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         u64 min_vdisktime;
  89         struct cfq_ttime ttime;
  90 };
  91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  92                         .ttime = {.last_end_request = jiffies,},}
  93
  94 /*
  95  * Per process-grouping structure
  96  */
  97 struct cfq_queue {
  98         /* reference count */
  99         int ref;
 100         /* various state flags, see below */
 101         unsigned int flags;
 102         /* parent cfq_data */
 103         struct cfq_data *cfqd;
 104         /* service_tree member */
 105         struct rb_node rb_node;
 106         /* service_tree key */
 107         unsigned long rb_key;
 108         /* prio tree member */
 109         struct rb_node p_node;
 110         /* prio tree root we belong to, if any */
 111         struct rb_root *p_root;
 112         /* sorted list of pending requests */
 113         struct rb_root sort_list;
 114         /* if fifo isn't expired, next request to serve */
 115         struct request *next_rq;
 116         /* requests queued in sort_list */
 117         int queued[2];
 118         /* currently allocated requests */
 119         int allocated[2];
 120         /* fifo list of requests in sort_list */
 121         struct list_head fifo;
 122
 123         /* time when queue got scheduled in to dispatch first request. */
 124         unsigned long dispatch_start;
 125         unsigned int allocated_slice;
 126         unsigned int slice_dispatch;
 127         /* time when first request from queue completed and slice started. */
 128         unsigned long slice_start;
 129         unsigned long slice_end;
 130         long slice_resid;
 131
 132         /* pending priority requests */
 133         int prio_pending;
 134         /* number of requests that are on the dispatch list or inside driver */
 135         int dispatched;
 136
 137         /* io prio of this group */
 138         unsigned short ioprio, org_ioprio;
 139         unsigned short ioprio_class;
 140
 141         pid_t pid;
 142
 143         u32 seek_history;
 144         sector_t last_request_pos;
 145
 146         struct cfq_rb_root *service_tree;
 147         struct cfq_queue *new_cfqq;
 148         struct cfq_group *cfqg;
 149         /* Number of sectors dispatched from queue in single dispatch round */
 150         unsigned long nr_sectors;
 151 };
 152
 153 /*
 154  * First index in the service_trees.
 155  * IDLE is handled separately, so it has negative index
 156  */
 157 enum wl_class_t {
 158         BE_WORKLOAD = 0,
 159         RT_WORKLOAD = 1,
 160         IDLE_WORKLOAD = 2,
 161         CFQ_PRIO_NR,
 162 };
 163
 164 /*
 165  * Second index in the service_trees.
 166  */
 167 enum wl_type_t {
 168         ASYNC_WORKLOAD = 0,
 169         SYNC_NOIDLE_WORKLOAD = 1,
 170         SYNC_WORKLOAD = 2
 171 };
 172
 173 struct cfqg_stats {
 174 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 175         /* total bytes transferred */
 176         struct blkg_rwstat              service_bytes;
 177         /* total IOs serviced, post merge */
 178         struct blkg_rwstat              serviced;
 179         /* number of ios merged */
 180         struct blkg_rwstat              merged;
 181         /* total time spent on device in ns, may not be accurate w/ queueing */
 182         struct blkg_rwstat              service_time;
 183         /* total time spent waiting in scheduler queue in ns */
 184         struct blkg_rwstat              wait_time;
 185         /* number of IOs queued up */
 186         struct blkg_rwstat              queued;
 187         /* total sectors transferred */
 188         struct blkg_stat                sectors;
 189         /* total disk time and nr sectors dispatched by this group */
 190         struct blkg_stat                time;
 191 #ifdef CONFIG_DEBUG_BLK_CGROUP
 192         /* time not charged to this cgroup */
 193         struct blkg_stat                unaccounted_time;
 194         /* sum of number of ios queued across all samples */
 195         struct blkg_stat                avg_queue_size_sum;
 196         /* count of samples taken for average */
 197         struct blkg_stat                avg_queue_size_samples;
 198         /* how many times this group has been removed from service tree */
 199         struct blkg_stat                dequeue;
 200         /* total time spent waiting for it to be assigned a timeslice. */
 201         struct blkg_stat                group_wait_time;
 202         /* time spent idling for this blkcg_gq */
 203         struct blkg_stat                idle_time;
 204         /* total time with empty current active q with other requests queued */
 205         struct blkg_stat                empty_time;
 206         /* fields after this shouldn't be cleared on stat reset */
 207         uint64_t                        start_group_wait_time;
 208         uint64_t                        start_idle_time;
 209         uint64_t                        start_empty_time;
 210         uint16_t                        flags;
 211 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 212 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 213 };
 214
 215 /* This is per cgroup per device grouping structure */
 216 struct cfq_group {
 217         /* must be the first member */
 218         struct blkg_policy_data pd;
 219
 220         /* group service_tree member */
 221         struct rb_node rb_node;
 222
 223         /* group service_tree key */
 224         u64 vdisktime;
 225
 226         /*
 227          * The number of active cfqgs and sum of their weights under this
 228          * cfqg.  This covers this cfqg's leaf_weight and all children's
 229          * weights, but does not cover weights of further descendants.
 230          *
 231          * If a cfqg is on the service tree, it's active.  An active cfqg
 232          * also activates its parent and contributes to the children_weight
 233          * of the parent.
 234          */
 235         int nr_active;
 236         unsigned int children_weight;
 237
 238         /*
 239          * vfraction is the fraction of vdisktime that the tasks in this
 240          * cfqg are entitled to.  This is determined by compounding the
 241          * ratios walking up from this cfqg to the root.
 242          *
 243          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 244          * vfractions on a service tree is approximately 1.  The sum may
 245          * deviate a bit due to rounding errors and fluctuations caused by
 246          * cfqgs entering and leaving the service tree.
 247          */
 248         unsigned int vfraction;
 249
 250         /*
 251          * There are two weights - (internal) weight is the weight of this
 252          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 253          * this cfqg against the child cfqgs.  For the root cfqg, both
 254          * weights are kept in sync for backward compatibility.
 255          */
 256         unsigned int weight;
 257         unsigned int new_weight;
 258         unsigned int dev_weight;
 259
 260         unsigned int leaf_weight;
 261         unsigned int new_leaf_weight;
 262         unsigned int dev_leaf_weight;
 263
 264         /* number of cfqq currently on this group */
 265         int nr_cfqq;
 266
 267         /*
 268          * Per group busy queues average. Useful for workload slice calc. We
 269          * create the array for each prio class but at run time it is used
 270          * only for RT and BE class and slot for IDLE class remains unused.
 271          * This is primarily done to avoid confusion and a gcc warning.
 272          */
 273         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 274         /*
 275          * rr lists of queues with requests. We maintain service trees for
 276          * RT and BE classes. These trees are subdivided in subclasses
 277          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 278          * class there is no subclassification and all the cfq queues go on
 279          * a single tree service_tree_idle.
 280          * Counts are embedded in the cfq_rb_root
 281          */
 282         struct cfq_rb_root service_trees[2][3];
 283         struct cfq_rb_root service_tree_idle;
 284
 285         unsigned long saved_wl_slice;
 286         enum wl_type_t saved_wl_type;
 287         enum wl_class_t saved_wl_class;
 288
 289         /* number of requests that are on the dispatch list or inside driver */
 290         int dispatched;
 291         struct cfq_ttime ttime;
 292         struct cfqg_stats stats;        /* stats for this cfqg */
 293         struct cfqg_stats dead_stats;   /* stats pushed from dead children */
 294 };
 295
 296 struct cfq_io_cq {
 297         struct io_cq            icq;            /* must be the first member */
 298         struct cfq_queue        *cfqq[2];
 299         struct cfq_ttime        ttime;
 300         int                     ioprio;         /* the current ioprio */
 301 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 302         uint64_t                blkcg_id;       /* the current blkcg ID */
 303 #endif
 304 };
 305
 306 /*
 307  * Per block device queue structure
 308  */
 309 struct cfq_data {
 310         struct request_queue *queue;
 311         /* Root service tree for cfq_groups */
 312         struct cfq_rb_root grp_service_tree;
 313         struct cfq_group *root_group;
 314
 315         /*
 316          * The priority currently being served
 317          */
 318         enum wl_class_t serving_wl_class;
 319         enum wl_type_t serving_wl_type;
 320         unsigned long workload_expires;
 321         struct cfq_group *serving_group;
 322
 323         /*
 324          * Each priority tree is sorted by next_request position.  These
 325          * trees are used when determining if two or more queues are
 326          * interleaving requests (see cfq_close_cooperator).
 327          */
 328         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 329
 330         unsigned int busy_queues;
 331         unsigned int busy_sync_queues;
 332
 333         int rq_in_driver;
 334         int rq_in_flight[2];
 335
 336         /*
 337          * queue-depth detection
 338          */
 339         int rq_queued;
 340         int hw_tag;
 341         /*
 342          * hw_tag can be
 343          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 344          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 345          *  0 => no NCQ
 346          */
 347         int hw_tag_est_depth;
 348         unsigned int hw_tag_samples;
 349
 350         /*
 351          * idle window management
 352          */
 353         struct timer_list idle_slice_timer;
 354         struct work_struct unplug_work;
 355
 356         struct cfq_queue *active_queue;
 357         struct cfq_io_cq *active_cic;
 358
 359         /*
 360          * async queue for each priority case
 361          */
 362         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 363         struct cfq_queue *async_idle_cfqq;
 364
 365         sector_t last_position;
 366
 367         /*
 368          * tunables, see top of file
 369          */
 370         unsigned int cfq_quantum;
 371         unsigned int cfq_fifo_expire[2];
 372         unsigned int cfq_back_penalty;
 373         unsigned int cfq_back_max;
 374         unsigned int cfq_slice[2];
 375         unsigned int cfq_slice_async_rq;
 376         unsigned int cfq_slice_idle;
 377         unsigned int cfq_group_idle;
 378         unsigned int cfq_latency;
 379         unsigned int cfq_target_latency;
 380
 381         /*
 382          * Fallback dummy cfqq for extreme OOM conditions
 383          */
 384         struct cfq_queue oom_cfqq;
 385
 386         unsigned long last_delayed_sync;
 387 };
 388
 389 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 390
 391 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 392                                             enum wl_class_t class,
 393                                             enum wl_type_t type)
 394 {
 395         if (!cfqg)
 396                 return NULL;
 397
 398         if (class == IDLE_WORKLOAD)
 399                 return &cfqg->service_tree_idle;
 400
 401         return &cfqg->service_trees[class][type];
 402 }
 403
 404 enum cfqq_state_flags {
 405         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 406         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 407         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 408         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 409         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 410         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 411         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 412         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 413         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 414         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 415         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 416         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 417         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 418 };
 419
 420 #define CFQ_CFQQ_FNS(name)                                              \
 421 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 422 {                                                                       \
 423         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 424 }                                                                       \
 425 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 426 {                                                                       \
 427         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 428 }                                                                       \
 429 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 430 {                                                                       \
 431         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 432 }
 433
 434 CFQ_CFQQ_FNS(on_rr);
 435 CFQ_CFQQ_FNS(wait_request);
 436 CFQ_CFQQ_FNS(must_dispatch);
 437 CFQ_CFQQ_FNS(must_alloc_slice);
 438 CFQ_CFQQ_FNS(fifo_expire);
 439 CFQ_CFQQ_FNS(idle_window);
 440 CFQ_CFQQ_FNS(prio_changed);
 441 CFQ_CFQQ_FNS(slice_new);
 442 CFQ_CFQQ_FNS(sync);
 443 CFQ_CFQQ_FNS(coop);
 444 CFQ_CFQQ_FNS(split_coop);
 445 CFQ_CFQQ_FNS(deep);
 446 CFQ_CFQQ_FNS(wait_busy);
 447 #undef CFQ_CFQQ_FNS
 448
 449 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 450 {
 451         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 452 }
 453
 454 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 455 {
 456         return pd_to_blkg(&cfqg->pd);
 457 }
 458
 459 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 460
 461 /* cfqg stats flags */
 462 enum cfqg_stats_flags {
 463         CFQG_stats_waiting = 0,
 464         CFQG_stats_idling,
 465         CFQG_stats_empty,
 466 };
 467
 468 #define CFQG_FLAG_FNS(name)                                             \
 469 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 470 {                                                                       \
 471         stats->flags |= (1 << CFQG_stats_##name);                       \
 472 }                                                                       \
 473 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 474 {                                                                       \
 475         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 476 }                                                                       \
 477 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 478 {                                                                       \
 479         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 480 }                                                                       \
 481
 482 CFQG_FLAG_FNS(waiting)
 483 CFQG_FLAG_FNS(idling)
 484 CFQG_FLAG_FNS(empty)
 485 #undef CFQG_FLAG_FNS
 486
 487 /* This should be called with the queue_lock held. */
 488 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 489 {
 490         unsigned long long now;
 491
 492         if (!cfqg_stats_waiting(stats))
 493                 return;
 494
 495         now = sched_clock();
 496         if (time_after64(now, stats->start_group_wait_time))
 497                 blkg_stat_add(&stats->group_wait_time,
 498                               now - stats->start_group_wait_time);
 499         cfqg_stats_clear_waiting(stats);
 500 }
 501
 502 /* This should be called with the queue_lock held. */
 503 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 504                                                  struct cfq_group *curr_cfqg)
 505 {
 506         struct cfqg_stats *stats = &cfqg->stats;
 507
 508         if (cfqg_stats_waiting(stats))
 509                 return;
 510         if (cfqg == curr_cfqg)
 511                 return;
 512         stats->start_group_wait_time = sched_clock();
 513         cfqg_stats_mark_waiting(stats);
 514 }
 515
 516 /* This should be called with the queue_lock held. */
 517 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 518 {
 519         unsigned long long now;
 520
 521         if (!cfqg_stats_empty(stats))
 522                 return;
 523
 524         now = sched_clock();
 525         if (time_after64(now, stats->start_empty_time))
 526                 blkg_stat_add(&stats->empty_time,
 527                               now - stats->start_empty_time);
 528         cfqg_stats_clear_empty(stats);
 529 }
 530
 531 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 532 {
 533         blkg_stat_add(&cfqg->stats.dequeue, 1);
 534 }
 535
 536 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 537 {
 538         struct cfqg_stats *stats = &cfqg->stats;
 539
 540         if (blkg_rwstat_total(&stats->queued))
 541                 return;
 542
 543         /*
 544          * group is already marked empty. This can happen if cfqq got new
 545          * request in parent group and moved to this group while being added
 546          * to service tree. Just ignore the event and move on.
