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_subsys_state *css,
1611                                     struct cftype *cft, struct seq_file *sf)
1612 {
1613         blkcg_print_blkgs(sf, css_to_blkcg(css), cfqg_prfill_weight_device,
1614                           &blkcg_policy_cfq, 0, false);
1615         return 0;
1616 }
1617
1618 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1619                                           struct blkg_policy_data *pd, int off)
1620 {
1621         struct cfq_group *cfqg = pd_to_cfqg(pd);
1622
1623         if (!cfqg->dev_leaf_weight)
1624                 return 0;
1625         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1626 }
1627
1628 static int cfqg_print_leaf_weight_device(struct cgroup_subsys_state *css,
1629                                          struct cftype *cft,
1630                                          struct seq_file *sf)
1631 {
1632         blkcg_print_blkgs(sf, css_to_blkcg(css), cfqg_prfill_leaf_weight_device,
1633                           &blkcg_policy_cfq, 0, false);
1634         return 0;
1635 }
1636
1637 static int cfq_print_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1638                             struct seq_file *sf)
1639 {
1640         seq_printf(sf, "%u\n", css_to_blkcg(css)->cfq_weight);
1641         return 0;
1642 }
1643
1644 static int cfq_print_leaf_weight(struct cgroup_subsys_state *css,
1645                                  struct cftype *cft, struct seq_file *sf)
1646 {
1647         seq_printf(sf, "%u\n", css_to_blkcg(css)->cfq_leaf_weight);
1648         return 0;
1649 }
1650
1651 static int __cfqg_set_weight_device(struct cgroup_subsys_state *css,
1652                                     struct cftype *cft, const char *buf,
1653                                     bool is_leaf_weight)
1654 {
1655         struct blkcg *blkcg = css_to_blkcg(css);
1656         struct blkg_conf_ctx ctx;
1657         struct cfq_group *cfqg;
1658         int ret;
1659
1660         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1661         if (ret)
1662                 return ret;
1663
1664         ret = -EINVAL;
1665         cfqg = blkg_to_cfqg(ctx.blkg);
1666         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1667                 if (!is_leaf_weight) {
1668                         cfqg->dev_weight = ctx.v;
1669                         cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1670                 } else {
1671                         cfqg->dev_leaf_weight = ctx.v;
1672                         cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1673                 }
1674                 ret = 0;
1675         }
1676
1677         blkg_conf_finish(&ctx);
1678         return ret;
1679 }
1680
1681 static int cfqg_set_weight_device(struct cgroup_subsys_state *css,
1682                                   struct cftype *cft, const char *buf)
1683 {
1684         return __cfqg_set_weight_device(css, cft, buf, false);
1685 }
1686
1687 static int cfqg_set_leaf_weight_device(struct cgroup_subsys_state *css,
1688                                        struct cftype *cft, const char *buf)
1689 {
1690         return __cfqg_set_weight_device(css, cft, buf, true);
1691 }
1692
1693 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1694                             u64 val, bool is_leaf_weight)
1695 {
1696         struct blkcg *blkcg = css_to_blkcg(css);
1697         struct blkcg_gq *blkg;
1698
1699         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1700                 return -EINVAL;
1701
1702         spin_lock_irq(&blkcg->lock);
1703
1704         if (!is_leaf_weight)
1705                 blkcg->cfq_weight = val;
1706         else
1707                 blkcg->cfq_leaf_weight = val;
1708
1709         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1710                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1711
1712                 if (!cfqg)
1713                         continue;
1714
1715                 if (!is_leaf_weight) {
1716                         if (!cfqg->dev_weight)
1717                                 cfqg->new_weight = blkcg->cfq_weight;
1718                 } else {
1719                         if (!cfqg->dev_leaf_weight)
1720                                 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1721                 }
1722         }
1723
1724         spin_unlock_irq(&blkcg->lock);
1725         return 0;
1726 }
1727
1728 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1729                           u64 val)
1730 {
1731         return __cfq_set_weight(css, cft, val, false);
1732 }
1733
1734 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1735                                struct cftype *cft, u64 val)
1736 {
1737         return __cfq_set_weight(css, cft, val, true);
1738 }
1739
1740 static int cfqg_print_stat(struct cgroup_subsys_state *css, struct cftype *cft,
1741                            struct seq_file *sf)
1742 {
1743         struct blkcg *blkcg = css_to_blkcg(css);
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_subsys_state *css,
1751                              struct cftype *cft, struct seq_file *sf)
1752 {
1753         struct blkcg *blkcg = css_to_blkcg(css);
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_subsys_state *css,
1777                                      struct cftype *cft, struct seq_file *sf)
1778 {
1779         struct blkcg *blkcg = css_to_blkcg(css);
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_subsys_state *css,
1787                                        struct cftype *cft, struct seq_file *sf)
1788 {
1789         struct blkcg *blkcg = css_to_blkcg(css);
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_subsys_state *css,
1814                                      struct cftype *cft, struct seq_file *sf)
1815 {
1816         struct blkcg *blkcg = css_to_blkcg(css);
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                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2275
2276         return NULL;
2277 }
2278
2279 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2280 {
2281         struct cfq_data *cfqd = q->elevator->elevator_data;
2282
2283         cfqd->rq_in_driver++;
2284         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2285                                                 cfqd->rq_in_driver);
