice[0] = cfq_slice_async;
4032         cfqd->cfq_slice[1] = cfq_slice_sync;
4033         cfqd->cfq_target_latency = cfq_target_latency;
4034         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4035         cfqd->cfq_slice_idle = cfq_slice_idle;
4036         cfqd->cfq_group_idle = cfq_group_idle;
4037         cfqd->cfq_latency = 1;
4038         cfqd->hw_tag = -1;
4039         /*
4040          * we optimistically start assuming sync ops weren't delayed in last
4041          * second, in order to have larger depth for async operations.
4042          */
4043         cfqd->last_delayed_sync = jiffies - HZ;
4044         return 0;
4045
4046 out_free:
4047         kfree(cfqd);
4048         return ret;
4049 }
4050
4051 /*
4052  * sysfs parts below -->
4053  */
4054 static ssize_t
4055 cfq_var_show(unsigned int var, char *page)
4056 {
4057         return sprintf(page, "%d\n", var);
4058 }
4059
4060 static ssize_t
4061 cfq_var_store(unsigned int *var, const char *page, size_t count)
4062 {
4063         char *p = (char *) page;
4064
4065         *var = simple_strtoul(p, &p, 10);
4066         return count;
4067 }
4068
4069 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4070 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4071 {                                                                       \
4072         struct cfq_data *cfqd = e->elevator_data;                       \
4073         unsigned int __data = __VAR;                                    \
4074         if (__CONV)                                                     \
4075                 __data = jiffies_to_msecs(__data);                      \
4076         return cfq_var_show(__data, (page));                            \
4077 }
4078 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4079 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4080 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4081 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4082 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4083 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4084 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4085 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4086 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4087 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4088 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4089 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4090 #undef SHOW_FUNCTION
4091
4092 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4093 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4094 {                                                                       \
4095         struct cfq_data *cfqd = e->elevator_data;                       \
4096         unsigned int __data;                                            \
4097         int ret = cfq_var_store(&__data, (page), count);                \
4098         if (__data < (MIN))                                             \
4099                 __data = (MIN);                                         \
4100         else if (__data > (MAX))                                        \
4101                 __data = (MAX);                                         \
4102         if (__CONV)                                                     \
4103                 *(__PTR) = msecs_to_jiffies(__data);                    \
4104         else                                                            \
4105                 *(__PTR) = __data;                                      \
4106         return ret;                                                     \
4107 }
4108 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4109 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4110                 UINT_MAX, 1);
4111 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4112                 UINT_MAX, 1);
4113 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4114 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4115                 UINT_MAX, 0);
4116 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4117 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4118 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4119 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4120 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4121                 UINT_MAX, 0);
4122 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4123 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4124 #undef STORE_FUNCTION
4125
4126 #define CFQ_ATTR(name) \
4127         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4128
4129 static struct elv_fs_entry cfq_attrs[] = {
4130         CFQ_ATTR(quantum),
4131         CFQ_ATTR(fifo_expire_sync),
4132         CFQ_ATTR(fifo_expire_async),
4133         CFQ_ATTR(back_seek_max),
4134         CFQ_ATTR(back_seek_penalty),
4135         CFQ_ATTR(slice_sync),
4136         CFQ_ATTR(slice_async),
4137         CFQ_ATTR(slice_async_rq),
4138         CFQ_ATTR(slice_idle),
4139         CFQ_ATTR(group_idle),
4140         CFQ_ATTR(low_latency),
4141         CFQ_ATTR(target_latency),
4142         __ATTR_NULL
4143 };
4144
4145 static struct elevator_type iosched_cfq = {
4146         .ops = {
4147                 .elevator_merge_fn =            cfq_merge,
4148                 .elevator_merged_fn =           cfq_merged_request,
4149                 .elevator_merge_req_fn =        cfq_merged_requests,
4150                 .elevator_allow_merge_fn =      cfq_allow_merge,
4151                 .elevator_bio_merged_fn =       cfq_bio_merged,
4152                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4153                 .elevator_add_req_fn =          cfq_insert_request,
4154                 .elevator_activate_req_fn =     cfq_activate_request,
4155                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4156                 .elevator_completed_req_fn =    cfq_completed_request,
4157                 .elevator_former_req_fn =       elv_rb_former_request,
4158                 .elevator_latter_req_fn =       elv_rb_latter_request,
4159                 .elevator_init_icq_fn =         cfq_init_icq,
4160                 .elevator_exit_icq_fn =         cfq_exit_icq,
4161                 .elevator_set_req_fn =          cfq_set_request,
4162                 .elevator_put_req_fn =          cfq_put_request,
4163                 .elevator_may_queue_fn =        cfq_may_queue,
4164                 .elevator_init_fn =             cfq_init_queue,
4165                 .elevator_exit_fn =             cfq_exit_queue,
4166         },
4167         .icq_size       =       sizeof(struct cfq_io_cq),
4168         .icq_align      =       __alignof__(struct cfq_io_cq),
4169         .elevator_attrs =       cfq_attrs,
4170         .elevator_name  =       "cfq",
4171         .elevator_owner =       THIS_MODULE,
4172 };
4173
4174 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4175 static struct blkcg_policy blkcg_policy_cfq = {
4176         .pd_size                = sizeof(struct cfq_group),
4177         .cftypes                = cfq_blkcg_files,
4178
4179         .pd_init_fn             = cfq_pd_init,
4180         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4181 };
4182 #endif
4183
4184 static int __init cfq_init(void)
4185 {
4186         int ret;
4187
4188         /*
4189          * could be 0 on HZ < 1000 setups
4190          */
4191         if (!cfq_slice_async)
4192                 cfq_slice_async = 1;
4193         if (!cfq_slice_idle)
4194                 cfq_slice_idle = 1;
4195
4196 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4197         if (!cfq_group_idle)
4198                 cfq_group_idle = 1;
4199
4200         ret = blkcg_policy_register(&blkcg_policy_cfq);
4201         if (ret)
4202                 return ret;
4203 #else
4204         cfq_group_idle = 0;
4205 #endif
4206
4207         ret = -ENOMEM;
4208         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4209         if (!cfq_pool)
4210                 goto err_pol_unreg;
4211
4212         ret = elv_register(&iosched_cfq);
4213         if (ret)
4214                 goto err_free_pool;
4215
4216         return 0;
4217
4218 err_free_pool:
4219         kmem_cache_destroy(cfq_pool);
4220 err_pol_unreg:
4221 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4222         blkcg_policy_unregister(&blkcg_policy_cfq);
4223 #endif
4224         return ret;
4225 }
4226
4227 static void __exit cfq_exit(void)
4228 {
4229 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4230         blkcg_policy_unregister(&blkcg_policy_cfq);
4231 #endif
4232         elv_unregister(&iosched_cfq);
4233         kmem_cache_destroy(cfq_pool);
4234 }
4235
4236 module_init(cfq_init);
4237 module_exit(cfq_exit);
4238
4239 MODULE_AUTHOR("Jens Axboe");
4240 MODULE_LICENSE("GPL");
4241 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");