 547          */
 548         if (cfqg_stats_empty(stats))
 549                 return;
 550
 551         stats->start_empty_time = sched_clock();
 552         cfqg_stats_mark_empty(stats);
 553 }
 554
 555 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 556 {
 557         struct cfqg_stats *stats = &cfqg->stats;
 558
 559         if (cfqg_stats_idling(stats)) {
 560                 unsigned long long now = sched_clock();
 561
 562                 if (time_after64(now, stats->start_idle_time))
 563                         blkg_stat_add(&stats->idle_time,
 564                                       now - stats->start_idle_time);
 565                 cfqg_stats_clear_idling(stats);
 566         }
 567 }
 568
 569 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 570 {
 571         struct cfqg_stats *stats = &cfqg->stats;
 572
 573         BUG_ON(cfqg_stats_idling(stats));
 574
 575         stats->start_idle_time = sched_clock();
 576         cfqg_stats_mark_idling(stats);
 577 }
 578
 579 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 580 {
 581         struct cfqg_stats *stats = &cfqg->stats;
 582
 583         blkg_stat_add(&stats->avg_queue_size_sum,
 584                       blkg_rwstat_total(&stats->queued));
 585         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 586         cfqg_stats_update_group_wait_time(stats);
 587 }
 588
 589 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 590
 591 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 592 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 593 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 594 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 595 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 596 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 597 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 598
 599 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 600
 601 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 602
 603 static struct blkcg_policy blkcg_policy_cfq;
 604
 605 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 606 {
 607         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 608 }
 609
 610 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 611 {
 612         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 613
 614         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 615 }
 616
 617 static inline void cfqg_get(struct cfq_group *cfqg)
 618 {
 619         return blkg_get(cfqg_to_blkg(cfqg));
 620 }
 621
 622 static inline void cfqg_put(struct cfq_group *cfqg)
 623 {
 624         return blkg_put(cfqg_to_blkg(cfqg));
 625 }
 626
 627 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 628         char __pbuf[128];                                               \
 629                                                                         \
 630         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 631         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 632                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 633                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 634                           __pbuf, ##args);                              \
 635 } while (0)
 636
 637 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 638         char __pbuf[128];                                               \
 639                                                                         \
 640         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 641         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 642 } while (0)
 643
 644 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 645                                             struct cfq_group *curr_cfqg, int rw)
 646 {
 647         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 648         cfqg_stats_end_empty_time(&cfqg->stats);
 649         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 650 }
 651
 652 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 653                         unsigned long time, unsigned long unaccounted_time)
 654 {
 655         blkg_stat_add(&cfqg->stats.time, time);
 656 #ifdef CONFIG_DEBUG_BLK_CGROUP
 657         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 658 #endif
 659 }
 660
 661 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 662 {
 663         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 664 }
 665
 666 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 667 {
 668         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 669 }
 670
 671 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 672                                               uint64_t bytes, int rw)
 673 {
 674         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
 675         blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
 676         blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
 677 }
 678
 679 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 680                         uint64_t start_time, uint64_t io_start_time, int rw)
 681 {
 682         struct cfqg_stats *stats = &cfqg->stats;
 683         unsigned long long now = sched_clock();
 684
 685         if (time_after64(now, io_start_time))
 686                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 687         if (time_after64(io_start_time, start_time))
 688                 blkg_rwstat_add(&stats->wait_time, rw,
 689                                 io_start_time - start_time);
 690 }
 691
 692 /* @stats = 0 */
 693 static void cfqg_stats_reset(struct cfqg_stats *stats)
 694 {
 695         /* queued stats shouldn't be cleared */
 696         blkg_rwstat_reset(&stats->service_bytes);
 697         blkg_rwstat_reset(&stats->serviced);
 698         blkg_rwstat_reset(&stats->merged);
 699         blkg_rwstat_reset(&stats->service_time);
 700         blkg_rwstat_reset(&stats->wait_time);
 701         blkg_stat_reset(&stats->time);
 702 #ifdef CONFIG_DEBUG_BLK_CGROUP
 703         blkg_stat_reset(&stats->unaccounted_time);
 704         blkg_stat_reset(&stats->avg_queue_size_sum);
 705         blkg_stat_reset(&stats->avg_queue_size_samples);
 706         blkg_stat_reset(&stats->dequeue);
 707         blkg_stat_reset(&stats->group_wait_time);
 708         blkg_stat_reset(&stats->idle_time);
 709         blkg_stat_reset(&stats->empty_time);
 710 #endif
 711 }
 712
 713 /* @to += @from */
 714 static void cfqg_stats_merge(struct cfqg_stats *to, struct cfqg_stats *from)
 715 {
 716         /* queued stats shouldn't be cleared */
 717         blkg_rwstat_merge(&to->service_bytes, &from->service_bytes);
 718         blkg_rwstat_merge(&to->serviced, &from->serviced);
 719         blkg_rwstat_merge(&to->merged, &from->merged);
 720         blkg_rwstat_merge(&to->service_time, &from->service_time);
 721         blkg_rwstat_merge(&to->wait_time, &from->wait_time);
 722         blkg_stat_merge(&from->time, &from->time);
 723 #ifdef CONFIG_DEBUG_BLK_CGROUP
 724         blkg_stat_merge(&to->unaccounted_time, &from->unaccounted_time);
 725         blkg_stat_merge(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 726         blkg_stat_merge(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 727         blkg_stat_merge(&to->dequeue, &from->dequeue);
 728         blkg_stat_merge(&to->group_wait_time, &from->group_wait_time);
 729         blkg_stat_merge(&to->idle_time, &from->idle_time);
 730         blkg_stat_merge(&to->empty_time, &from->empty_time);
 731 #endif
 732 }
 733
 734 /*
 735  * Transfer @cfqg's stats to its parent's dead_stats so that the ancestors'
 736  * recursive stats can still account for the amount used by this cfqg after
 737  * it's gone.
 738  */
 739 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 740 {
 741         struct cfq_group *parent = cfqg_parent(cfqg);
 742
 743         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 744
 745         if (unlikely(!parent))
 746                 return;
 747
 748         cfqg_stats_merge(&parent->dead_stats, &cfqg->stats);
 749         cfqg_stats_merge(&parent->dead_stats, &cfqg->dead_stats);
 750         cfqg_stats_reset(&cfqg->stats);
 751         cfqg_stats_reset(&cfqg->dead_stats);
 752 }
 753
 754 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 755
 756 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 757 static inline void cfqg_get(struct cfq_group *cfqg) { }
 758 static inline void cfqg_put(struct cfq_group *cfqg) { }
 759
 760 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 761         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 762                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 763                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 764                                 ##args)
 765 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 766
 767 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 768                         struct cfq_group *curr_cfqg, int rw) { }
 769 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 770                         unsigned long time, unsigned long unaccounted_time) { }
 771 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 772 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 773 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 774                                               uint64_t bytes, int rw) { }
 775 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 776                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 777
 778 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 779
 780 #define cfq_log(cfqd, fmt, args...)     \
 781         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 782
 783 /* Traverses through cfq group service trees */
 784 #define for_each_cfqg_st(cfqg, i, j, st) \
 785         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 786                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 787                         : &cfqg->service_tree_idle; \
 788                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 789                         (i == IDLE_WORKLOAD && j == 0); \
 790                         j++, st = i < IDLE_WORKLOAD ? \
 791                         &cfqg->service_trees[i][j]: NULL) \
 792
 793 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 794         struct cfq_ttime *ttime, bool group_idle)
 795 {
 796         unsigned long slice;
 797         if (!sample_valid(ttime->ttime_samples))
 798                 return false;
 799         if (group_idle)
 800                 slice = cfqd->cfq_group_idle;
 801         else
 802                 slice = cfqd->cfq_slice_idle;
 803         return ttime->ttime_mean > slice;
 804 }
 805
 806 static inline bool iops_mode(struct cfq_data *cfqd)
 807 {
 808         /*
 809          * If we are not idling on queues and it is a NCQ drive, parallel
 810          * execution of requests is on and measuring time is not possible
 811          * in most of the cases until and unless we drive shallower queue
 812          * depths and that becomes a performance bottleneck. In such cases
 813          * switch to start providing fairness in terms of number of IOs.
 814          */
 815         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 816                 return true;
 817         else
 818                 return false;
 819 }
 820
 821 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 822 {
 823         if (cfq_class_idle(cfqq))
 824                 return IDLE_WORKLOAD;
 825         if (cfq_class_rt(cfqq))
 826                 return RT_WORKLOAD;
 827         return BE_WORKLOAD;
 828 }
 829
 830
 831 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 832 {
 833         if (!cfq_cfqq_sync(cfqq))
 834                 return ASYNC_WORKLOAD;
 835         if (!cfq_cfqq_idle_window(cfqq))
 836                 return SYNC_NOIDLE_WORKLOAD;
 837         return SYNC_WORKLOAD;
 838 }
 839
 840 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 841                                         struct cfq_data *cfqd,
 842                                         struct cfq_group *cfqg)
 843 {
 844         if (wl_class == IDLE_WORKLOAD)
 845                 return cfqg->service_tree_idle.count;
 846
 847         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 848                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 849                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 850 }
 851
 852 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 853                                         struct cfq_group *cfqg)
 854 {
 855         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 856                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 857 }
 858
 859 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 860 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 861                                        struct cfq_io_cq *cic, struct bio *bio,
 862                                        gfp_t gfp_mask);
 863
 864 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 865 {
 866         /* cic->icq is the first member, %NULL will convert to %NULL */
 867         return container_of(icq, struct cfq_io_cq, icq);
 868 }
 869
 870 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 871                                                struct io_context *ioc)
 872 {
 873         if (ioc)
 874                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 875         return NULL;
 876 }
 877
 878 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 879 {
 880         return cic->cfqq[is_sync];
 881 }
 882
 883 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 884                                 bool is_sync)
 885 {
 886         cic->cfqq[is_sync] = cfqq;
 887 }
 888
 889 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 890 {
 891         return cic->icq.q->elevator->elevator_data;
 892 }
 893
 894 /*
 895  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 896  * set (in which case it could also be direct WRITE).
 897  */
 898 static inline bool cfq_bio_sync(struct bio *bio)
 899 {
 900         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 901 }
 902
 903 /*
 904  * scheduler run of queue, if there are requests pending and no one in the
 905  * driver that will restart queueing
 906  */
 907 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 908 {
 909         if (cfqd->busy_queues) {
 910                 cfq_log(cfqd, "schedule dispatch");
 911                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 912         }
 913 }
 914
 915 /*
 916  * Scale schedule slice based on io priority. Use the sync time slice only
 917  * if a queue is marked sync and has sync io queued. A sync queue with async
 918  * io only, should not get full sync slice length.
 919  */
 920 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 921                                  unsigned short prio)
 922 {
 923         const int base_slice = cfqd->cfq_slice[sync];
 924
 925         WARN_ON(prio >= IOPRIO_BE_NR);
 926
 927         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 928 }
 929
 930 static inline int
 931 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 932 {
 933         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 934 }
 935
 936 /**
 937  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 938  * @charge: disk time being charged
 939  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 940  *
 941  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 942  * scaling is inversely proportional.
 943  *
 944  * scaled = charge / vfraction
 945  *
 946  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 947  */
 948 static inline u64 cfqg_scale_charge(unsigned long charge,
 949                                     unsigned int vfraction)
 950 {
 951         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 952
 953         /* charge / vfraction */
 954         c <<= CFQ_SERVICE_SHIFT;
 955         do_div(c, vfraction);
 956         return c;
 957 }
 958
 959 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 960 {
 961         s64 delta = (s64)(vdisktime - min_vdisktime);
 962         if (delta > 0)
 963                 min_vdisktime = vdisktime;
 964
 965         return min_vdisktime;
 966 }
 967
 968 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 969 {
 970         s64 delta = (s64)(vdisktime - min_vdisktime);
 971         if (delta < 0)
 972                 min_vdisktime = vdisktime;
 973
 974         return min_vdisktime;
 975 }
 976
 977 static void update_min_vdisktime(struct cfq_rb_root *st)
 978 {
 979         struct cfq_group *cfqg;
 980
 981         if (st->left) {
 982                 cfqg = rb_entry_cfqg(st->left);
 983                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 984                                                   cfqg->vdisktime);
 985         }
 986 }
 987
 988 /*
 989  * get averaged number of queues of RT/BE priority.
 990  * average is updated, with a formula that gives more weight to higher numbers,
 991  * to quickly follows sudden increases and decrease slowly
 992  */
 993
 994 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
 995                                         struct cfq_group *cfqg, bool rt)
 996 {
 997         unsigned min_q, max_q;
 998         unsigned mult  = cfq_hist_divisor - 1;
 999         unsigned round = cfq_hist_divisor / 2;
1000         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1001
1002         min_q = min(cfqg->busy_queues_avg[rt], busy);
1003         max_q = max(cfqg->busy_queues_avg[rt], busy);
1004         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1005                 cfq_hist_divisor;
1006         return cfqg->busy_queues_avg[rt];
1007 }
1008
1009 static inline unsigned
1010 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1011 {
1012         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1013 }
1014
1015 static inline unsigned
1016 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1017 {
1018         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1019         if (cfqd->cfq_latency) {
1020                 /*
1021                  * interested queues (we consider only the ones with the same
1022                  * priority class in the cfq group)
1023                  */
1024                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1025                                                 cfq_class_rt(cfqq));
1026                 unsigned sync_slice = cfqd->cfq_slice[1];
1027                 unsigned expect_latency = sync_slice * iq;
1028                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1029
1030                 if (expect_latency > group_slice) {
1031                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1032                         /* scale low_slice according to IO priority
1033                          * and sync vs async */
1034                         unsigned low_slice =
1035                                 min(slice, base_low_slice * slice / sync_slice);
1036                         /* the adapted slice value is scaled to fit all iqs
1037                          * into the target latency */
1038                         slice = max(slice * group_slice / expect_latency,
1039                                     low_slice);
1040                 }
1041         }
1042         return slice;
1043 }
1044
1045 static inline void
1046 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1047 {
1048         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1049
1050         cfqq->slice_start = jiffies;
1051         cfqq->slice_end = jiffies + slice;
1052         cfqq->allocated_slice = slice;
1053         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1054 }
1055
1056 /*
1057  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1058  * isn't valid until the first request from the dispatch is activated
1059  * and the slice time set.
1060  */
1061 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1062 {
1063         if (cfq_cfqq_slice_new(cfqq))
1064                 return false;
1065         if (time_before(jiffies, cfqq->slice_end))
1066                 return false;
1067
1068         return true;
1069 }
1070
1071 /*
1072  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1073  * We choose the request that is closest to the head right now. Distance
1074  * behind the head is penalized and only allowed to a certain extent.
1075  */
1076 static struct request *
1077 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1078 {
1079         sector_t s1, s2, d1 = 0, d2 = 0;
1080         unsigned long back_max;
1081 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1082 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1083         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1084
1085         if (rq1 == NULL || rq1 == rq2)
1086                 return rq2;
1087         if (rq2 == NULL)
1088                 return rq1;
1089
1090         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1091                 return rq_is_sync(rq1) ? rq1 : rq2;
1092
1093         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1094                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1095
1096         s1 = blk_rq_pos(rq1);
1097         s2 = blk_rq_pos(rq2);
1098
1099         /*
1100          * by definition, 1KiB is 2 sectors
1101          */
1102         back_max = cfqd->cfq_back_max * 2;
1103
1104         /*
1105          * Strict one way elevator _except_ in the case where we allow
1106          * short backward seeks which are biased as twice the cost of a
1107          * similar forward seek.
1108          */
1109         if (s1 >= last)
1110                 d1 = s1 - last;
1111         else if (s1 + back_max >= last)
1112                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1113         else
1114                 wrap |= CFQ_RQ1_WRAP;
1115
1116         if (s2 >= last)
1117                 d2 = s2 - last;
1118         else if (s2 + back_max >= last)
1119                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1120         else
1121                 wrap |= CFQ_RQ2_WRAP;
1122
1123         /* Found required data */
1124
1125         /*
1126          * By doing switch() on the bit mask "wrap" we avoid having to
1127          * check two variables for all permutations: --> faster!
1128          */
1129         switch (wrap) {
1130         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1131                 if (d1 < d2)
1132                         return rq1;
1133                 else if (d2 < d1)
1134                         return rq2;
1135                 else {
1136                         if (s1 >= s2)
1137                                 return rq1;
1138                         else
1139                                 return rq2;
1140                 }
1141
1142         case CFQ_RQ2_WRAP:
1143                 return rq1;
1144         case CFQ_RQ1_WRAP:
1145                 return rq2;
1146         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1147         default:
1148                 /*
1149                  * Since both rqs are wrapped,
1150                  * start with the one that's further behind head
1151                  * (--> only *one* back seek required),
1152                  * since back seek takes more time than forward.
1153                  */
1154                 if (s1 <= s2)
1155                         return rq1;
1156                 else
1157                         return rq2;
1158         }
1159 }
1160
1161 /*
1162  * The below is leftmost cache rbtree addon
1163  */
1164 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1165 {
1166         /* Service tree is empty */
1167         if (!root->count)
1168                 return NULL;
1169
1170         if (!root->left)
1171                 root->left = rb_first(&root->rb);
1172
1173         if (root->left)
1174                 return rb_entry(root->left, struct cfq_queue, rb_node);
1175
1176         return NULL;
1177 }
1178
1179 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1180 {
1181         if (!root->left)
1182                 root->left = rb_first(&root->rb);
1183
1184         if (root->left)
1185                 return rb_entry_cfqg(root->left);
1186
1187         return NULL;
1188 }
1189
1190 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1191 {
1192         rb_erase(n, root);
1193         RB_CLEAR_NODE(n);
1194 }
1195
1196 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1197 {
1198         if (root->left == n)
1199                 root->left = NULL;
1200         rb_erase_init(n, &root->rb);
1201         --root->count;
1202 }
1203
1204 /*
1205  * would be nice to take fifo expire time into account as well
1206  */
1207 static struct request *
1208 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1209                   struct request *last)
1210 {
1211         struct rb_node *rbnext = rb_next(&last->rb_node);
1212         struct rb_node *rbprev = rb_prev(&last->rb_node);
1213         struct request *next = NULL, *prev = NULL;
1214
1215         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1216
1217         if (rbprev)
1218                 prev = rb_entry_rq(rbprev);
1219
1220         if (rbnext)
1221                 next = rb_entry_rq(rbnext);
1222         else {
1223                 rbnext = rb_first(&cfqq->sort_list);
1224                 if (rbnext && rbnext != &last->rb_node)
1225                         next = rb_entry_rq(rbnext);
1226         }
1227
1228         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1229 }
1230
1231 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1232                                       struct cfq_queue *cfqq)
1233 {
1234         /*
1235          * just an approximation, should be ok.