2286
2287         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2288 }
2289
2290 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2291 {
2292         struct cfq_data *cfqd = q->elevator->elevator_data;
2293
2294         WARN_ON(!cfqd->rq_in_driver);
2295         cfqd->rq_in_driver--;
2296         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2297                                                 cfqd->rq_in_driver);
2298 }
2299
2300 static void cfq_remove_request(struct request *rq)
2301 {
2302         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2303
2304         if (cfqq->next_rq == rq)
2305                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2306
2307         list_del_init(&rq->queuelist);
2308         cfq_del_rq_rb(rq);
2309
2310         cfqq->cfqd->rq_queued--;
2311         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2312         if (rq->cmd_flags & REQ_PRIO) {
2313                 WARN_ON(!cfqq->prio_pending);
2314                 cfqq->prio_pending--;
2315         }
2316 }
2317
2318 static int cfq_merge(struct request_queue *q, struct request **req,
2319                      struct bio *bio)
2320 {
2321         struct cfq_data *cfqd = q->elevator->elevator_data;
2322         struct request *__rq;
2323
2324         __rq = cfq_find_rq_fmerge(cfqd, bio);
2325         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2326                 *req = __rq;
2327                 return ELEVATOR_FRONT_MERGE;
2328         }
2329
2330         return ELEVATOR_NO_MERGE;
2331 }
2332
2333 static void cfq_merged_request(struct request_queue *q, struct request *req,
2334                                int type)
2335 {
2336         if (type == ELEVATOR_FRONT_MERGE) {
2337                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2338
2339                 cfq_reposition_rq_rb(cfqq, req);
2340         }
2341 }
2342
2343 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2344                                 struct bio *bio)
2345 {
2346         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2347 }
2348
2349 static void
2350 cfq_merged_requests(struct request_queue *q, struct request *rq,
2351                     struct request *next)
2352 {
2353         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2354         struct cfq_data *cfqd = q->elevator->elevator_data;
2355
2356         /*
2357          * reposition in fifo if next is older than rq
2358          */
2359         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2360             time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2361             cfqq == RQ_CFQQ(next)) {
2362                 list_move(&rq->queuelist, &next->queuelist);
2363                 rq_set_fifo_time(rq, rq_fifo_time(next));
2364         }
2365
2366         if (cfqq->next_rq == next)
2367                 cfqq->next_rq = rq;
2368         cfq_remove_request(next);
2369         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2370
2371         cfqq = RQ_CFQQ(next);
2372         /*
2373          * all requests of this queue are merged to other queues, delete it
2374          * from the service tree. If it's the active_queue,
2375          * cfq_dispatch_requests() will choose to expire it or do idle
2376          */
2377         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2378             cfqq != cfqd->active_queue)
2379                 cfq_del_cfqq_rr(cfqd, cfqq);
2380 }
2381
2382 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2383                            struct bio *bio)
2384 {
2385         struct cfq_data *cfqd = q->elevator->elevator_data;
2386         struct cfq_io_cq *cic;
2387         struct cfq_queue *cfqq;
2388
2389         /*
2390          * Disallow merge of a sync bio into an async request.
2391          */
2392         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2393                 return false;
2394
2395         /*
2396          * Lookup the cfqq that this bio will be queued with and allow
2397          * merge only if rq is queued there.
2398          */
2399         cic = cfq_cic_lookup(cfqd, current->io_context);
2400         if (!cic)
2401                 return false;
2402
2403         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2404         return cfqq == RQ_CFQQ(rq);
2405 }
2406
2407 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2408 {
2409         del_timer(&cfqd->idle_slice_timer);
2410         cfqg_stats_update_idle_time(cfqq->cfqg);
2411 }
2412
2413 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2414                                    struct cfq_queue *cfqq)
2415 {
2416         if (cfqq) {
2417                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2418                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2419                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2420                 cfqq->slice_start = 0;
2421                 cfqq->dispatch_start = jiffies;
2422                 cfqq->allocated_slice = 0;
2423                 cfqq->slice_end = 0;
2424                 cfqq->slice_dispatch = 0;
2425                 cfqq->nr_sectors = 0;
2426
2427                 cfq_clear_cfqq_wait_request(cfqq);
2428                 cfq_clear_cfqq_must_dispatch(cfqq);
2429                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2430                 cfq_clear_cfqq_fifo_expire(cfqq);
2431                 cfq_mark_cfqq_slice_new(cfqq);
2432
2433                 cfq_del_timer(cfqd, cfqq);
2434         }
2435
2436         cfqd->active_queue = cfqq;
2437 }
2438
2439 /*
2440  * current cfqq expired its slice (or was too idle), select new one
2441  */
2442 static void
2443 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2444                     bool timed_out)
2445 {
2446         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2447
2448         if (cfq_cfqq_wait_request(cfqq))
2449                 cfq_del_timer(cfqd, cfqq);
2450
2451         cfq_clear_cfqq_wait_request(cfqq);
2452         cfq_clear_cfqq_wait_busy(cfqq);
2453
2454         /*
2455          * If this cfqq is shared between multiple processes, check to
2456          * make sure that those processes are still issuing I/Os within
2457          * the mean seek distance.  If not, it may be time to break the
2458          * queues apart again.
2459          */
2460         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2461                 cfq_mark_cfqq_split_coop(cfqq);
2462
2463         /*
2464          * store what was left of this slice, if the queue idled/timed out
2465          */
2466         if (timed_out) {
2467                 if (cfq_cfqq_slice_new(cfqq))
2468                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2469                 else
2470                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2471                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2472         }
2473
2474         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2475
2476         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2477                 cfq_del_cfqq_rr(cfqd, cfqq);
2478
2479         cfq_resort_rr_list(cfqd, cfqq);
2480
2481         if (cfqq == cfqd->active_queue)
2482                 cfqd->active_queue = NULL;
2483
2484         if (cfqd->active_cic) {
2485                 put_io_context(cfqd->active_cic->icq.ioc);
2486                 cfqd->active_cic = NULL;
2487         }
2488 }
2489
2490 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2491 {
2492         struct cfq_queue *cfqq = cfqd->active_queue;
2493
2494         if (cfqq)
2495                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2496 }
2497
2498 /*
2499  * Get next queue for service. Unless we have a queue preemption,
2500  * we'll simply select the first cfqq in the service tree.
2501  */
2502 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2503 {
2504         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2505                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2506
2507         if (!cfqd->rq_queued)
2508                 return NULL;
2509
2510         /* There is nothing to dispatch */
2511         if (!st)
2512                 return NULL;
2513         if (RB_EMPTY_ROOT(&st->rb))
2514                 return NULL;
2515         return cfq_rb_first(st);
2516 }
2517
2518 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2519 {
2520         struct cfq_group *cfqg;
2521         struct cfq_queue *cfqq;
2522         int i, j;
2523         struct cfq_rb_root *st;
2524
2525         if (!cfqd->rq_queued)
2526                 return NULL;
2527
2528         cfqg = cfq_get_next_cfqg(cfqd);
2529         if (!cfqg)
2530                 return NULL;
2531
2532         for_each_cfqg_st(cfqg, i, j, st)
2533                 if ((cfqq = cfq_rb_first(st)) != NULL)
2534                         return cfqq;
2535         return NULL;
2536 }
2537
2538 /*
2539  * Get and set a new active queue for service.
2540  */
2541 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2542                                               struct cfq_queue *cfqq)
2543 {
2544         if (!cfqq)
2545                 cfqq = cfq_get_next_queue(cfqd);
2546
2547         __cfq_set_active_queue(cfqd, cfqq);
2548         return cfqq;
2549 }
2550
2551 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2552                                           struct request *rq)
2553 {
2554         if (blk_rq_pos(rq) >= cfqd->last_position)
2555                 return blk_rq_pos(rq) - cfqd->last_position;
2556         else
2557                 return cfqd->last_position - blk_rq_pos(rq);
2558 }
2559
2560 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2561                                struct request *rq)
2562 {
2563         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2564 }
2565
2566 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2567                                     struct cfq_queue *cur_cfqq)
2568 {
2569         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2570         struct rb_node *parent, *node;
2571         struct cfq_queue *__cfqq;
2572         sector_t sector = cfqd->last_position;
2573
2574         if (RB_EMPTY_ROOT(root))
2575                 return NULL;
2576
2577         /*
2578          * First, if we find a request starting at the end of the last
2579          * request, choose it.
2580          */
2581         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2582         if (__cfqq)
2583                 return __cfqq;
2584
2585         /*
2586          * If the exact sector wasn't found, the parent of the NULL leaf
2587          * will contain the closest sector.
2588          */
2589         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2590         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2591                 return __cfqq;
2592
2593         if (blk_rq_pos(__cfqq->next_rq) < sector)
2594                 node = rb_next(&__cfqq->p_node);
2595         else
2596                 node = rb_prev(&__cfqq->p_node);
2597         if (!node)
2598                 return NULL;
2599
2600         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2601         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2602                 return __cfqq;
2603
2604         return NULL;
2605 }
2606
2607 /*
2608  * cfqd - obvious
2609  * cur_cfqq - passed in so that we don't decide that the current queue is
2610  *            closely cooperating with itself.