1236          */
1237         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1238                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1239 }
1240
1241 static inline s64
1242 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1243 {
1244         return cfqg->vdisktime - st->min_vdisktime;
1245 }
1246
1247 static void
1248 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1249 {
1250         struct rb_node **node = &st->rb.rb_node;
1251         struct rb_node *parent = NULL;
1252         struct cfq_group *__cfqg;
1253         s64 key = cfqg_key(st, cfqg);
1254         int left = 1;
1255
1256         while (*node != NULL) {
1257                 parent = *node;
1258                 __cfqg = rb_entry_cfqg(parent);
1259
1260                 if (key < cfqg_key(st, __cfqg))
1261                         node = &parent->rb_left;
1262                 else {
1263                         node = &parent->rb_right;
1264                         left = 0;
1265                 }
1266         }
1267
1268         if (left)
1269                 st->left = &cfqg->rb_node;
1270
1271         rb_link_node(&cfqg->rb_node, parent, node);
1272         rb_insert_color(&cfqg->rb_node, &st->rb);
1273 }
1274
1275 static void
1276 cfq_update_group_weight(struct cfq_group *cfqg)
1277 {
1278         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1279
1280         if (cfqg->new_weight) {
1281                 cfqg->weight = cfqg->new_weight;
1282                 cfqg->new_weight = 0;
1283         }
1284
1285         if (cfqg->new_leaf_weight) {
1286                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1287                 cfqg->new_leaf_weight = 0;
1288         }
1289 }
1290
1291 static void
1292 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1293 {
1294         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1295         struct cfq_group *pos = cfqg;
1296         struct cfq_group *parent;
1297         bool propagate;
1298
1299         /* add to the service tree */
1300         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1301
1302         cfq_update_group_weight(cfqg);
1303         __cfq_group_service_tree_add(st, cfqg);
1304
1305         /*
1306          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1307          * entitled to.  vfraction is calculated by walking the tree
1308          * towards the root calculating the fraction it has at each level.
1309          * The compounded ratio is how much vfraction @cfqg owns.
1310          *
1311          * Start with the proportion tasks in this cfqg has against active
1312          * children cfqgs - its leaf_weight against children_weight.
1313          */
1314         propagate = !pos->nr_active++;
1315         pos->children_weight += pos->leaf_weight;
1316         vfr = vfr * pos->leaf_weight / pos->children_weight;
1317
1318         /*
1319          * Compound ->weight walking up the tree.  Both activation and
1320          * vfraction calculation are done in the same loop.  Propagation
1321          * stops once an already activated node is met.  vfraction
1322          * calculation should always continue to the root.
1323          */
1324         while ((parent = cfqg_parent(pos))) {
1325                 if (propagate) {
1326                         propagate = !parent->nr_active++;
1327                         parent->children_weight += pos->weight;
1328                 }
1329                 vfr = vfr * pos->weight / parent->children_weight;
1330                 pos = parent;
1331         }
1332
1333         cfqg->vfraction = max_t(unsigned, vfr, 1);
1334 }
1335
1336 static void
1337 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1338 {
1339         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1340         struct cfq_group *__cfqg;
1341         struct rb_node *n;
1342
1343         cfqg->nr_cfqq++;
1344         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1345                 return;
1346
1347         /*
1348          * Currently put the group at the end. Later implement something
1349          * so that groups get lesser vtime based on their weights, so that
1350          * if group does not loose all if it was not continuously backlogged.
1351          */
1352         n = rb_last(&st->rb);
1353         if (n) {
1354                 __cfqg = rb_entry_cfqg(n);
1355                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1356         } else
1357                 cfqg->vdisktime = st->min_vdisktime;
1358         cfq_group_service_tree_add(st, cfqg);
1359 }
1360
1361 static void
1362 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1363 {
1364         struct cfq_group *pos = cfqg;
1365         bool propagate;
1366
1367         /*
1368          * Undo activation from cfq_group_service_tree_add().  Deactivate
1369          * @cfqg and propagate deactivation upwards.
1370          */
1371         propagate = !--pos->nr_active;
1372         pos->children_weight -= pos->leaf_weight;
1373
1374         while (propagate) {
1375                 struct cfq_group *parent = cfqg_parent(pos);
1376
1377                 /* @pos has 0 nr_active at this point */
1378                 WARN_ON_ONCE(pos->children_weight);
1379                 pos->vfraction = 0;
1380
1381                 if (!parent)
1382                         break;
1383
1384                 propagate = !--parent->nr_active;
1385                 parent->children_weight -= pos->weight;
1386                 pos = parent;
1387         }
1388
1389         /* remove from the service tree */
1390         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1391                 cfq_rb_erase(&cfqg->rb_node, st);
1392 }
1393
1394 static void
1395 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1396 {
1397         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1398
1399         BUG_ON(cfqg->nr_cfqq < 1);
1400         cfqg->nr_cfqq--;
1401
1402         /* If there are other cfq queues under this group, don't delete it */
1403         if (cfqg->nr_cfqq)
1404                 return;
1405
1406         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1407         cfq_group_service_tree_del(st, cfqg);
1408         cfqg->saved_wl_slice = 0;
1409         cfqg_stats_update_dequeue(cfqg);
1410 }
1411
1412 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1413                                                 unsigned int *unaccounted_time)
1414 {
1415         unsigned int slice_used;
1416
1417         /*
1418          * Queue got expired before even a single request completed or
1419          * got expired immediately after first request completion.
1420          */
1421         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1422                 /*
1423                  * Also charge the seek time incurred to the group, otherwise
1424                  * if there are mutiple queues in the group, each can dispatch
1425                  * a single request on seeky media and cause lots of seek time
1426                  * and group will never know it.
1427                  */
1428                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1429                                         1);
1430         } else {
1431                 slice_used = jiffies - cfqq->slice_start;
1432                 if (slice_used > cfqq->allocated_slice) {
1433                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1434                         slice_used = cfqq->allocated_slice;
1435                 }
1436                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1437                         *unaccounted_time += cfqq->slice_start -
1438                                         cfqq->dispatch_start;
1439         }
1440
1441         return slice_used;
1442 }
1443
1444 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1445                                 struct cfq_queue *cfqq)
1446 {
1447         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1448         unsigned int used_sl, charge, unaccounted_sl = 0;
1449         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1450                         - cfqg->service_tree_idle.count;
1451         unsigned int vfr;
1452
1453         BUG_ON(nr_sync < 0);
1454         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1455
1456         if (iops_mode(cfqd))
1457                 charge = cfqq->slice_dispatch;
1458         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1459                 charge = cfqq->allocated_slice;
1460
1461         /*
1462          * Can't update vdisktime while on service tree and cfqg->vfraction
1463          * is valid only while on it.  Cache vfr, leave the service tree,
1464          * update vdisktime and go back on.  The re-addition to the tree
1465          * will also update the weights as necessary.
1466          */
1467         vfr = cfqg->vfraction;
1468         cfq_group_service_tree_del(st, cfqg);
1469         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1470         cfq_group_service_tree_add(st, cfqg);
1471
1472         /* This group is being expired. Save the context */
1473         if (time_after(cfqd->workload_expires, jiffies)) {
1474                 cfqg->saved_wl_slice = cfqd->workload_expires
1475                                                 - jiffies;
1476                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1477                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1478         } else
1479                 cfqg->saved_wl_slice = 0;
1480
1481         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1482                                         st->min_vdisktime);
1483         cfq_log_cfqq(cfqq->cfqd, cfqq,
1484                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1485                      used_sl, cfqq->slice_dispatch, charge,
1486                      iops_mode(cfqd), cfqq->nr_sectors);
1487         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1488         cfqg_stats_set_start_empty_time(cfqg);
1489 }
1490
1491 /**
1492  * cfq_init_cfqg_base - initialize base part of a cfq_group
1493  * @cfqg: cfq_group to initialize
1494  *
1495  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1496  * is enabled or not.
1497  */
1498 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1499 {
1500         struct cfq_rb_root *st;
1501         int i, j;
1502
1503         for_each_cfqg_st(cfqg, i, j, st)
1504                 *st = CFQ_RB_ROOT;
1505         RB_CLEAR_NODE(&cfqg->rb_node);
1506
1507         cfqg->ttime.last_end_request = jiffies;
1508 }
1509
1510 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1511 static void cfq_pd_init(struct blkcg_gq *blkg)
1512 {
1513         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1514
1515         cfq_init_cfqg_base(cfqg);
1516         cfqg->weight = blkg->blkcg->cfq_weight;
1517         cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
1518 }
1519
1520 static void cfq_pd_offline(struct blkcg_gq *blkg)
1521 {
1522         /*
1523          * @blkg is going offline and will be ignored by
1524          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1525          * that they don't get lost.  If IOs complete after this point, the
1526          * stats for them will be lost.  Oh well...
1527          */
1528         cfqg_stats_xfer_dead(blkg_to_cfqg(blkg));
1529 }
1530
1531 /* offset delta from cfqg->stats to cfqg->dead_stats */
1532 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1533                                         offsetof(struct cfq_group, stats);
1534
1535 /* to be used by recursive prfill, sums live and dead stats recursively */
1536 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1537 {
1538         u64 sum = 0;
1539
1540         sum += blkg_stat_recursive_sum(pd, off);
1541         sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1542         return sum;
1543 }
1544
1545 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1546 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1547                                                        int off)
1548 {
1549         struct blkg_rwstat a, b;
1550
1551         a = blkg_rwstat_recursive_sum(pd, off);
1552         b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1553         blkg_rwstat_merge(&a, &b);
1554         return a;
1555 }
1556
1557 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1558 {
1559         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1560
1561         cfqg_stats_reset(&cfqg->stats);
1562         cfqg_stats_reset(&cfqg->dead_stats);
1563 }
1564
1565 /*
1566  * Search for the cfq group current task belongs to. request_queue lock must
1567  * be held.
1568  */
1569 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1570                                                 struct blkcg *blkcg)
1571 {
1572         struct request_queue *q = cfqd->queue;
1573         struct cfq_group *cfqg = NULL;
1574
1575         /* avoid lookup for the common case where there's no blkcg */
1576         if (blkcg == &blkcg_root) {
1577                 cfqg = cfqd->root_group;
1578         } else {
1579                 struct blkcg_gq *blkg;
1580
1581                 blkg = blkg_lookup_create(blkcg, q);
1582                 if (!IS_ERR(blkg))
1583                         cfqg = blkg_to_cfqg(blkg);
1584         }
1585
1586         return cfqg;
1587 }
1588
1589 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1590 {
1591         /* Currently, all async queues are mapped to root group */
1592         if (!cfq_cfqq_sync(cfqq))
1593                 cfqg = cfqq->cfqd->root_group;
1594
1595         cfqq->cfqg = cfqg;
1596         /* cfqq reference on cfqg */
1597         cfqg_get(cfqg);
1598 }
1599
1600 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1601                                      struct blkg_policy_data *pd, int off)
1602 {
1603         struct cfq_group *cfqg = pd_to_cfqg(pd);
1604
1605         if (!cfqg->dev_weight)
1606                 return 0;
1607         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1608 }
1609
1610 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1611                                     struct seq_file *sf)
1612 {
1613         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1614                           cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1615                           false);
1616         return 0;
1617 }
1618
1619 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1620                                           struct blkg_policy_data *pd, int off)
1621 {
1622         struct cfq_group *cfqg = pd_to_cfqg(pd);
1623
1624         if (!cfqg->dev_leaf_weight)
1625                 return 0;
1626         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1627 }
1628
1629 static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1630                                          struct cftype *cft,
1631                                          struct seq_file *sf)
1632 {
1633         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1634                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1635                           false);
1636         return 0;
1637 }
1638
1639 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1640                             struct seq_file *sf)
1641 {
1642         seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1643         return 0;
1644 }
1645
1646 static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1647                                  struct seq_file *sf)
1648 {
1649         seq_printf(sf, "%u\n",
1650                    cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1651         return 0;
1652 }
1653
1654 static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1655                                     const char *buf, bool is_leaf_weight)
1656 {
1657         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1658         struct blkg_conf_ctx ctx;
1659         struct cfq_group *cfqg;
1660         int ret;
1661
1662         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1663         if (ret)
1664                 return ret;
1665
1666         ret = -EINVAL;
1667         cfqg = blkg_to_cfqg(ctx.blkg);
1668         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1669                 if (!is_leaf_weight) {
1670                         cfqg->dev_weight = ctx.v;
1671                         cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1672                 } else {
1673                         cfqg->dev_leaf_weight = ctx.v;
1674                         cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1675                 }
1676                 ret = 0;
1677         }
1678
1679         blkg_conf_finish(&ctx);
1680         return ret;
1681 }
1682
1683 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1684                                   const char *buf)
1685 {
1686         return __cfqg_set_weight_device(cgrp, cft, buf, false);
1687 }
1688
1689 static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1690                                        const char *buf)
1691 {
1692         return __cfqg_set_weight_device(cgrp, cft, buf, true);
1693 }
1694
1695 static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1696                             bool is_leaf_weight)
1697 {
1698         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1699         struct blkcg_gq *blkg;
1700
1701         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1702                 return -EINVAL;
1703
1704         spin_lock_irq(&blkcg->lock);
1705
1706         if (!is_leaf_weight)
1707                 blkcg->cfq_weight = val;
1708         else
1709                 blkcg->cfq_leaf_weight = val;
1710
1711         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1712                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1713
1714                 if (!cfqg)
1715                         continue;
1716
1717                 if (!is_leaf_weight) {
1718                         if (!cfqg->dev_weight)
1719                                 cfqg->new_weight = blkcg->cfq_weight;
1720                 } else {
1721                         if (!cfqg->dev_leaf_weight)
1722                                 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1723                 }
1724         }
1725
1726         spin_unlock_irq(&blkcg->lock);
1727         return 0;
1728 }
1729
1730 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1731 {
1732         return __cfq_set_weight(cgrp, cft, val, false);
1733 }
1734
1735 static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1736 {
1737         return __cfq_set_weight(cgrp, cft, val, true);
1738 }
1739
1740 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1741                            struct seq_file *sf)
1742 {
1743         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1744
1745         blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1746                           cft->private, false);
1747         return 0;
1748 }
1749
1750 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1751                              struct seq_file *sf)
1752 {
1753         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1754
1755         blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1756                           cft->private, true);
1757         return 0;
1758 }
1759
1760 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1761                                       struct blkg_policy_data *pd, int off)
1762 {
1763         u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1764
1765         return __blkg_prfill_u64(sf, pd, sum);
1766 }
1767
1768 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1769                                         struct blkg_policy_data *pd, int off)
1770 {
1771         struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1772
1773         return __blkg_prfill_rwstat(sf, pd, &sum);
1774 }
1775
1776 static int cfqg_print_stat_recursive(struct cgroup *cgrp, struct cftype *cft,
1777                                      struct seq_file *sf)
1778 {
1779         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1780
1781         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_stat_recursive,
1782                           &blkcg_policy_cfq, cft->private, false);
1783         return 0;
1784 }
1785
1786 static int cfqg_print_rwstat_recursive(struct cgroup *cgrp, struct cftype *cft,
1787                                        struct seq_file *sf)
1788 {
1789         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1790
1791         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_rwstat_recursive,
1792                           &blkcg_policy_cfq, cft->private, true);
1793         return 0;
1794 }
1795
1796 #ifdef CONFIG_DEBUG_BLK_CGROUP
1797 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1798                                       struct blkg_policy_data *pd, int off)
1799 {
1800         struct cfq_group *cfqg = pd_to_cfqg(pd);
1801         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1802         u64 v = 0;
1803
1804         if (samples) {
1805                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1806                 do_div(v, samples);
1807         }
1808         __blkg_prfill_u64(sf, pd, v);
1809         return 0;
1810 }
1811
1812 /* print avg_queue_size */
1813 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1814                                      struct seq_file *sf)
1815 {