2611  *
2612  * So, basically we're assuming that that cur_cfqq has dispatched at least
2613  * one request, and that cfqd->last_position reflects a position on the disk
2614  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2615  * assumption.
2616  */
2617 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2618                                               struct cfq_queue *cur_cfqq)
2619 {
2620         struct cfq_queue *cfqq;
2621
2622         if (cfq_class_idle(cur_cfqq))
2623                 return NULL;
2624         if (!cfq_cfqq_sync(cur_cfqq))
2625                 return NULL;
2626         if (CFQQ_SEEKY(cur_cfqq))
2627                 return NULL;
2628
2629         /*
2630          * Don't search priority tree if it's the only queue in the group.
2631          */
2632         if (cur_cfqq->cfqg->nr_cfqq == 1)
2633                 return NULL;
2634
2635         /*
2636          * We should notice if some of the queues are cooperating, eg
2637          * working closely on the same area of the disk. In that case,
2638          * we can group them together and don't waste time idling.
2639          */
2640         cfqq = cfqq_close(cfqd, cur_cfqq);
2641         if (!cfqq)
2642                 return NULL;
2643
2644         /* If new queue belongs to different cfq_group, don't choose it */
2645         if (cur_cfqq->cfqg != cfqq->cfqg)
2646                 return NULL;
2647
2648         /*
2649          * It only makes sense to merge sync queues.
2650          */
2651         if (!cfq_cfqq_sync(cfqq))
2652                 return NULL;
2653         if (CFQQ_SEEKY(cfqq))
2654                 return NULL;
2655
2656         /*
2657          * Do not merge queues of different priority classes
2658          */
2659         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2660                 return NULL;
2661
2662         return cfqq;
2663 }
2664
2665 /*
2666  * Determine whether we should enforce idle window for this queue.
2667  */
2668
2669 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2670 {
2671         enum wl_class_t wl_class = cfqq_class(cfqq);
2672         struct cfq_rb_root *st = cfqq->service_tree;
2673
2674         BUG_ON(!st);
2675         BUG_ON(!st->count);
2676
2677         if (!cfqd->cfq_slice_idle)
2678                 return false;
2679
2680         /* We never do for idle class queues. */
2681         if (wl_class == IDLE_WORKLOAD)
2682                 return false;
2683
2684         /* We do for queues that were marked with idle window flag. */
2685         if (cfq_cfqq_idle_window(cfqq) &&
2686            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2687                 return true;
2688
2689         /*
2690          * Otherwise, we do only if they are the last ones
2691          * in their service tree.
2692          */
2693         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2694            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2695                 return true;
2696         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2697         return false;
2698 }
2699
2700 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2701 {
2702         struct cfq_queue *cfqq = cfqd->active_queue;
2703         struct cfq_io_cq *cic;
2704         unsigned long sl, group_idle = 0;
2705
2706         /*
2707          * SSD device without seek penalty, disable idling. But only do so
2708          * for devices that support queuing, otherwise we still have a problem
2709          * with sync vs async workloads.
2710          */
2711         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2712                 return;
2713
2714         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2715         WARN_ON(cfq_cfqq_slice_new(cfqq));
2716
2717         /*
2718          * idle is disabled, either manually or by past process history
2719          */
2720         if (!cfq_should_idle(cfqd, cfqq)) {
2721                 /* no queue idling. Check for group idling */
2722                 if (cfqd->cfq_group_idle)
2723                         group_idle = cfqd->cfq_group_idle;
2724                 else
2725                         return;
2726         }
2727
2728         /*
2729          * still active requests from this queue, don't idle
2730          */
2731         if (cfqq->dispatched)
2732                 return;
2733
2734         /*
2735          * task has exited, don't wait
2736          */
2737         cic = cfqd->active_cic;
2738         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2739                 return;
2740
2741         /*
2742          * If our average think time is larger than the remaining time
2743          * slice, then don't idle. This avoids overrunning the allotted
2744          * time slice.
2745          */
2746         if (sample_valid(cic->ttime.ttime_samples) &&
2747             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2748                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2749                              cic->ttime.ttime_mean);
2750                 return;
2751         }
2752
2753         /* There are other queues in the group, don't do group idle */
2754         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2755                 return;
2756
2757         cfq_mark_cfqq_wait_request(cfqq);
2758
2759         if (group_idle)
2760                 sl = cfqd->cfq_group_idle;
2761         else
2762                 sl = cfqd->cfq_slice_idle;
2763
2764         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2765         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2766         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2767                         group_idle ? 1 : 0);
2768 }
2769
2770 /*
2771  * Move request from internal lists to the request queue dispatch list.
2772  */
2773 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2774 {
2775         struct cfq_data *cfqd = q->elevator->elevator_data;
2776         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2777
2778         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2779
2780         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2781         cfq_remove_request(rq);
2782         cfqq->dispatched++;
2783         (RQ_CFQG(rq))->dispatched++;
2784         elv_dispatch_sort(q, rq);
2785
2786         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2787         cfqq->nr_sectors += blk_rq_sectors(rq);
2788         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2789 }
2790
2791 /*
2792  * return expired entry, or NULL to just start from scratch in rbtree
2793  */
2794 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2795 {
2796         struct request *rq = NULL;
2797
2798         if (cfq_cfqq_fifo_expire(cfqq))
2799                 return NULL;
2800
2801         cfq_mark_cfqq_fifo_expire(cfqq);
2802
2803         if (list_empty(&cfqq->fifo))
2804                 return NULL;
2805
2806         rq = rq_entry_fifo(cfqq->fifo.next);
2807         if (time_before(jiffies, rq_fifo_time(rq)))
2808                 rq = NULL;
2809
2810         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2811         return rq;
2812 }
2813
2814 static inline int
2815 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2816 {
2817         const int base_rq = cfqd->cfq_slice_async_rq;
2818
2819         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2820
2821         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2822 }
2823
2824 /*
2825  * Must be called with the queue_lock held.
2826  */
2827 static int cfqq_process_refs(struct cfq_queue *cfqq)
2828 {
2829         int process_refs, io_refs;
2830
2831         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2832         process_refs = cfqq->ref - io_refs;
2833         BUG_ON(process_refs < 0);
2834         return process_refs;
2835 }
2836
2837 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2838 {
2839         int process_refs, new_process_refs;
2840         struct cfq_queue *__cfqq;
2841
2842         /*
2843          * If there are no process references on the new_cfqq, then it is
2844          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2845          * chain may have dropped their last reference (not just their
2846          * last process reference).
2847          */
2848         if (!cfqq_process_refs(new_cfqq))
2849                 return;
2850
2851         /* Avoid a circular list and skip interim queue merges */
2852         while ((__cfqq = new_cfqq->new_cfqq)) {
2853                 if (__cfqq == cfqq)
2854                         return;
2855                 new_cfqq = __cfqq;
2856         }
2857
2858         process_refs = cfqq_process_refs(cfqq);
2859         new_process_refs = cfqq_process_refs(new_cfqq);
2860         /*
2861          * If the process for the cfqq has gone away, there is no
2862          * sense in merging the queues.
2863          */
2864         if (process_refs == 0 || new_process_refs == 0)
2865                 return;
2866
2867         /*
2868          * Merge in the direction of the lesser amount of work.
2869          */
2870         if (new_process_refs >= process_refs) {
2871                 cfqq->new_cfqq = new_cfqq;
2872                 new_cfqq->ref += process_refs;
2873         } else {
2874                 new_cfqq->new_cfqq = cfqq;
2875                 cfqq->ref += new_process_refs;
2876         }
2877 }
2878
2879 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2880                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2881 {
2882         struct cfq_queue *queue;
2883         int i;
2884         bool key_valid = false;
2885         unsigned long lowest_key = 0;
2886         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2887
2888         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2889                 /* select the one with lowest rb_key */
2890                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2891                 if (queue &&
2892                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2893                         lowest_key = queue->rb_key;
2894                         cur_best = i;
2895                         key_valid = true;
2896                 }
2897         }
2898
2899         return cur_best;
2900 }
2901
2902 static void
2903 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2904 {
2905         unsigned slice;
2906         unsigned count;
2907         struct cfq_rb_root *st;
2908         unsigned group_slice;
2909         enum wl_class_t original_class = cfqd->serving_wl_class;
2910
2911         /* Choose next priority. RT > BE > IDLE */
2912         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2913                 cfqd->serving_wl_class = RT_WORKLOAD;
2914         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2915                 cfqd->serving_wl_class = BE_WORKLOAD;
2916         else {
2917                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2918                 cfqd->workload_expires = jiffies + 1;
2919                 return;
2920         }
2921
2922         if (original_class != cfqd->serving_wl_class)
2923                 goto new_workload;
2924
2925         /*
2926          * For RT and BE, we have to choose also the type
2927          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2928          * expiration time
2929          */
2930         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2931         count = st->count;
2932
2933         /*
2934          * check workload expiration, and that we still have other queues ready
2935          */
2936         if (count && !time_after(jiffies, cfqd->workload_expires))
2937                 return;
2938
2939 new_workload:
2940         /* otherwise select new workload type */
2941         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2942                                         cfqd->serving_wl_class);
2943         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2944         count = st->count;
2945
2946         /*
2947          * the workload slice is computed as a fraction of target latency
2948          * proportional to the number of queues in that workload, over
2949          * all the queues in the same priority class
2950          */
2951         group_slice = cfq_group_slice(cfqd, cfqg);
2952
2953         slice = group_slice * count /
2954                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2955                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2956                                         cfqg));
2957
2958         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2959                 unsigned int tmp;
2960
2961                 /*
2962                  * Async queues are currently system wide. Just taking
2963                  * proportion of queues with-in same group will lead to higher
2964                  * async ratio system wide as generally root group is going
2965                  * to have higher weight. A more accurate thing would be to
2966                  * calculate system wide asnc/sync ratio.