1816         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1817
1818         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1819                           &blkcg_policy_cfq, 0, false);
1820         return 0;
1821 }
1822 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1823
1824 static struct cftype cfq_blkcg_files[] = {
1825         /* on root, weight is mapped to leaf_weight */
1826         {
1827                 .name = "weight_device",
1828                 .flags = CFTYPE_ONLY_ON_ROOT,
1829                 .read_seq_string = cfqg_print_leaf_weight_device,
1830                 .write_string = cfqg_set_leaf_weight_device,
1831                 .max_write_len = 256,
1832         },
1833         {
1834                 .name = "weight",
1835                 .flags = CFTYPE_ONLY_ON_ROOT,
1836                 .read_seq_string = cfq_print_leaf_weight,
1837                 .write_u64 = cfq_set_leaf_weight,
1838         },
1839
1840         /* no such mapping necessary for !roots */
1841         {
1842                 .name = "weight_device",
1843                 .flags = CFTYPE_NOT_ON_ROOT,
1844                 .read_seq_string = cfqg_print_weight_device,
1845                 .write_string = cfqg_set_weight_device,
1846                 .max_write_len = 256,
1847         },
1848         {
1849                 .name = "weight",
1850                 .flags = CFTYPE_NOT_ON_ROOT,
1851                 .read_seq_string = cfq_print_weight,
1852                 .write_u64 = cfq_set_weight,
1853         },
1854
1855         {
1856                 .name = "leaf_weight_device",
1857                 .read_seq_string = cfqg_print_leaf_weight_device,
1858                 .write_string = cfqg_set_leaf_weight_device,
1859                 .max_write_len = 256,
1860         },
1861         {
1862                 .name = "leaf_weight",
1863                 .read_seq_string = cfq_print_leaf_weight,
1864                 .write_u64 = cfq_set_leaf_weight,
1865         },
1866
1867         /* statistics, covers only the tasks in the cfqg */
1868         {
1869                 .name = "time",
1870                 .private = offsetof(struct cfq_group, stats.time),
1871                 .read_seq_string = cfqg_print_stat,
1872         },
1873         {
1874                 .name = "sectors",
1875                 .private = offsetof(struct cfq_group, stats.sectors),
1876                 .read_seq_string = cfqg_print_stat,
1877         },
1878         {
1879                 .name = "io_service_bytes",
1880                 .private = offsetof(struct cfq_group, stats.service_bytes),
1881                 .read_seq_string = cfqg_print_rwstat,
1882         },
1883         {
1884                 .name = "io_serviced",
1885                 .private = offsetof(struct cfq_group, stats.serviced),
1886                 .read_seq_string = cfqg_print_rwstat,
1887         },
1888         {
1889                 .name = "io_service_time",
1890                 .private = offsetof(struct cfq_group, stats.service_time),
1891                 .read_seq_string = cfqg_print_rwstat,
1892         },
1893         {
1894                 .name = "io_wait_time",
1895                 .private = offsetof(struct cfq_group, stats.wait_time),
1896                 .read_seq_string = cfqg_print_rwstat,
1897         },
1898         {
1899                 .name = "io_merged",
1900                 .private = offsetof(struct cfq_group, stats.merged),
1901                 .read_seq_string = cfqg_print_rwstat,
1902         },
1903         {
1904                 .name = "io_queued",
1905                 .private = offsetof(struct cfq_group, stats.queued),
1906                 .read_seq_string = cfqg_print_rwstat,
1907         },
1908
1909         /* the same statictics which cover the cfqg and its descendants */
1910         {
1911                 .name = "time_recursive",
1912                 .private = offsetof(struct cfq_group, stats.time),
1913                 .read_seq_string = cfqg_print_stat_recursive,
1914         },
1915         {
1916                 .name = "sectors_recursive",
1917                 .private = offsetof(struct cfq_group, stats.sectors),
1918                 .read_seq_string = cfqg_print_stat_recursive,
1919         },
1920         {
1921                 .name = "io_service_bytes_recursive",
1922                 .private = offsetof(struct cfq_group, stats.service_bytes),
1923                 .read_seq_string = cfqg_print_rwstat_recursive,
1924         },
1925         {
1926                 .name = "io_serviced_recursive",
1927                 .private = offsetof(struct cfq_group, stats.serviced),
1928                 .read_seq_string = cfqg_print_rwstat_recursive,
1929         },
1930         {
1931                 .name = "io_service_time_recursive",
1932                 .private = offsetof(struct cfq_group, stats.service_time),
1933                 .read_seq_string = cfqg_print_rwstat_recursive,
1934         },
1935         {
1936                 .name = "io_wait_time_recursive",
1937                 .private = offsetof(struct cfq_group, stats.wait_time),
1938                 .read_seq_string = cfqg_print_rwstat_recursive,
1939         },
1940         {
1941                 .name = "io_merged_recursive",
1942                 .private = offsetof(struct cfq_group, stats.merged),
1943                 .read_seq_string = cfqg_print_rwstat_recursive,
1944         },
1945         {
1946                 .name = "io_queued_recursive",
1947                 .private = offsetof(struct cfq_group, stats.queued),
1948                 .read_seq_string = cfqg_print_rwstat_recursive,
1949         },
1950 #ifdef CONFIG_DEBUG_BLK_CGROUP
1951         {
1952                 .name = "avg_queue_size",
1953                 .read_seq_string = cfqg_print_avg_queue_size,
1954         },
1955         {
1956                 .name = "group_wait_time",
1957                 .private = offsetof(struct cfq_group, stats.group_wait_time),
1958                 .read_seq_string = cfqg_print_stat,
1959         },
1960         {
1961                 .name = "idle_time",
1962                 .private = offsetof(struct cfq_group, stats.idle_time),
1963                 .read_seq_string = cfqg_print_stat,
1964         },
1965         {
1966                 .name = "empty_time",
1967                 .private = offsetof(struct cfq_group, stats.empty_time),
1968                 .read_seq_string = cfqg_print_stat,
1969         },
1970         {
1971                 .name = "dequeue",
1972                 .private = offsetof(struct cfq_group, stats.dequeue),
1973                 .read_seq_string = cfqg_print_stat,
1974         },
1975         {
1976                 .name = "unaccounted_time",
1977                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1978                 .read_seq_string = cfqg_print_stat,
1979         },
1980 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1981         { }     /* terminate */
1982 };
1983 #else /* GROUP_IOSCHED */
1984 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1985                                                 struct blkcg *blkcg)
1986 {
1987         return cfqd->root_group;
1988 }
1989
1990 static inline void
1991 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1992         cfqq->cfqg = cfqg;
1993 }
1994
1995 #endif /* GROUP_IOSCHED */
1996
1997 /*
1998  * The cfqd->service_trees holds all pending cfq_queue's that have
1999  * requests waiting to be processed. It is sorted in the order that
2000  * we will service the queues.
2001  */
2002 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2003                                  bool add_front)
2004 {
2005         struct rb_node **p, *parent;
2006         struct cfq_queue *__cfqq;
2007         unsigned long rb_key;
2008         struct cfq_rb_root *st;
2009         int left;
2010         int new_cfqq = 1;
2011
2012         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2013         if (cfq_class_idle(cfqq)) {
2014                 rb_key = CFQ_IDLE_DELAY;
2015                 parent = rb_last(&st->rb);
2016                 if (parent && parent != &cfqq->rb_node) {
2017                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2018                         rb_key += __cfqq->rb_key;
2019                 } else
2020                         rb_key += jiffies;
2021         } else if (!add_front) {
2022                 /*
2023                  * Get our rb key offset. Subtract any residual slice
2024                  * value carried from last service. A negative resid
2025                  * count indicates slice overrun, and this should position
2026                  * the next service time further away in the tree.
2027                  */
2028                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2029                 rb_key -= cfqq->slice_resid;
2030                 cfqq->slice_resid = 0;
2031         } else {
2032                 rb_key = -HZ;
2033                 __cfqq = cfq_rb_first(st);
2034                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2035         }
2036
2037         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2038                 new_cfqq = 0;
2039                 /*
2040                  * same position, nothing more to do
2041                  */
2042                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2043                         return;
2044
2045                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2046                 cfqq->service_tree = NULL;
2047         }
2048
2049         left = 1;
2050         parent = NULL;
2051         cfqq->service_tree = st;
2052         p = &st->rb.rb_node;
2053         while (*p) {
2054                 parent = *p;
2055                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2056
2057                 /*
2058                  * sort by key, that represents service time.
2059                  */
2060                 if (time_before(rb_key, __cfqq->rb_key))
2061                         p = &parent->rb_left;
2062                 else {
2063                         p = &parent->rb_right;
2064                         left = 0;
2065                 }
2066         }
2067
2068         if (left)
2069                 st->left = &cfqq->rb_node;
2070
2071         cfqq->rb_key = rb_key;
2072         rb_link_node(&cfqq->rb_node, parent, p);
2073         rb_insert_color(&cfqq->rb_node, &st->rb);
2074         st->count++;
2075         if (add_front || !new_cfqq)
2076                 return;
2077         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2078 }
2079
2080 static struct cfq_queue *
2081 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2082                      sector_t sector, struct rb_node **ret_parent,
2083                      struct rb_node ***rb_link)
2084 {
2085         struct rb_node **p, *parent;
2086         struct cfq_queue *cfqq = NULL;
2087
2088         parent = NULL;
2089         p = &root->rb_node;
2090         while (*p) {
2091                 struct rb_node **n;
2092
2093                 parent = *p;
2094                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2095
2096                 /*
2097                  * Sort strictly based on sector.  Smallest to the left,
2098                  * largest to the right.
2099                  */
2100                 if (sector > blk_rq_pos(cfqq->next_rq))
2101                         n = &(*p)->rb_right;
2102                 else if (sector < blk_rq_pos(cfqq->next_rq))
2103                         n = &(*p)->rb_left;
2104                 else
2105                         break;
2106                 p = n;
2107                 cfqq = NULL;
2108         }
2109
2110         *ret_parent = parent;
2111         if (rb_link)
2112                 *rb_link = p;
2113         return cfqq;
2114 }
2115
2116 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2117 {
2118         struct rb_node **p, *parent;
2119         struct cfq_queue *__cfqq;
2120
2121         if (cfqq->p_root) {
2122                 rb_erase(&cfqq->p_node, cfqq->p_root);
2123                 cfqq->p_root = NULL;
2124         }
2125
2126         if (cfq_class_idle(cfqq))
2127                 return;
2128         if (!cfqq->next_rq)
2129                 return;
2130
2131         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2132         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2133                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2134         if (!__cfqq) {
2135                 rb_link_node(&cfqq->p_node, parent, p);
2136                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2137         } else
2138                 cfqq->p_root = NULL;
2139 }
2140
2141 /*
2142  * Update cfqq's position in the service tree.
2143  */
2144 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2145 {
2146         /*
2147          * Resorting requires the cfqq to be on the RR list already.
2148          */
2149         if (cfq_cfqq_on_rr(cfqq)) {
2150                 cfq_service_tree_add(cfqd, cfqq, 0);
2151                 cfq_prio_tree_add(cfqd, cfqq);
2152         }
2153 }
2154
2155 /*
2156  * add to busy list of queues for service, trying to be fair in ordering
2157  * the pending list according to last request service
2158  */
2159 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2160 {
2161         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2162         BUG_ON(cfq_cfqq_on_rr(cfqq));
2163         cfq_mark_cfqq_on_rr(cfqq);
2164         cfqd->busy_queues++;
2165         if (cfq_cfqq_sync(cfqq))
2166                 cfqd->busy_sync_queues++;
2167
2168         cfq_resort_rr_list(cfqd, cfqq);
2169 }
2170
2171 /*
2172  * Called when the cfqq no longer has requests pending, remove it from
2173  * the service tree.
2174  */
2175 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2176 {
2177         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2178         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2179         cfq_clear_cfqq_on_rr(cfqq);
2180
2181         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2182                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2183                 cfqq->service_tree = NULL;
2184         }
2185         if (cfqq->p_root) {
2186                 rb_erase(&cfqq->p_node, cfqq->p_root);
2187                 cfqq->p_root = NULL;
2188         }
2189
2190         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2191         BUG_ON(!cfqd->busy_queues);
2192         cfqd->busy_queues--;
2193         if (cfq_cfqq_sync(cfqq))
2194                 cfqd->busy_sync_queues--;
2195 }
2196
2197 /*
2198  * rb tree support functions
2199  */
2200 static void cfq_del_rq_rb(struct request *rq)
2201 {
2202         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2203         const int sync = rq_is_sync(rq);
2204
2205         BUG_ON(!cfqq->queued[sync]);
2206         cfqq->queued[sync]--;
2207
2208         elv_rb_del(&cfqq->sort_list, rq);
2209
2210         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2211                 /*
2212                  * Queue will be deleted from service tree when we actually
2213                  * expire it later. Right now just remove it from prio tree
2214                  * as it is empty.
2215                  */
2216                 if (cfqq->p_root) {
2217                         rb_erase(&cfqq->p_node, cfqq->p_root);
2218                         cfqq->p_root = NULL;
2219                 }
2220         }
2221 }
2222
2223 static void cfq_add_rq_rb(struct request *rq)
2224 {
2225         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2226         struct cfq_data *cfqd = cfqq->cfqd;
2227         struct request *prev;
2228
2229         cfqq->queued[rq_is_sync(rq)]++;
2230
2231         elv_rb_add(&cfqq->sort_list, rq);
2232
2233         if (!cfq_cfqq_on_rr(cfqq))
2234                 cfq_add_cfqq_rr(cfqd, cfqq);
2235
2236         /*
2237          * check if this request is a better next-serve candidate
2238          */
2239         prev = cfqq->next_rq;
2240         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2241
2242         /*
2243          * adjust priority tree position, if ->next_rq changes
2244          */
2245         if (prev != cfqq->next_rq)
2246                 cfq_prio_tree_add(cfqd, cfqq);
2247
2248         BUG_ON(!cfqq->next_rq);
2249 }
2250
2251 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2252 {
2253         elv_rb_del(&cfqq->sort_list, rq);
2254         cfqq->queued[rq_is_sync(rq)]--;
2255         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2256         cfq_add_rq_rb(rq);
2257         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2258                                  rq->cmd_flags);
2259 }
2260
2261 static struct request *
2262 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2263 {
2264         struct task_struct *tsk = current;
2265         struct cfq_io_cq *cic;
2266         struct cfq_queue *cfqq;
2267
2268         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2269         if (!cic)
2270                 return NULL;
2271
2272         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2273         if (cfqq) {
2274                 sector_t sector = bio->bi_sector + bio_sectors(bio);
2275
2276                 return elv_rb_find(&cfqq->sort_list, sector);
2277         }
2278
2279         return NULL;
2280 }
2281
2282 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2283 {
2284         struct cfq_data *cfqd = q->elevator->elevator_data;
2285
2286         cfqd->rq_in_driver++;
2287         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2288                                                 cfqd->rq_in_driver);
2289
2290         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2291 }
2292
2293 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2294 {
2295         struct cfq_data *cfqd = q->elevator->elevator_data;
2296
2297         WARN_ON(!cfqd->rq_in_driver);
2298         cfqd->rq_in_driver--;
2299         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2300                                                 cfqd->rq_in_driver);
2301 }
2302
2303 static void cfq_remove_request(struct request *rq)
2304 {
2305         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2306
2307         if (cfqq->next_rq == rq)
2308                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2309
2310         list_del_init(&rq->queuelist);
2311         cfq_del_rq_rb(rq);
2312
2313         cfqq->cfqd->rq_queued--;
2314         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2315         if (rq->cmd_flags & REQ_PRIO) {
2316                 WARN_ON(!cfqq->prio_pending);
2317                 cfqq->prio_pending--;
2318         }
2319 }
2320
2321 static int cfq_merge(struct request_queue *q, struct request **req,
2322                      struct bio *bio)
2323 {
2324         struct cfq_data *cfqd = q->elevator->elevator_data;
2325         struct request *__rq;
2326
2327         __rq = cfq_find_rq_fmerge(cfqd, bio);
2328         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2329                 *req = __rq;
2330                 return ELEVATOR_FRONT_MERGE;
2331         }
2332
2333         return ELEVATOR_NO_MERGE;
2334 }
2335
2336 static void cfq_merged_request(struct request_queue *q, struct request *req,
2337                                int type)
2338 {
2339         if (type == ELEVATOR_FRONT_MERGE) {
2340                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2341
2342                 cfq_reposition_rq_rb(cfqq, req);
2343         }
2344 }
2345
2346 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2347                                 struct bio *bio)
2348 {
2349         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2350 }
2351
2352 static void
2353 cfq_merged_requests(struct request_queue *q, struct request *rq,
2354                     struct request *next)
2355 {
2356         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2357         struct cfq_data *cfqd = q->elevator->elevator_data;
2358
2359         /*
2360          * reposition in fifo if next is older than rq
2361          */
2362         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2363             time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2364             cfqq == RQ_CFQQ(next)) {
2365                 list_move(&rq->queuelist, &next->queuelist);
2366                 rq_set_fifo_time(rq, rq_fifo_time(next));
2367         }
2368
2369         if (cfqq->next_rq == next)
2370                 cfqq->next_rq = rq;
2371         cfq_remove_request(next);
2372         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2373
2374         cfqq = RQ_CFQQ(next);
2375         /*
2376          * all requests of this queue are merged to other queues, delete it
2377          * from the service tree. If it's the active_queue,
2378          * cfq_dispatch_requests() will choose to expire it or do idle
2379          */
2380         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2381             cfqq != cfqd->active_queue)
2382                 cfq_del_cfqq_rr(cfqd, cfqq);
2383 }
2384
2385 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2386                            struct bio *bio)
2387 {
2388         struct cfq_data *cfqd = q->elevator->elevator_data;
2389         struct cfq_io_cq *cic;
2390         struct cfq_queue *cfqq;
2391
2392         /*
2393          * Disallow merge of a sync bio into an async request.