2967                  */
2968                 tmp = cfqd->cfq_target_latency *
2969                         cfqg_busy_async_queues(cfqd, cfqg);
2970                 tmp = tmp/cfqd->busy_queues;
2971                 slice = min_t(unsigned, slice, tmp);
2972
2973                 /* async workload slice is scaled down according to
2974                  * the sync/async slice ratio. */
2975                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2976         } else
2977                 /* sync workload slice is at least 2 * cfq_slice_idle */
2978                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2979
2980         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2981         cfq_log(cfqd, "workload slice:%d", slice);
2982         cfqd->workload_expires = jiffies + slice;
2983 }
2984
2985 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2986 {
2987         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2988         struct cfq_group *cfqg;
2989
2990         if (RB_EMPTY_ROOT(&st->rb))
2991                 return NULL;
2992         cfqg = cfq_rb_first_group(st);
2993         update_min_vdisktime(st);
2994         return cfqg;
2995 }
2996
2997 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2998 {
2999         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3000
3001         cfqd->serving_group = cfqg;
3002
3003         /* Restore the workload type data */
3004         if (cfqg->saved_wl_slice) {
3005                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3006                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3007                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3008         } else
3009                 cfqd->workload_expires = jiffies - 1;
3010
3011         choose_wl_class_and_type(cfqd, cfqg);
3012 }
3013
3014 /*
3015  * Select a queue for service. If we have a current active queue,
3016  * check whether to continue servicing it, or retrieve and set a new one.
3017  */
3018 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3019 {
3020         struct cfq_queue *cfqq, *new_cfqq = NULL;
3021
3022         cfqq = cfqd->active_queue;
3023         if (!cfqq)
3024                 goto new_queue;
3025
3026         if (!cfqd->rq_queued)
3027                 return NULL;
3028
3029         /*
3030          * We were waiting for group to get backlogged. Expire the queue
3031          */
3032         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3033                 goto expire;
3034
3035         /*
3036          * The active queue has run out of time, expire it and select new.
3037          */
3038         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3039                 /*
3040                  * If slice had not expired at the completion of last request
3041                  * we might not have turned on wait_busy flag. Don't expire
3042                  * the queue yet. Allow the group to get backlogged.
3043                  *
3044                  * The very fact that we have used the slice, that means we
3045                  * have been idling all along on this queue and it should be
3046                  * ok to wait for this request to complete.
3047                  */
3048                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3049                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3050                         cfqq = NULL;
3051                         goto keep_queue;
3052                 } else
3053                         goto check_group_idle;
3054         }
3055
3056         /*
3057          * The active queue has requests and isn't expired, allow it to
3058          * dispatch.
3059          */
3060         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3061                 goto keep_queue;
3062
3063         /*
3064          * If another queue has a request waiting within our mean seek
3065          * distance, let it run.  The expire code will check for close
3066          * cooperators and put the close queue at the front of the service
3067          * tree.  If possible, merge the expiring queue with the new cfqq.
3068          */
3069         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3070         if (new_cfqq) {
3071                 if (!cfqq->new_cfqq)
3072                         cfq_setup_merge(cfqq, new_cfqq);
3073                 goto expire;
3074         }
3075
3076         /*
3077          * No requests pending. If the active queue still has requests in
3078          * flight or is idling for a new request, allow either of these
3079          * conditions to happen (or time out) before selecting a new queue.
3080          */
3081         if (timer_pending(&cfqd->idle_slice_timer)) {
3082                 cfqq = NULL;
3083                 goto keep_queue;
3084         }
3085
3086         /*
3087          * This is a deep seek queue, but the device is much faster than
3088          * the queue can deliver, don't idle
3089          **/
3090         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3091             (cfq_cfqq_slice_new(cfqq) ||
3092             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3093                 cfq_clear_cfqq_deep(cfqq);
3094                 cfq_clear_cfqq_idle_window(cfqq);
3095         }
3096
3097         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3098                 cfqq = NULL;
3099                 goto keep_queue;
3100         }
3101
3102         /*
3103          * If group idle is enabled and there are requests dispatched from
3104          * this group, wait for requests to complete.
3105          */
3106 check_group_idle:
3107         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3108             cfqq->cfqg->dispatched &&
3109             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3110                 cfqq = NULL;
3111                 goto keep_queue;
3112         }
3113
3114 expire:
3115         cfq_slice_expired(cfqd, 0);
3116 new_queue:
3117         /*
3118          * Current queue expired. Check if we have to switch to a new
3119          * service tree
3120          */
3121         if (!new_cfqq)
3122                 cfq_choose_cfqg(cfqd);
3123
3124         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3125 keep_queue:
3126         return cfqq;
3127 }
3128
3129 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3130 {
3131         int dispatched = 0;
3132
3133         while (cfqq->next_rq) {
3134                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3135                 dispatched++;
3136         }
3137
3138         BUG_ON(!list_empty(&cfqq->fifo));
3139
3140         /* By default cfqq is not expired if it is empty. Do it explicitly */
3141         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3142         return dispatched;
3143 }
3144
3145 /*
3146  * Drain our current requests. Used for barriers and when switching
3147  * io schedulers on-the-fly.
3148  */
3149 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3150 {
3151         struct cfq_queue *cfqq;
3152         int dispatched = 0;
3153
3154         /* Expire the timeslice of the current active queue first */
3155         cfq_slice_expired(cfqd, 0);
3156         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3157                 __cfq_set_active_queue(cfqd, cfqq);
3158                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3159         }
3160
3161         BUG_ON(cfqd->busy_queues);
3162
3163         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3164         return dispatched;
3165 }
3166
3167 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3168         struct cfq_queue *cfqq)
3169 {
3170         /* the queue hasn't finished any request, can't estimate */
3171         if (cfq_cfqq_slice_new(cfqq))
3172                 return true;
3173         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3174                 cfqq->slice_end))
3175                 return true;
3176
3177         return false;
3178 }
3179
3180 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3181 {
3182         unsigned int max_dispatch;
3183
3184         /*
3185          * Drain async requests before we start sync IO
3186          */
3187         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3188                 return false;
3189
3190         /*
3191          * If this is an async queue and we have sync IO in flight, let it wait
3192          */
3193         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3194                 return false;
3195
3196         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3197         if (cfq_class_idle(cfqq))
3198                 max_dispatch = 1;
3199
3200         /*
3201          * Does this cfqq already have too much IO in flight?
3202          */
3203         if (cfqq->dispatched >= max_dispatch) {
3204                 bool promote_sync = false;
3205                 /*
3206                  * idle queue must always only have a single IO in flight
3207                  */
3208                 if (cfq_class_idle(cfqq))
3209                         return false;
3210
3211                 /*
3212                  * If there is only one sync queue
3213                  * we can ignore async queue here and give the sync
3214                  * queue no dispatch limit. The reason is a sync queue can
3215                  * preempt async queue, limiting the sync queue doesn't make
3216                  * sense. This is useful for aiostress test.
3217                  */
3218                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3219                         promote_sync = true;
3220
3221                 /*
3222                  * We have other queues, don't allow more IO from this one
3223                  */
3224                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3225                                 !promote_sync)
3226                         return false;
3227
3228                 /*
3229                  * Sole queue user, no limit
3230                  */
3231                 if (cfqd->busy_queues == 1 || promote_sync)
3232                         max_dispatch = -1;
3233                 else
3234                         /*
3235                          * Normally we start throttling cfqq when cfq_quantum/2
3236                          * requests have been dispatched. But we can drive
3237                          * deeper queue depths at the beginning of slice
3238                          * subjected to upper limit of cfq_quantum.
3239                          * */
3240                         max_dispatch = cfqd->cfq_quantum;
3241         }
3242
3243         /*
3244          * Async queues must wait a bit before being allowed dispatch.
3245          * We also ramp up the dispatch depth gradually for async IO,
3246          * based on the last sync IO we serviced
3247          */
3248         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3249                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3250                 unsigned int depth;
3251
3252                 depth = last_sync / cfqd->cfq_slice[1];
3253                 if (!depth && !cfqq->dispatched)
3254                         depth = 1;
3255                 if (depth < max_dispatch)
3256                         max_dispatch = depth;
3257         }
3258
3259         /*
3260          * If we're below the current max, allow a dispatch
3261          */
3262         return cfqq->dispatched < max_dispatch;
3263 }
3264
3265 /*
3266  * Dispatch a request from cfqq, moving them to the request queue
3267  * dispatch list.
3268  */
3269 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3270 {
3271         struct request *rq;
3272
3273         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3274
3275         if (!cfq_may_dispatch(cfqd, cfqq))
3276                 return false;
3277
3278         /*
3279          * follow expired path, else get first next available
3280          */
3281         rq = cfq_check_fifo(cfqq);
3282         if (!rq)
3283                 rq = cfqq->next_rq;
3284
3285         /*
3286          * insert request into driver dispatch list
3287          */
3288         cfq_dispatch_insert(cfqd->queue, rq);
3289
3290         if (!cfqd->active_cic) {
3291                 struct cfq_io_cq *cic = RQ_CIC(rq);
3292
3293                 atomic_long_inc(&cic->icq.ioc->refcount);
3294                 cfqd->active_cic = cic;
3295         }
3296
3297         return true;
3298 }
3299
3300 /*
3301  * Find the cfqq that we need to service and move a request from that to the
3302  * dispatch list
3303  */
3304 static int cfq_dispatch_requests(struct request_queue *q, int force)
3305 {
3306         struct cfq_data *cfqd = q->elevator->elevator_data;
3307         struct cfq_queue *cfqq;
3308
3309         if (!cfqd->busy_queues)
3310                 return 0;
3311
3312         if (unlikely(force))
3313                 return cfq_forced_dispatch(cfqd);
3314
3315         cfqq = cfq_select_queue(cfqd);
3316         if (!cfqq)
3317                 return 0;
3318
3319         /*
3320          * Dispatch a request from this cfqq, if it is allowed
3321          */
3322         if (!cfq_dispatch_request(cfqd, cfqq))
3323                 return 0;
3324
3325         cfqq->slice_dispatch++;
3326         cfq_clear_cfqq_must_dispatch(cfqq);
3327
3328         /*
3329          * expire an async queue immediately if it has used up its slice. idle
3330          * queue always expire after 1 dispatch round.