2394          */
2395         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2396                 return false;
2397
2398         /*
2399          * Lookup the cfqq that this bio will be queued with and allow
2400          * merge only if rq is queued there.
2401          */
2402         cic = cfq_cic_lookup(cfqd, current->io_context);
2403         if (!cic)
2404                 return false;
2405
2406         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2407         return cfqq == RQ_CFQQ(rq);
2408 }
2409
2410 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2411 {
2412         del_timer(&cfqd->idle_slice_timer);
2413         cfqg_stats_update_idle_time(cfqq->cfqg);
2414 }
2415
2416 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2417                                    struct cfq_queue *cfqq)
2418 {
2419         if (cfqq) {
2420                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2421                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2422                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2423                 cfqq->slice_start = 0;
2424                 cfqq->dispatch_start = jiffies;
2425                 cfqq->allocated_slice = 0;
2426                 cfqq->slice_end = 0;
2427                 cfqq->slice_dispatch = 0;
2428                 cfqq->nr_sectors = 0;
2429
2430                 cfq_clear_cfqq_wait_request(cfqq);
2431                 cfq_clear_cfqq_must_dispatch(cfqq);
2432                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2433                 cfq_clear_cfqq_fifo_expire(cfqq);
2434                 cfq_mark_cfqq_slice_new(cfqq);
2435
2436                 cfq_del_timer(cfqd, cfqq);
2437         }
2438
2439         cfqd->active_queue = cfqq;
2440 }
2441
2442 /*
2443  * current cfqq expired its slice (or was too idle), select new one
2444  */
2445 static void
2446 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2447                     bool timed_out)
2448 {
2449         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2450
2451         if (cfq_cfqq_wait_request(cfqq))
2452                 cfq_del_timer(cfqd, cfqq);
2453
2454         cfq_clear_cfqq_wait_request(cfqq);
2455         cfq_clear_cfqq_wait_busy(cfqq);
2456
2457         /*
2458          * If this cfqq is shared between multiple processes, check to
2459          * make sure that those processes are still issuing I/Os within
2460          * the mean seek distance.  If not, it may be time to break the
2461          * queues apart again.
2462          */
2463         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2464                 cfq_mark_cfqq_split_coop(cfqq);
2465
2466         /*
2467          * store what was left of this slice, if the queue idled/timed out
2468          */
2469         if (timed_out) {
2470                 if (cfq_cfqq_slice_new(cfqq))
2471                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2472                 else
2473                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2474                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2475         }
2476
2477         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2478
2479         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2480                 cfq_del_cfqq_rr(cfqd, cfqq);
2481
2482         cfq_resort_rr_list(cfqd, cfqq);
2483
2484         if (cfqq == cfqd->active_queue)
2485                 cfqd->active_queue = NULL;
2486
2487         if (cfqd->active_cic) {
2488                 put_io_context(cfqd->active_cic->icq.ioc);
2489                 cfqd->active_cic = NULL;
2490         }
2491 }
2492
2493 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2494 {
2495         struct cfq_queue *cfqq = cfqd->active_queue;
2496
2497         if (cfqq)
2498                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2499 }
2500
2501 /*
2502  * Get next queue for service. Unless we have a queue preemption,
2503  * we'll simply select the first cfqq in the service tree.
2504  */
2505 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2506 {
2507         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2508                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2509
2510         if (!cfqd->rq_queued)
2511                 return NULL;
2512
2513         /* There is nothing to dispatch */
2514         if (!st)
2515                 return NULL;
2516         if (RB_EMPTY_ROOT(&st->rb))
2517                 return NULL;
2518         return cfq_rb_first(st);
2519 }
2520
2521 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2522 {
2523         struct cfq_group *cfqg;
2524         struct cfq_queue *cfqq;
2525         int i, j;
2526         struct cfq_rb_root *st;
2527
2528         if (!cfqd->rq_queued)
2529                 return NULL;
2530
2531         cfqg = cfq_get_next_cfqg(cfqd);
2532         if (!cfqg)
2533                 return NULL;
2534
2535         for_each_cfqg_st(cfqg, i, j, st)
2536                 if ((cfqq = cfq_rb_first(st)) != NULL)
2537                         return cfqq;
2538         return NULL;
2539 }
2540
2541 /*
2542  * Get and set a new active queue for service.
2543  */
2544 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2545                                               struct cfq_queue *cfqq)
2546 {
2547         if (!cfqq)
2548                 cfqq = cfq_get_next_queue(cfqd);
2549
2550         __cfq_set_active_queue(cfqd, cfqq);
2551         return cfqq;
2552 }
2553
2554 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2555                                           struct request *rq)
2556 {
2557         if (blk_rq_pos(rq) >= cfqd->last_position)
2558                 return blk_rq_pos(rq) - cfqd->last_position;
2559         else
2560                 return cfqd->last_position - blk_rq_pos(rq);
2561 }
2562
2563 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2564                                struct request *rq)
2565 {
2566         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2567 }
2568
2569 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2570                                     struct cfq_queue *cur_cfqq)
2571 {
2572         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2573         struct rb_node *parent, *node;
2574         struct cfq_queue *__cfqq;
2575         sector_t sector = cfqd->last_position;
2576
2577         if (RB_EMPTY_ROOT(root))
2578                 return NULL;
2579
2580         /*
2581          * First, if we find a request starting at the end of the last
2582          * request, choose it.
2583          */
2584         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2585         if (__cfqq)
2586                 return __cfqq;
2587
2588         /*
2589          * If the exact sector wasn't found, the parent of the NULL leaf
2590          * will contain the closest sector.
2591          */
2592         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2593         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2594                 return __cfqq;
2595
2596         if (blk_rq_pos(__cfqq->next_rq) < sector)
2597                 node = rb_next(&__cfqq->p_node);
2598         else
2599                 node = rb_prev(&__cfqq->p_node);
2600         if (!node)
2601                 return NULL;
2602
2603         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2604         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2605                 return __cfqq;
2606
2607         return NULL;
2608 }
2609
2610 /*
2611  * cfqd - obvious
2612  * cur_cfqq - passed in so that we don't decide that the current queue is
2613  *            closely cooperating with itself.
2614  *
2615  * So, basically we're assuming that that cur_cfqq has dispatched at least
2616  * one request, and that cfqd->last_position reflects a position on the disk
2617  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2618  * assumption.
2619  */
2620 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2621                                               struct cfq_queue *cur_cfqq)
2622 {
2623         struct cfq_queue *cfqq;
2624
2625         if (cfq_class_idle(cur_cfqq))
2626                 return NULL;
2627         if (!cfq_cfqq_sync(cur_cfqq))
2628                 return NULL;
2629         if (CFQQ_SEEKY(cur_cfqq))
2630                 return NULL;
2631
2632         /*
2633          * Don't search priority tree if it's the only queue in the group.
2634          */
2635         if (cur_cfqq->cfqg->nr_cfqq == 1)
2636                 return NULL;
2637
2638         /*
2639          * We should notice if some of the queues are cooperating, eg
2640          * working closely on the same area of the disk. In that case,
2641          * we can group them together and don't waste time idling.
2642          */
2643         cfqq = cfqq_close(cfqd, cur_cfqq);
2644         if (!cfqq)
2645                 return NULL;
2646
2647         /* If new queue belongs to different cfq_group, don't choose it */
2648         if (cur_cfqq->cfqg != cfqq->cfqg)
2649                 return NULL;
2650
2651         /*
2652          * It only makes sense to merge sync queues.
2653          */
2654         if (!cfq_cfqq_sync(cfqq))
2655                 return NULL;
2656         if (CFQQ_SEEKY(cfqq))
2657                 return NULL;
2658
2659         /*
2660          * Do not merge queues of different priority classes
2661          */
2662         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2663                 return NULL;
2664
2665         return cfqq;
2666 }
2667
2668 /*
2669  * Determine whether we should enforce idle window for this queue.
2670  */
2671
2672 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2673 {
2674         enum wl_class_t wl_class = cfqq_class(cfqq);
2675         struct cfq_rb_root *st = cfqq->service_tree;
2676
2677         BUG_ON(!st);
2678         BUG_ON(!st->count);
2679
2680         if (!cfqd->cfq_slice_idle)
2681                 return false;
2682
2683         /* We never do for idle class queues. */
2684         if (wl_class == IDLE_WORKLOAD)
2685                 return false;
2686
2687         /* We do for queues that were marked with idle window flag. */
2688         if (cfq_cfqq_idle_window(cfqq) &&
2689            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2690                 return true;
2691
2692         /*
2693          * Otherwise, we do only if they are the last ones
2694          * in their service tree.
2695          */
2696         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2697            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2698                 return true;
2699         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2700         return false;
2701 }
2702
2703 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2704 {
2705         struct cfq_queue *cfqq = cfqd->active_queue;
2706         struct cfq_io_cq *cic;
2707         unsigned long sl, group_idle = 0;
2708
2709         /*
2710          * SSD device without seek penalty, disable idling. But only do so
2711          * for devices that support queuing, otherwise we still have a problem
2712          * with sync vs async workloads.
2713          */
2714         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2715                 return;
2716
2717         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2718         WARN_ON(cfq_cfqq_slice_new(cfqq));
2719
2720         /*
2721          * idle is disabled, either manually or by past process history
2722          */
2723         if (!cfq_should_idle(cfqd, cfqq)) {
2724                 /* no queue idling. Check for group idling */
2725                 if (cfqd->cfq_group_idle)
2726                         group_idle = cfqd->cfq_group_idle;
2727                 else
2728                         return;
2729         }
2730
2731         /*
2732          * still active requests from this queue, don't idle
2733          */
2734         if (cfqq->dispatched)
2735                 return;
2736
2737         /*
2738          * task has exited, don't wait
2739          */
2740         cic = cfqd->active_cic;
2741         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2742                 return;
2743
2744         /*
2745          * If our average think time is larger than the remaining time
2746          * slice, then don't idle. This avoids overrunning the allotted
2747          * time slice.
2748          */
2749         if (sample_valid(cic->ttime.ttime_samples) &&
2750             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2751                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2752                              cic->ttime.ttime_mean);
2753                 return;
2754         }
2755
2756         /* There are other queues in the group, don't do group idle */
2757         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2758                 return;
2759
2760         cfq_mark_cfqq_wait_request(cfqq);
2761
2762         if (group_idle)
2763                 sl = cfqd->cfq_group_idle;
2764         else
2765                 sl = cfqd->cfq_slice_idle;
2766
2767         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2768         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2769         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2770                         group_idle ? 1 : 0);
2771 }
2772
2773 /*
2774  * Move request from internal lists to the request queue dispatch list.
2775  */
2776 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2777 {
2778         struct cfq_data *cfqd = q->elevator->elevator_data;
2779         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2780
2781         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2782
2783         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2784         cfq_remove_request(rq);
2785         cfqq->dispatched++;
2786         (RQ_CFQG(rq))->dispatched++;
2787         elv_dispatch_sort(q, rq);
2788
2789         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2790         cfqq->nr_sectors += blk_rq_sectors(rq);
2791         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2792 }
2793
2794 /*
2795  * return expired entry, or NULL to just start from scratch in rbtree
2796  */
2797 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2798 {
2799         struct request *rq = NULL;
2800
2801         if (cfq_cfqq_fifo_expire(cfqq))
2802                 return NULL;
2803
2804         cfq_mark_cfqq_fifo_expire(cfqq);
2805
2806         if (list_empty(&cfqq->fifo))
2807                 return NULL;
2808
2809         rq = rq_entry_fifo(cfqq->fifo.next);
2810         if (time_before(jiffies, rq_fifo_time(rq)))
2811                 rq = NULL;
2812
2813         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2814         return rq;
2815 }
2816
2817 static inline int
2818 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2819 {
2820         const int base_rq = cfqd->cfq_slice_async_rq;
2821
2822         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2823
2824         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2825 }
2826
2827 /*
2828  * Must be called with the queue_lock held.
2829  */
2830 static int cfqq_process_refs(struct cfq_queue *cfqq)
2831 {
2832         int process_refs, io_refs;
2833
2834         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2835         process_refs = cfqq->ref - io_refs;
2836         BUG_ON(process_refs < 0);
2837         return process_refs;
2838 }
2839
2840 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2841 {
2842         int process_refs, new_process_refs;
2843         struct cfq_queue *__cfqq;
2844
2845         /*
2846          * If there are no process references on the new_cfqq, then it is
2847          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2848          * chain may have dropped their last reference (not just their
2849          * last process reference).
2850          */
2851         if (!cfqq_process_refs(new_cfqq))
2852                 return;
2853
2854         /* Avoid a circular list and skip interim queue merges */
2855         while ((__cfqq = new_cfqq->new_cfqq)) {
2856                 if (__cfqq == cfqq)
2857                         return;
2858                 new_cfqq = __cfqq;
2859         }
2860
2861         process_refs = cfqq_process_refs(cfqq);
2862         new_process_refs = cfqq_process_refs(new_cfqq);
2863         /*
2864          * If the process for the cfqq has gone away, there is no
2865          * sense in merging the queues.
2866          */
2867         if (process_refs == 0 || new_process_refs == 0)
2868                 return;
2869
2870         /*
2871          * Merge in the direction of the lesser amount of work.
2872          */
2873         if (new_process_refs >= process_refs) {
2874                 cfqq->new_cfqq = new_cfqq;
2875                 new_cfqq->ref += process_refs;
2876         } else {
2877                 new_cfqq->new_cfqq = cfqq;
2878                 cfqq->ref += new_process_refs;
2879         }
2880 }
2881
2882 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2883                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2884 {
2885         struct cfq_queue *queue;
2886         int i;
2887         bool key_valid = false;
2888         unsigned long lowest_key = 0;
2889         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2890
2891         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2892                 /* select the one with lowest rb_key */
2893                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2894                 if (queue &&
2895                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2896                         lowest_key = queue->rb_key;
2897                         cur_best = i;
2898                         key_valid = true;
2899                 }
2900         }
2901
2902         return cur_best;
2903 }
2904
2905 static void
2906 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2907 {
2908         unsigned slice;
2909         unsigned count;
2910         struct cfq_rb_root *st;
2911         unsigned group_slice;
2912         enum wl_class_t original_class = cfqd->serving_wl_class;
2913
2914         /* Choose next priority. RT > BE > IDLE */
2915         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2916                 cfqd->serving_wl_class = RT_WORKLOAD;
2917         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2918                 cfqd->serving_wl_class = BE_WORKLOAD;
2919         else {
2920                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2921                 cfqd->workload_expires = jiffies + 1;
2922                 return;
2923         }
2924
2925         if (original_class != cfqd->serving_wl_class)
2926                 goto new_workload;
2927
2928         /*
2929          * For RT and BE, we have to choose also the type
2930          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2931          * expiration time
2932          */
2933         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2934         count = st->count;
2935
2936         /*
2937          * check workload expiration, and that we still have other queues ready
2938          */
2939         if (count && !time_after(jiffies, cfqd->workload_expires))
2940                 return;
2941
2942 new_workload:
2943         /* otherwise select new workload type */
2944         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2945                                         cfqd->serving_wl_class);
2946         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2947         count = st->count;
2948
2949         /*
2950          * the workload slice is computed as a fraction of target latency
2951          * proportional to the number of queues in that workload, over
2952          * all the queues in the same priority class
2953          */
2954         group_slice = cfq_group_slice(cfqd, cfqg);
2955
2956         slice = group_slice * count /
2957                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2958                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2959                                         cfqg));
2960
2961         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2962                 unsigned int tmp;
2963
2964                 /*
2965                  * Async queues are currently system wide. Just taking
2966                  * proportion of queues with-in same group will lead to higher
2967                  * async ratio system wide as generally root group is going
2968                  * to have higher weight. A more accurate thing would be to
2969                  * calculate system wide asnc/sync ratio.
2970                  */
2971                 tmp = cfqd->cfq_target_latency *
2972                         cfqg_busy_async_queues(cfqd, cfqg);
2973                 tmp = tmp/cfqd->busy_queues;
2974                 slice = min_t(unsigned, slice, tmp);
2975
2976                 /* async workload slice is scaled down according to
2977                  * the sync/async slice ratio. */
2978                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2979         } else
2980                 /* sync workload slice is at least 2 * cfq_slice_idle */
2981                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2982
2983         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2984         cfq_log(cfqd, "workload slice:%d", slice);
2985         cfqd->workload_expires = jiffies + slice;
2986 }
2987
2988 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2989 {
2990         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2991         struct cfq_group *cfqg;
2992
2993         if (RB_EMPTY_ROOT(&st->rb))
2994                 return NULL;
2995         cfqg = cfq_rb_first_group(st);
2996         update_min_vdisktime(st);
2997         return cfqg;
2998 }
2999
3000 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3001 {
3002         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3003
3004         cfqd->serving_group = cfqg;
3005
3006         /* Restore the workload type data */
3007         if (cfqg->saved_wl_slice) {
3008                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3009                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3010                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3011         } else
3012                 cfqd->workload_expires = jiffies - 1;
3013
3014         choose_wl_class_and_type(cfqd, cfqg);
3015 }
3016
3017 /*
3018  * Select a queue for service. If we have a current active queue,
3019  * check whether to continue servicing it, or retrieve and set a new one.
3020  */
3021 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3022 {
3023         struct cfq_queue *cfqq, *new_cfqq = NULL;
3024
3025         cfqq = cfqd->active_queue;
3026         if (!cfqq)
3027                 goto new_queue;
3028
3029         if (!cfqd->rq_queued)
3030                 return NULL;
3031
3032         /*
3033          * We were waiting for group to get backlogged. Expire the queue
3034          */
3035         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3036                 goto expire;
3037
3038         /*
3039          * The active queue has run out of time, expire it and select new.
3040          */
3041         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3042                 /*
3043                  * If slice had not expired at the completion of last request
3044                  * we might not have turned on wait_busy flag. Don't expire
3045                  * the queue yet. Allow the group to get backlogged.
3046                  *
3047                  * The very fact that we have used the slice, that means we
3048                  * have been idling all along on this queue and it should be
3049                  * ok to wait for this request to complete.
3050                  */
3051                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3052                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3053                         cfqq = NULL;
3054                         goto keep_queue;
3055                 } else
3056                         goto check_group_idle;
3057         }
3058
3059         /*
3060          * The active queue has requests and isn't expired, allow it to
3061          * dispatch.
3062          */
3063         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3064                 goto keep_queue;
3065
3066         /*
3067          * If another queue has a request waiting within our mean seek
3068          * distance, let it run.  The expire code will check for close
3069          * cooperators and put the close queue at the front of the service
3070          * tree.  If possible, merge the expiring queue with the new cfqq.
3071          */
3072         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3073         if (new_cfqq) {
3074                 if (!cfqq->new_cfqq)
3075                         cfq_setup_merge(cfqq, new_cfqq);
3076                 goto expire;
3077         }
3078
3079         /*
3080          * No requests pending. If the active queue still has requests in
3081          * flight or is idling for a new request, allow either of these
3082          * conditions to happen (or time out) before selecting a new queue.
3083          */
3084         if (timer_pending(&cfqd->idle_slice_timer)) {
3085                 cfqq = NULL;
3086                 goto keep_queue;
3087         }
3088
3089         /*
3090          * This is a deep seek queue, but the device is much faster than
3091          * the queue can deliver, don't idle
3092          **/
3093         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3094             (cfq_cfqq_slice_new(cfqq) ||
3095             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3096                 cfq_clear_cfqq_deep(cfqq);
3097                 cfq_clear_cfqq_idle_window(cfqq);
3098         }
3099
3100         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3101                 cfqq = NULL;
3102                 goto keep_queue;
3103         }
3104
3105         /*
3106          * If group idle is enabled and there are requests dispatched from
3107          * this group, wait for requests to complete.
3108          */
3109 check_group_idle:
3110         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3111             cfqq->cfqg->dispatched &&
3112             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3113                 cfqq = NULL;
3114                 goto keep_queue;
3115         }
3116
3117 expire:
3118         cfq_slice_expired(cfqd, 0);
3119 new_queue:
3120         /*
3121          * Current queue expired. Check if we have to switch to a new
3122          * service tree
3123          */
3124         if (!new_cfqq)
3125                 cfq_choose_cfqg(cfqd);
3126
3127         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3128 keep_queue:
3129         return cfqq;
3130 }
3131
3132 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3133 {
3134         int dispatched = 0;
3135
3136         while (cfqq->next_rq) {
3137                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3138                 dispatched++;
3139         }
3140
3141         BUG_ON(!list_empty(&cfqq->fifo));
3142
3143         /* By default cfqq is not expired if it is empty. Do it explicitly */
3144         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3145         return dispatched;
3146 }
3147
3148 /*
3149  * Drain our current requests. Used for barriers and when switching
3150  * io schedulers on-the-fly.
3151  */
3152 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3153 {
3154         struct cfq_queue *cfqq;
3155         int dispatched = 0;
3156
3157         /* Expire the timeslice of the current active queue first */
3158         cfq_slice_expired(cfqd, 0);
3159         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3160                 __cfq_set_active_queue(cfqd, cfqq);
3161                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3162         }
3163
3164         BUG_ON(cfqd->busy_queues);
3165
3166         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3167         return dispatched;
3168 }
3169
3170 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3171         struct cfq_queue *cfqq)
3172 {
3173         /* the queue hasn't finished any request, can't estimate */
3174         if (cfq_cfqq_slice_new(cfqq))
3175                 return true;
3176         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3177                 cfqq->slice_end))
3178                 return true;
3179
3180         return false;
3181 }
3182
3183 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3184 {
3185         unsigned int max_dispatch;
3186
3187         /*
3188          * Drain async requests before we start sync IO
3189          */
3190         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3191                 return false;
3192
3193         /*
3194          * If this is an async queue and we have sync IO in flight, let it wait
3195          */
3196         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3197                 return false;
3198
3199         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3200         if (cfq_class_idle(cfqq))
3201                 max_dispatch = 1;
3202
3203         /*
3204          * Does this cfqq already have too much IO in flight?
3205          */
3206         if (cfqq->dispatched >= max_dispatch) {
3207                 bool promote_sync = false;
3208                 /*
3209                  * idle queue must always only have a single IO in flight
3210                  */
3211                 if (cfq_class_idle(cfqq))
3212                         return false;
3213
3214                 /*
3215                  * If there is only one sync queue
3216                  * we can ignore async queue here and give the sync
3217                  * queue no dispatch limit. The reason is a sync queue can
3218                  * preempt async queue, limiting the sync queue doesn't make
3219                  * sense. This is useful for aiostress test.
3220                  */
3221                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3222                         promote_sync = true;
3223
3224                 /*
3225                  * We have other queues, don't allow more IO from this one
3226                  */
3227                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3228                                 !promote_sync)
3229                         return false;
3230
3231                 /*
3232                  * Sole queue user, no limit
3233                  */
3234                 if (cfqd->busy_queues == 1 || promote_sync)
3235                         max_dispatch = -1;
3236                 else
3237                         /*
3238                          * Normally we start throttling cfqq when cfq_quantum/2
3239                          * requests have been dispatched. But we can drive
3240                          * deeper queue depths at the beginning of slice
3241                          * subjected to upper limit of cfq_quantum.
3242                          * */
3243                         max_dispatch = cfqd->cfq_quantum;
3244         }
3245
3246         /*
3247          * Async queues must wait a bit before being allowed dispatch.
3248          * We also ramp up the dispatch depth gradually for async IO,
3249          * based on the last sync IO we serviced
3250          */
3251         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3252                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3253                 unsigned int depth;
3254
3255                 depth = last_sync / cfqd->cfq_slice[1];
3256                 if (!depth && !cfqq->dispatched)
3257                         depth = 1;
3258                 if (depth < max_dispatch)
3259                         max_dispatch = depth;
3260         }
3261
3262         /*
3263          * If we're below the current max, allow a dispatch
3264          */
3265         return cfqq->dispatched < max_dispatch;
3266 }
3267
3268 /*
3269  * Dispatch a request from cfqq, moving them to the request queue
3270  * dispatch list.
3271  */
3272 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3273 {
3274         struct request *rq;
3275
3276         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3277
3278         if (!cfq_may_dispatch(cfqd, cfqq))
3279                 return false;
3280
3281         /*
3282          * follow expired path, else get first next available
3283          */
3284         rq = cfq_check_fifo(cfqq);
3285         if (!rq)
3286                 rq = cfqq->next_rq;
3287
3288         /*
3289          * insert request into driver dispatch list
3290          */
3291         cfq_dispatch_insert(cfqd->queue, rq);
3292
3293         if (!cfqd->active_cic) {
3294                 struct cfq_io_cq *cic = RQ_CIC(rq);
3295
3296                 atomic_long_inc(&cic->icq.ioc->refcount);
3297                 cfqd->active_cic = cic;
3298         }
3299
3300         return true;
3301 }
3302
3303 /*
3304  * Find the cfqq that we need to service and move a request from that to the
3305  * dispatch list
3306  */
3307 static int cfq_dispatch_requests(struct request_queue *q, int force)
3308 {
3309         struct cfq_data *cfqd = q->elevator->elevator_data;
3310         struct cfq_queue *cfqq;
3311
3312         if (!cfqd->busy_queues)
3313                 return 0;
3314
3315         if (unlikely(force))
3316                 return cfq_forced_dispatch(cfqd);
3317
3318         cfqq = cfq_select_queue(cfqd);
3319         if (!cfqq)
3320                 return 0;
3321
3322         /*
3323          * Dispatch a request from this cfqq, if it is allowed
3324          */
3325         if (!cfq_dispatch_request(cfqd, cfqq))
3326                 return 0;
3327
3328         cfqq->slice_dispatch++;
3329         cfq_clear_cfqq_must_dispatch(cfqq);
3330
3331         /*
3332          * expire an async queue immediately if it has used up its slice. idle
3333          * queue always expire after 1 dispatch round.
3334          */
3335         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3336             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3337             cfq_class_idle(cfqq))) {
3338                 cfqq->slice_end = jiffies + 1;
3339                 cfq_slice_expired(cfqd, 0);
3340         }
3341
3342         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3343         return 1;
3344 }
3345
3346 /*
3347  * task holds one reference to the queue, dropped when task exits. each rq
3348  * in-flight on this queue also holds a reference, dropped when rq is freed.
3349  *
3350  * Each cfq queue took a reference on the parent group. Drop it now.
3351  * queue lock must be held here.
3352  */
3353 static void cfq_put_queue(struct cfq_queue *cfqq)
3354 {
3355         struct cfq_data *cfqd = cfqq->cfqd;
3356         struct cfq_group *cfqg;
3357
3358         BUG_ON(cfqq->ref <= 0);
3359
3360         cfqq->ref--;
3361         if (cfqq->ref)
3362                 return;
3363
3364         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3365         BUG_ON(rb_first(&cfqq->sort_list));
3366         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3367         cfqg = cfqq->cfqg;
3368
3369         if (unlikely(cfqd->active_queue == cfqq)) {
3370                 __cfq_slice_expired(cfqd, cfqq, 0);
3371                 cfq_schedule_dispatch(cfqd);
3372         }
3373
3374         BUG_ON(cfq_cfqq_on_rr(cfqq));
3375         kmem_cache_free(cfq_pool, cfqq);
3376         cfqg_put(cfqg);
3377 }
3378
3379 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3380 {
3381         struct cfq_queue *__cfqq, *next;
3382
3383         /*
3384          * If this queue was scheduled to merge with another queue, be
3385          * sure to drop the reference taken on that queue (and others in
3386          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3387          */
3388         __cfqq = cfqq->new_cfqq;
3389         while (__cfqq) {
3390                 if (__cfqq == cfqq) {
3391                         WARN(1, "cfqq->new_cfqq loop detected\n");
3392                         break;
3393                 }
3394                 next = __cfqq->new_cfqq;
3395                 cfq_put_queue(__cfqq);
3396                 __cfqq = next;
3397         }
3398 }
3399
3400 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3401 {
3402         if (unlikely(cfqq == cfqd->active_queue)) {
3403                 __cfq_slice_expired(cfqd, cfqq, 0);
3404                 cfq_schedule_dispatch(cfqd);
3405         }
3406
3407         cfq_put_cooperator(cfqq);
3408
3409         cfq_put_queue(cfqq);
3410 }
3411
3412 static void cfq_init_icq(struct io_cq *icq)
3413 {
3414         struct cfq_io_cq *cic = icq_to_cic(icq);
3415
3416         cic->ttime.last_end_request = jiffies;
3417 }
3418
3419 static void cfq_exit_icq(struct io_cq *icq)
3420 {
3421         struct cfq_io_cq *cic = icq_to_cic(icq);
3422         struct cfq_data *cfqd = cic_to_cfqd(cic);
3423
3424         if (cic->cfqq[BLK_RW_ASYNC]) {
3425                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3426                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3427         }
3428
3429         if (cic->cfqq[BLK_RW_SYNC]) {
3430                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3431                 cic->cfqq[BLK_RW_SYNC] = NULL;
3432         }
3433 }
3434
3435 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3436 {
3437         struct task_struct *tsk = current;
3438         int ioprio_class;
3439
3440         if (!cfq_cfqq_prio_changed(cfqq))
3441                 return;
3442
3443         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3444         switch (ioprio_class) {
3445         default:
3446                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3447         case IOPRIO_CLASS_NONE:
3448                 /*
3449                  * no prio set, inherit CPU scheduling settings
3450                  */
3451                 cfqq->ioprio = task_nice_ioprio(tsk);
3452                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3453                 break;
3454         case IOPRIO_CLASS_RT:
3455                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3456                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3457                 break;
3458         case IOPRIO_CLASS_BE:
3459                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3460                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3461                 break;
3462         case IOPRIO_CLASS_IDLE:
3463                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3464                 cfqq->ioprio = 7;
3465                 cfq_clear_cfqq_idle_window(cfqq);
3466                 break;
3467         }
3468
3469         /*
3470          * keep track of original prio settings in case we have to temporarily
3471          * elevate the priority of this queue
3472          */
3473         cfqq->org_ioprio = cfqq->ioprio;
3474         cfq_clear_cfqq_prio_changed(cfqq);
3475 }
3476
3477 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3478 {
3479         int ioprio = cic->icq.ioc->ioprio;
3480         struct cfq_data *cfqd = cic_to_cfqd(cic);
3481         struct cfq_queue *cfqq;
3482
3483         /*
3484          * Check whether ioprio has changed.  The condition may trigger
3485          * spuriously on a newly created cic but there's no harm.
3486          */
3487         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3488                 return;
3489
3490         cfqq = cic->cfqq[BLK_RW_ASYNC];
3491         if (cfqq) {
3492                 struct cfq_queue *new_cfqq;
3493                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3494                                          GFP_ATOMIC);
3495                 if (new_cfqq) {
3496                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3497                         cfq_put_queue(cfqq);
3498                 }
3499         }
3500
3501         cfqq = cic->cfqq[BLK_RW_SYNC];
3502         if (cfqq)
3503                 cfq_mark_cfqq_prio_changed(cfqq);
3504
3505         cic->ioprio = ioprio;
3506 }
3507
3508 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3509                           pid_t pid, bool is_sync)
3510 {
3511         RB_CLEAR_NODE(&cfqq->rb_node);
3512         RB_CLEAR_NODE(&cfqq->p_node);
3513         INIT_LIST_HEAD(&cfqq->fifo);
3514
3515         cfqq->ref = 0;
3516         cfqq->cfqd = cfqd;
3517
3518         cfq_mark_cfqq_prio_changed(cfqq);
3519
3520         if (is_sync) {
3521                 if (!cfq_class_idle(cfqq))
3522                         cfq_mark_cfqq_idle_window(cfqq);
3523                 cfq_mark_cfqq_sync(cfqq);
3524         }
3525         cfqq->pid = pid;
3526 }
3527
3528 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3529 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3530 {
3531         struct cfq_data *cfqd = cic_to_cfqd(cic);
3532         struct cfq_queue *sync_cfqq;
3533         uint64_t id;
3534
3535         rcu_read_lock();
3536         id = bio_blkcg(bio)->id;
3537         rcu_read_unlock();
3538
3539         /*
3540          * Check whether blkcg has changed.  The condition may trigger
3541          * spuriously on a newly created cic but there's no harm.
3542          */
3543         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3544                 return;
3545
3546         sync_cfqq = cic_to_cfqq(cic, 1);
3547         if (sync_cfqq) {
3548                 /*
3549                  * Drop reference to sync queue. A new sync queue will be
3550                  * assigned in new group upon arrival of a fresh request.
3551                  */
3552                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3553                 cic_set_cfqq(cic, NULL, 1);
3554                 cfq_put_queue(sync_cfqq);
3555         }
3556
3557         cic->blkcg_id = id;
3558 }
3559 #else
3560 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3561 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3562
3563 static struct cfq_queue *
3564 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3565                      struct bio *bio, gfp_t gfp_mask)
3566 {
3567         struct blkcg *blkcg;
3568         struct cfq_queue *cfqq, *new_cfqq = NULL;
3569         struct cfq_group *cfqg;
3570
3571 retry:
3572         rcu_read_lock();
3573
3574         blkcg = bio_blkcg(bio);
3575         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3576         cfqq = cic_to_cfqq(cic, is_sync);
3577
3578         /*
3579          * Always try a new alloc if we fell back to the OOM cfqq
3580          * originally, since it should just be a temporary situation.
3581          */
3582         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3583                 cfqq = NULL;
3584                 if (new_cfqq) {
3585                         cfqq = new_cfqq;
3586                         new_cfqq = NULL;
3587                 } else if (gfp_mask & __GFP_WAIT) {
3588                         rcu_read_unlock();
3589                         spin_unlock_irq(cfqd->queue->queue_lock);
3590                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3591                                         gfp_mask | __GFP_ZERO,
3592                                         cfqd->queue->node);
3593                         spin_lock_irq(cfqd->queue->queue_lock);
3594                         if (new_cfqq)
3595                                 goto retry;
3596                         else
3597                                 return &cfqd->oom_cfqq;
3598                 } else {
3599                         cfqq = kmem_cache_alloc_node(cfq_pool,
3600                                         gfp_mask | __GFP_ZERO,
3601                                         cfqd->queue->node);
3602                 }
3603
3604                 if (cfqq) {
3605                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3606                         cfq_init_prio_data(cfqq, cic);
3607                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3608                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3609                 } else
3610                         cfqq = &cfqd->oom_cfqq;
3611         }
3612
3613         if (new_cfqq)
3614                 kmem_cache_free(cfq_pool, new_cfqq);
3615
3616         rcu_read_unlock();
3617         return cfqq;
3618 }
3619
3620 static struct cfq_queue **
3621 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3622 {
3623         switch (ioprio_class) {
3624         case IOPRIO_CLASS_RT:
3625                 return &cfqd->async_cfqq[0][ioprio];
3626         case IOPRIO_CLASS_NONE:
3627                 ioprio = IOPRIO_NORM;
3628                 /* fall through */
3629         case IOPRIO_CLASS_BE:
3630                 return &cfqd->async_cfqq[1][ioprio];
3631         case IOPRIO_CLASS_IDLE:
3632                 return &cfqd->async_idle_cfqq;
3633         default:
3634                 BUG();
3635         }
3636 }
3637
3638 static struct cfq_queue *
3639 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3640               struct bio *bio, gfp_t gfp_mask)
3641 {
3642         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3643         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3644         struct cfq_queue **async_cfqq = NULL;
3645         struct cfq_queue *cfqq = NULL;
3646
3647         if (!is_sync) {
3648                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3649                 cfqq = *async_cfqq;
3650         }
3651
3652         if (!cfqq)
3653                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3654
3655         /*
3656          * pin the queue now that it's allocated, scheduler exit will prune it
3657          */
3658         if (!is_sync && !(*async_cfqq)) {
3659                 cfqq->ref++;
3660                 *async_cfqq = cfqq;
3661         }
3662
3663         cfqq->ref++;
3664         return cfqq;
3665 }
3666
3667 static void
3668 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3669 {
3670         unsigned long elapsed = jiffies - ttime->last_end_request;
3671         elapsed = min(elapsed, 2UL * slice_idle);
3672
3673         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3674         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3675         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3676 }
3677
3678 static void
3679 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3680                         struct cfq_io_cq *cic)
3681 {
3682         if (cfq_cfqq_sync(cfqq)) {
3683                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3684                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3685                         cfqd->cfq_slice_idle);
3686         }
3687 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3688         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3689 #endif
3690 }
3691
3692 static void
3693 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3694                        struct request *rq)
3695 {
3696         sector_t sdist = 0;
3697         sector_t n_sec = blk_rq_sectors(rq);
3698         if (cfqq->last_request_pos) {
3699                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3700                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3701                 else
3702                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3703         }
3704
3705         cfqq->seek_history <<= 1;
3706         if (blk_queue_nonrot(cfqd->queue))
3707                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3708         else
3709                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3710 }
3711
3712 /*
3713  * Disable idle window if the process thinks too long or seeks so much that
3714  * it doesn't matter
3715  */
3716 static void
3717 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3718                        struct cfq_io_cq *cic)
3719 {
3720         int old_idle, enable_idle;
3721
3722         /*
3723          * Don't idle for async or idle io prio class
3724          */
3725         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3726                 return;
3727
3728         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3729
3730         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3731                 cfq_mark_cfqq_deep(cfqq);
3732
3733         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3734                 enable_idle = 0;
3735         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3736                  !cfqd->cfq_slice_idle ||
3737                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3738                 enable_idle = 0;
3739         else if (sample_valid(cic->ttime.ttime_samples)) {
3740                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3741                         enable_idle = 0;
3742                 else
3743                         enable_idle = 1;
3744         }
3745
3746         if (old_idle != enable_idle) {
3747                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3748                 if (enable_idle)
3749                         cfq_mark_cfqq_idle_window(cfqq);
3750                 else
3751                         cfq_clear_cfqq_idle_window(cfqq);
3752         }
3753 }
3754
3755 /*
3756  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3757  * no or if we aren't sure, a 1 will cause a preempt.
3758  */
3759 static bool
3760 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3761                    struct request *rq)
3762 {
3763         struct cfq_queue *cfqq;
3764
3765         cfqq = cfqd->active_queue;
3766         if (!cfqq)
3767                 return false;
3768
3769         if (cfq_class_idle(new_cfqq))
3770                 return false;
3771
3772         if (cfq_class_idle(cfqq))
3773                 return true;
3774
3775         /*
3776          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3777          */
3778         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3779                 return false;
3780
3781         /*
3782          * if the new request is sync, but the currently running queue is
3783          * not, let the sync request have priority.
3784          */
3785         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3786                 return true;
3787
3788         if (new_cfqq->cfqg != cfqq->cfqg)
3789                 return false;
3790
3791         if (cfq_slice_used(cfqq))
3792                 return true;
3793
3794         /* Allow preemption only if we are idling on sync-noidle tree */
3795         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3796             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3797             new_cfqq->service_tree->count == 2 &&
3798             RB_EMPTY_ROOT(&cfqq->sort_list))
3799                 return true;
3800
3801         /*
3802          * So both queues are sync. Let the new request get disk time if
3803          * it's a metadata request and the current queue is doing regular IO.
3804          */
3805         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3806                 return true;
3807
3808         /*
3809          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3810          */
3811         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3812                 return true;
3813
3814         /* An idle queue should not be idle now for some reason */
3815         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3816                 return true;
3817
3818         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3819                 return false;
3820
3821         /*
3822          * if this request is as-good as one we would expect from the
3823          * current cfqq, let it preempt
3824          */
3825         if (cfq_rq_close(cfqd, cfqq, rq))
3826                 return true;
3827
3828         return false;
3829 }
3830
3831 /*
3832  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3833  * let it have half of its nominal slice.
3834  */
3835 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3836 {
3837         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3838
3839         cfq_log_cfqq(cfqd, cfqq, "preempt");
3840         cfq_slice_expired(cfqd, 1);
3841
3842         /*
3843          * workload type is changed, don't save slice, otherwise preempt
3844          * doesn't happen
3845          */
3846         if (old_type != cfqq_type(cfqq))
3847                 cfqq->cfqg->saved_wl_slice = 0;
3848
3849         /*
3850          * Put the new queue at the front of the of the current list,
3851          * so we know that it will be selected next.
3852          */
3853         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3854
3855         cfq_service_tree_add(cfqd, cfqq, 1);
3856
3857         cfqq->slice_end = 0;
3858         cfq_mark_cfqq_slice_new(cfqq);
3859 }
3860
3861 /*
3862  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3863  * something we should do about it
3864  */
3865 static void
3866 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3867                 struct request *rq)
3868 {
3869         struct cfq_io_cq *cic = RQ_CIC(rq);
3870
3871         cfqd->rq_queued++;
3872         if (rq->cmd_flags & REQ_PRIO)
3873                 cfqq->prio_pending++;
3874
3875         cfq_update_io_thinktime(cfqd, cfqq, cic);
3876         cfq_update_io_seektime(cfqd, cfqq, rq);
3877         cfq_update_idle_window(cfqd, cfqq, cic);
3878
3879         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3880
3881         if (cfqq == cfqd->active_queue) {
3882                 /*
3883                  * Remember that we saw a request from this process, but
3884                  * don't start queuing just yet. Otherwise we risk seeing lots
3885                  * of tiny requests, because we disrupt the normal plugging
3886                  * and merging. If the request is already larger than a single
3887                  * page, let it rip immediately. For that case we assume that
3888                  * merging is already done. Ditto for a busy system that
3889                  * has other work pending, don't risk delaying until the
3890                  * idle timer unplug to continue working.
3891                  */
3892                 if (cfq_cfqq_wait_request(cfqq)) {
3893                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3894                             cfqd->busy_queues > 1) {
3895                                 cfq_del_timer(cfqd, cfqq);
3896                                 cfq_clear_cfqq_wait_request(cfqq);
3897                                 __blk_run_queue(cfqd->queue);
3898                         } else {
3899                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3900                                 cfq_mark_cfqq_must_dispatch(cfqq);
3901                         }
3902                 }
3903         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3904                 /*
3905                  * not the active queue - expire current slice if it is
3906                  * idle and has expired it's mean thinktime or this new queue
3907                  * has some old slice time left and is of higher priority or
3908                  * this new queue is RT and the current one is BE
3909                  */
3910                 cfq_preempt_queue(cfqd, cfqq);
3911                 __blk_run_queue(cfqd->queue);
3912         }
3913 }
3914
3915 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3916 {
3917         struct cfq_data *cfqd = q->elevator->elevator_data;
3918         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3919
3920         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3921         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3922
3923         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3924         list_add_tail(&rq->queuelist, &cfqq->fifo);
3925         cfq_add_rq_rb(rq);
3926         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3927                                  rq->cmd_flags);
3928         cfq_rq_enqueued(cfqd, cfqq, rq);
3929 }
3930
3931 /*
3932  * Update hw_tag based on peak queue depth over 50 samples under
3933  * sufficient load.
3934  */
3935 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3936 {
3937         struct cfq_queue *cfqq = cfqd->active_queue;
3938
3939         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3940                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3941
3942         if (cfqd->hw_tag == 1)
3943                 return;
3944
3945         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3946             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3947                 return;
3948
3949         /*
3950          * If active queue hasn't enough requests and can idle, cfq might not
3951          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3952          * case
3953          */
3954         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3955             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3956             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3957                 return;
3958
3959         if (cfqd->hw_tag_samples++ < 50)
3960                 return;
3961
3962         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3963                 cfqd->hw_tag = 1;
3964         else
3965                 cfqd->hw_tag = 0;
3966 }
3967
3968 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3969 {
3970         struct cfq_io_cq *cic = cfqd->active_cic;
3971
3972         /* If the queue already has requests, don't wait */
3973         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3974                 return false;
3975
3976         /* If there are other queues in the group, don't wait */
3977         if (cfqq->cfqg->nr_cfqq > 1)
3978                 return false;
3979
3980         /* the only queue in the group, but think time is big */
3981         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3982                 return false;
3983
3984         if (cfq_slice_used(cfqq))
3985                 return true;
3986
3987         /* if slice left is less than think time, wait busy */
3988         if (cic && sample_valid(cic->ttime.ttime_samples)
3989             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3990                 return true;
3991
3992         /*
3993          * If think times is less than a jiffy than ttime_mean=0 and above
3994          * will not be true. It might happen that slice has not expired yet
3995          * but will expire soon (4-5 ns) during select_queue(). To cover the
3996          * case where think time is less than a jiffy, mark the queue wait
3997          * busy if only 1 jiffy is left in the slice.
3998          */
3999         if (cfqq->slice_end - jiffies == 1)
4000                 return true;
4001
4002         return false;
4003 }
4004
4005 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4006 {
4007         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4008         struct cfq_data *cfqd = cfqq->cfqd;
4009         const int sync = rq_is_sync(rq);
4010         unsigned long now;
4011
4012         now = jiffies;
4013         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4014                      !!(rq->cmd_flags & REQ_NOIDLE));
4015
4016         cfq_update_hw_tag(cfqd);
4017
4018         WARN_ON(!cfqd->rq_in_driver);
4019         WARN_ON(!cfqq->dispatched);
4020         cfqd->rq_in_driver--;
4021         cfqq->dispatched--;
4022         (RQ_CFQG(rq))->dispatched--;
4023         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4024                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4025
4026         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4027
4028         if (sync) {
4029                 struct cfq_rb_root *st;
4030
4031                 RQ_CIC(rq)->ttime.last_end_request = now;
4032
4033                 if (cfq_cfqq_on_rr(cfqq))
4034                         st = cfqq->service_tree;
4035                 else
4036                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4037                                         cfqq_type(cfqq));
4038
4039                 st->ttime.last_end_request = now;
4040                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4041                         cfqd->last_delayed_sync = now;
4042         }
4043
4044 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4045         cfqq->cfqg->ttime.last_end_request = now;
4046 #endif
4047
4048         /*
4049          * If this is the active queue, check if it needs to be expired,
4050          * or if we want to idle in case it has no pending requests.
4051          */
4052         if (cfqd->active_queue == cfqq) {
4053                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4054
4055                 if (cfq_cfqq_slice_new(cfqq)) {
4056                         cfq_set_prio_slice(cfqd, cfqq);
4057                         cfq_clear_cfqq_slice_new(cfqq);
4058                 }
4059
4060                 /*
4061                  * Should we wait for next request to come in before we expire
4062                  * the queue.
4063                  */
4064                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4065                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4066                         if (!cfqd->cfq_slice_idle)
4067                                 extend_sl = cfqd->cfq_group_idle;
4068                         cfqq->slice_end = jiffies + extend_sl;
4069                         cfq_mark_cfqq_wait_busy(cfqq);
4070                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4071                 }
4072
4073                 /*
4074                  * Idling is not enabled on:
4075                  * - expired queues
4076                  * - idle-priority queues
4077                  * - async queues
4078                  * - queues with still some requests queued
4079                  * - when there is a close cooperator
4080                  */
4081                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4082                         cfq_slice_expired(cfqd, 1);
4083                 else if (sync && cfqq_empty &&
4084                          !cfq_close_cooperator(cfqd, cfqq)) {
4085                         cfq_arm_slice_timer(cfqd);
4086                 }
4087         }
4088
4089         if (!cfqd->rq_in_driver)
4090                 cfq_schedule_dispatch(cfqd);
4091 }
4092
4093 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4094 {
4095         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4096                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4097                 return ELV_MQUEUE_MUST;
4098         }
4099
4100         return ELV_MQUEUE_MAY;
4101 }
4102
4103 static int cfq_may_queue(struct request_queue *q, int rw)
4104 {
4105         struct cfq_data *cfqd = q->elevator->elevator_data;
4106         struct task_struct *tsk = current;
4107         struct cfq_io_cq *cic;
4108         struct cfq_queue *cfqq;
4109
4110         /*
4111          * don't force setup of a queue from here, as a call to may_queue
4112          * does not necessarily imply that a request actually will be queued.
4113          * so just lookup a possibly existing queue, or return 'may queue'
4114          * if that fails
4115          */
4116         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4117         if (!cic)
4118                 return ELV_MQUEUE_MAY;
4119
4120         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4121         if (cfqq) {
4122                 cfq_init_prio_data(cfqq, cic);
4123
4124                 return __cfq_may_queue(cfqq);
4125         }
4126
4127         return ELV_MQUEUE_MAY;
4128 }
4129
4130 /*
4131  * queue lock held here
4132  */
4133 static void cfq_put_request(struct request *rq)
4134 {
4135         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4136
4137         if (cfqq) {
4138                 const int rw = rq_data_dir(rq);
4139
4140                 BUG_ON(!cfqq->allocated[rw]);
4141                 cfqq->allocated[rw]--;
4142
4143                 /* Put down rq reference on cfqg */
4144                 cfqg_put(RQ_CFQG(rq));
4145                 rq->elv.priv[0] = NULL;
4146                 rq->elv.priv[1] = NULL;
4147
4148                 cfq_put_queue(cfqq);
4149         }
4150 }
4151
4152 static struct cfq_queue *
4153 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4154                 struct cfq_queue *cfqq)
4155 {
4156         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4157         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4158         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4159         cfq_put_queue(cfqq);
4160         return cic_to_cfqq(cic, 1);
4161 }
4162
4163 /*
4164  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4165  * was the last process referring to said cfqq.
4166  */
4167 static struct cfq_queue *
4168 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4169 {
4170         if (cfqq_process_refs(cfqq) == 1) {
4171                 cfqq->pid = current->pid;
4172                 cfq_clear_cfqq_coop(cfqq);
4173                 cfq_clear_cfqq_split_coop(cfqq);
4174                 return cfqq;
4175         }
4176
4177         cic_set_cfqq(cic, NULL, 1);
4178
4179         cfq_put_cooperator(cfqq);
4180
4181         cfq_put_queue(cfqq);
4182         return NULL;
4183 }
4184 /*
4185  * Allocate cfq data structures associated with this request.
4186  */
4187 static int
4188 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4189                 gfp_t gfp_mask)
4190 {
4191         struct cfq_data *cfqd = q->elevator->elevator_data;
4192         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4193         const int rw = rq_data_dir(rq);
4194         const bool is_sync = rq_is_sync(rq);
4195         struct cfq_queue *cfqq;
4196
4197         might_sleep_if(gfp_mask & __GFP_WAIT);
4198
4199         spin_lock_irq(q->queue_lock);
4200
4201         check_ioprio_changed(cic, bio);
4202         check_blkcg_changed(cic, bio);
4203 new_queue:
4204         cfqq = cic_to_cfqq(cic, is_sync);
4205         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4206                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4207                 cic_set_cfqq(cic, cfqq, is_sync);
4208         } else {
4209                 /*
4210                  * If the queue was seeky for too long, break it apart.
4211                  */
4212                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4213                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4214                         cfqq = split_cfqq(cic, cfqq);
4215                         if (!cfqq)
4216                                 goto new_queue;
4217                 }
4218
4219                 /*
4220                  * Check to see if this queue is scheduled to merge with
4221                  * another, closely cooperating queue.  The merging of
4222                  * queues happens here as it must be done in process context.
4223                  * The reference on new_cfqq was taken in merge_cfqqs.
4224                  */
4225                 if (cfqq->new_cfqq)
4226                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4227         }
4228
4229         cfqq->allocated[rw]++;
4230
4231         cfqq->ref++;
4232         cfqg_get(cfqq->cfqg);
4233         rq->elv.priv[0] = cfqq;
4234         rq->elv.priv[1] = cfqq->cfqg;
4235         spin_unlock_irq(q->queue_lock);
4236         return 0;
4237 }
4238
4239 static void cfq_kick_queue(struct work_struct *work)
4240 {
4241         struct cfq_data *cfqd =
4242                 container_of(work, struct cfq_data, unplug_work);
4243         struct request_queue *q = cfqd->queue;
4244
4245         spin_lock_irq(q->queue_lock);
4246         __blk_run_queue(cfqd->queue);
4247         spin_unlock_irq(q->queue_lock);
4248 }
4249
4250 /*
4251  * Timer running if the active_queue is currently idling inside its time slice
4252  */
4253 static void cfq_idle_slice_timer(unsigned long data)
4254 {
4255         struct cfq_data *cfqd = (struct cfq_data *) data;
4256         struct cfq_queue *cfqq;
4257         unsigned long flags;
4258         int timed_out = 1;
4259
4260         cfq_log(cfqd, "idle timer fired");
4261
4262         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4263
4264         cfqq = cfqd->active_queue;
4265         if (cfqq) {
4266                 timed_out = 0;
4267
4268                 /*
4269                  * We saw a request before the queue expired, let it through
4270                  */
4271                 if (cfq_cfqq_must_dispatch(cfqq))
4272                         goto out_kick;
4273
4274                 /*
4275                  * expired
4276                  */
4277                 if (cfq_slice_used(cfqq))
4278                         goto expire;
4279
4280                 /*
4281                  * only expire and reinvoke request handler, if there are
4282                  * other queues with pending requests
4283                  */
4284                 if (!cfqd->busy_queues)
4285                         goto out_cont;
4286
4287                 /*
4288                  * not expired and it has a request pending, let it dispatch
4289                  */
4290                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4291                         goto out_kick;
4292
4293                 /*
4294                  * Queue depth flag is reset only when the idle didn't succeed
4295                  */
4296                 cfq_clear_cfqq_deep(cfqq);
4297         }
4298 expire:
4299         cfq_slice_expired(cfqd, timed_out);
4300 out_kick:
4301         cfq_schedule_dispatch(cfqd);
4302 out_cont:
4303         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4304 }
4305
4306 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4307 {
4308         del_timer_sync(&cfqd->idle_slice_timer);
4309         cancel_work_sync(&cfqd->unplug_work);
4310 }
4311
4312 static void cfq_put_async_queues(struct cfq_data *cfqd)
4313 {
4314         int i;
4315
4316         for (i = 0; i < IOPRIO_BE_NR; i++) {
4317                 if (cfqd->async_cfqq[0][i])
4318                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4319                 if (cfqd->async_cfqq[1][i])
4320                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4321         }
4322
4323         if (cfqd->async_idle_cfqq)
4324                 cfq_put_queue(cfqd->async_idle_cfqq);
4325 }
4326
4327 static void cfq_exit_queue(struct elevator_queue *e)
4328 {
4329         struct cfq_data *cfqd = e->elevator_data;
4330         struct request_queue *q = cfqd->queue;
4331
4332         cfq_shutdown_timer_wq(cfqd);
4333
4334         spin_lock_irq(q->queue_lock);
4335
4336         if (cfqd->active_queue)
4337                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4338
4339         cfq_put_async_queues(cfqd);
4340
4341         spin_unlock_irq(q->queue_lock);
4342
4343         cfq_shutdown_timer_wq(cfqd);
4344
4345 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4346         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4347 #else
4348         kfree(cfqd->root_group);
4349 #endif
4350         kfree(cfqd);
4351 }
4352
4353 static int cfq_init_queue(struct request_queue *q)
4354 {
4355         struct cfq_data *cfqd;
4356         struct blkcg_gq *blkg __maybe_unused;
4357         int i, ret;
4358
4359         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4360         if (!cfqd)
4361                 return -ENOMEM;
4362
4363         cfqd->queue = q;
4364         q->elevator->elevator_data = cfqd;
4365
4366         /* Init root service tree */
4367         cfqd->grp_service_tree = CFQ_RB_ROOT;
4368
4369         /* Init root group and prefer root group over other groups by default */
4370 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4371         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4372         if (ret)
4373                 goto out_free;
4374
4375         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4376 #else
4377         ret = -ENOMEM;
4378         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4379                                         GFP_KERNEL, cfqd->queue->node);
4380         if (!cfqd->root_group)
4381                 goto out_free;
4382
4383         cfq_init_cfqg_base(cfqd->root_group);
4384 #endif
4385         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4386         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4387
4388         /*
4389          * Not strictly needed (since RB_ROOT just clears the node and we
4390          * zeroed cfqd on alloc), but better be safe in case someone decides
4391          * to add magic to the rb code
4392          */
4393         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4394                 cfqd->prio_trees[i] = RB_ROOT;
4395
4396         /*
4397          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4398          * Grab a permanent reference to it, so that the normal code flow
4399          * will not attempt to free it.  oom_cfqq is linked to root_group
4400          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4401          * the reference from linking right away.
4402          */
4403         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4404         cfqd->oom_cfqq.ref++;
4405
4406         spin_lock_irq(q->queue_lock);
4407         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4408         cfqg_put(cfqd->root_group);
4409         spin_unlock_irq(q->queue_lock);
4410
4411         init_timer(&cfqd->idle_slice_timer);
4412         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4413         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4414
4415         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4416
4417         cfqd->cfq_quantum = cfq_quantum;
4418         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4419         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4420         cfqd->cfq_back_max = cfq_back_max;
4421         cfqd->cfq_back_penalty = cfq_back_penalty;
4422         cfqd->cfq_slice[0] = cfq_slice_async;
4423         cfqd->cfq_slice[1] = cfq_slice_sync;
4424         cfqd->cfq_target_latency = cfq_target_latency;
4425         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4426         cfqd->cfq_slice_idle = cfq_slice_idle;
4427         cfqd->cfq_group_idle = cfq_group_idle;
4428         cfqd->cfq_latency = 1;
4429         cfqd->hw_tag = -1;
4430         /*
4431          * we optimistically start assuming sync ops weren't delayed in last
4432          * second, in order to have larger depth for async operations.
4433          */
4434         cfqd->last_delayed_sync = jiffies - HZ;
4435         return 0;
4436
4437 out_free:
4438         kfree(cfqd);
4439         return ret;
4440 }
4441
4442 /*
4443  * sysfs parts below -->
4444  */
4445 static ssize_t
4446 cfq_var_show(unsigned int var, char *page)
4447 {
4448         return sprintf(page, "%d\n", var);
4449 }
4450
4451 static ssize_t
4452 cfq_var_store(unsigned int *var, const char *page, size_t count)
4453 {
4454         char *p = (char *) page;
4455
4456         *var = simple_strtoul(p, &p, 10);
4457         return count;
4458 }
4459
4460 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4461 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4462 {                                                                       \
4463         struct cfq_data *cfqd = e->elevator_data;                       \
4464         unsigned int __data = __VAR;                                    \
4465         if (__CONV)                                                     \
4466                 __data = jiffies_to_msecs(__data);                      \
4467         return cfq_var_show(__data, (page));                            \
4468 }
4469 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4470 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4471 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4472 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4473 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4474 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4475 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4476 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4477 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4478 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4479 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4480 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4481 #undef SHOW_FUNCTION
4482
4483 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4484 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4485 {                                                                       \
4486         struct cfq_data *cfqd = e->elevator_data;                       \
4487         unsigned int __data;                                            \
4488         int ret = cfq_var_store(&__data, (page), count);                \
4489         if (__data < (MIN))                                             \
4490                 __data = (MIN);                                         \
4491         else if (__data > (MAX))                                        \
4492                 __data = (MAX);                                         \
4493         if (__CONV)                                                     \
4494                 *(__PTR) = msecs_to_jiffies(__data);                    \
4495         else                                                            \
4496                 *(__PTR) = __data;                                      \
4497         return ret;                                                     \
4498 }
4499 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4500 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4501                 UINT_MAX, 1);
4502 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4503                 UINT_MAX, 1);
4504 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4505 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4506                 UINT_MAX, 0);
4507 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4508 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4509 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4510 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4511 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4512                 UINT_MAX, 0);
4513 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4514 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4515 #undef STORE_FUNCTION
4516
4517 #define CFQ_ATTR(name) \
4518         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4519
4520 static struct elv_fs_entry cfq_attrs[] = {
4521         CFQ_ATTR(quantum),
4522         CFQ_ATTR(fifo_expire_sync),
4523         CFQ_ATTR(fifo_expire_async),
4524         CFQ_ATTR(back_seek_max),
4525         CFQ_ATTR(back_seek_penalty),
4526         CFQ_ATTR(slice_sync),
4527         CFQ_ATTR(slice_async),
4528         CFQ_ATTR(slice_async_rq),
4529         CFQ_ATTR(slice_idle),
4530         CFQ_ATTR(group_idle),
4531         CFQ_ATTR(low_latency),
4532         CFQ_ATTR(target_latency),
4533         __ATTR_NULL
4534 };
4535
4536 static struct elevator_type iosched_cfq = {
4537         .ops = {
4538                 .elevator_merge_fn =            cfq_merge,
4539                 .elevator_merged_fn =           cfq_merged_request,
4540                 .elevator_merge_req_fn =        cfq_merged_requests,
4541                 .elevator_allow_merge_fn =      cfq_allow_merge,
4542                 .elevator_bio_merged_fn =       cfq_bio_merged,
4543                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4544                 .elevator_add_req_fn =          cfq_insert_request,
4545                 .elevator_activate_req_fn =     cfq_activate_request,
4546                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4547                 .elevator_completed_req_fn =    cfq_completed_request,
4548                 .elevator_former_req_fn =       elv_rb_former_request,
4549                 .elevator_latter_req_fn =       elv_rb_latter_request,
4550                 .elevator_init_icq_fn =         cfq_init_icq,
4551                 .elevator_exit_icq_fn =         cfq_exit_icq,
4552                 .elevator_set_req_fn =          cfq_set_request,
4553                 .elevator_put_req_fn =          cfq_put_request,
4554                 .elevator_may_queue_fn =        cfq_may_queue,
4555                 .elevator_init_fn =             cfq_init_queue,
4556                 .elevator_exit_fn =             cfq_exit_queue,
4557         },
4558         .icq_size       =       sizeof(struct cfq_io_cq),
4559         .icq_align      =       __alignof__(struct cfq_io_cq),
4560         .elevator_attrs =       cfq_attrs,
4561         .elevator_name  =       "cfq",
4562         .elevator_owner =       THIS_MODULE,
4563 };
4564
4565 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4566 static struct blkcg_policy blkcg_policy_cfq = {
4567         .pd_size                = sizeof(struct cfq_group),
4568         .cftypes                = cfq_blkcg_files,
4569
4570         .pd_init_fn             = cfq_pd_init,
4571         .pd_offline_fn          = cfq_pd_offline,
4572         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4573 };
4574 #endif
4575
4576 static int __init cfq_init(void)
4577 {
4578         int ret;
4579
4580         /*
4581          * could be 0 on HZ < 1000 setups
4582          */
4583         if (!cfq_slice_async)
4584                 cfq_slice_async = 1;
4585         if (!cfq_slice_idle)
4586                 cfq_slice_idle = 1;
4587
4588 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4589         if (!cfq_group_idle)
4590                 cfq_group_idle = 1;
4591
4592         ret = blkcg_policy_register(&blkcg_policy_cfq);
4593         if (ret)
4594                 return ret;
4595 #else
4596         cfq_group_idle = 0;
4597 #endif
4598
4599         ret = -ENOMEM;
4600         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4601         if (!cfq_pool)
4602                 goto err_pol_unreg;
4603
4604         ret = elv_register(&iosched_cfq);
4605         if (ret)
4606                 goto err_free_pool;
4607
4608         return 0;
4609
4610 err_free_pool:
4611         kmem_cache_destroy(cfq_pool);
4612 err_pol_unreg:
4613 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4614         blkcg_policy_unregister(&blkcg_policy_cfq);
4615 #endif
4616         return ret;
4617 }
4618
4619 static void __exit cfq_exit(void)
4620 {
4621 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4622         blkcg_policy_unregister(&blkcg_policy_cfq);
4623 #endif
4624         elv_unregister(&iosched_cfq);
4625         kmem_cache_destroy(cfq_pool);
4626 }
4627
4628 module_init(cfq_init);
4629 module_exit(cfq_exit);
4630
4631 MODULE_AUTHOR("Jens Axboe");
4632 MODULE_LICENSE("GPL");
4633 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");