3331          */
3332         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3333             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3334             cfq_class_idle(cfqq))) {
3335                 cfqq->slice_end = jiffies + 1;
3336                 cfq_slice_expired(cfqd, 0);
3337         }
3338
3339         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3340         return 1;
3341 }
3342
3343 /*
3344  * task holds one reference to the queue, dropped when task exits. each rq
3345  * in-flight on this queue also holds a reference, dropped when rq is freed.
3346  *
3347  * Each cfq queue took a reference on the parent group. Drop it now.
3348  * queue lock must be held here.
3349  */
3350 static void cfq_put_queue(struct cfq_queue *cfqq)
3351 {
3352         struct cfq_data *cfqd = cfqq->cfqd;
3353         struct cfq_group *cfqg;
3354
3355         BUG_ON(cfqq->ref <= 0);
3356
3357         cfqq->ref--;
3358         if (cfqq->ref)
3359                 return;
3360
3361         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3362         BUG_ON(rb_first(&cfqq->sort_list));
3363         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3364         cfqg = cfqq->cfqg;
3365
3366         if (unlikely(cfqd->active_queue == cfqq)) {
3367                 __cfq_slice_expired(cfqd, cfqq, 0);
3368                 cfq_schedule_dispatch(cfqd);
3369         }
3370
3371         BUG_ON(cfq_cfqq_on_rr(cfqq));
3372         kmem_cache_free(cfq_pool, cfqq);
3373         cfqg_put(cfqg);
3374 }
3375
3376 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3377 {
3378         struct cfq_queue *__cfqq, *next;
3379
3380         /*
3381          * If this queue was scheduled to merge with another queue, be
3382          * sure to drop the reference taken on that queue (and others in
3383          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3384          */
3385         __cfqq = cfqq->new_cfqq;
3386         while (__cfqq) {
3387                 if (__cfqq == cfqq) {
3388                         WARN(1, "cfqq->new_cfqq loop detected\n");
3389                         break;
3390                 }
3391                 next = __cfqq->new_cfqq;
3392                 cfq_put_queue(__cfqq);
3393                 __cfqq = next;
3394         }
3395 }
3396
3397 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3398 {
3399         if (unlikely(cfqq == cfqd->active_queue)) {
3400                 __cfq_slice_expired(cfqd, cfqq, 0);
3401                 cfq_schedule_dispatch(cfqd);
3402         }
3403
3404         cfq_put_cooperator(cfqq);
3405
3406         cfq_put_queue(cfqq);
3407 }
3408
3409 static void cfq_init_icq(struct io_cq *icq)
3410 {
3411         struct cfq_io_cq *cic = icq_to_cic(icq);
3412
3413         cic->ttime.last_end_request = jiffies;
3414 }
3415
3416 static void cfq_exit_icq(struct io_cq *icq)
3417 {
3418         struct cfq_io_cq *cic = icq_to_cic(icq);
3419         struct cfq_data *cfqd = cic_to_cfqd(cic);
3420
3421         if (cic->cfqq[BLK_RW_ASYNC]) {
3422                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3423                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3424         }
3425
3426         if (cic->cfqq[BLK_RW_SYNC]) {
3427                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3428                 cic->cfqq[BLK_RW_SYNC] = NULL;
3429         }
3430 }
3431
3432 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3433 {
3434         struct task_struct *tsk = current;
3435         int ioprio_class;
3436
3437         if (!cfq_cfqq_prio_changed(cfqq))
3438                 return;
3439
3440         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3441         switch (ioprio_class) {
3442         default:
3443                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3444         case IOPRIO_CLASS_NONE:
3445                 /*
3446                  * no prio set, inherit CPU scheduling settings
3447                  */
3448                 cfqq->ioprio = task_nice_ioprio(tsk);
3449                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3450                 break;
3451         case IOPRIO_CLASS_RT:
3452                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3453                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3454                 break;
3455         case IOPRIO_CLASS_BE:
3456                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3457                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3458                 break;
3459         case IOPRIO_CLASS_IDLE:
3460                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3461                 cfqq->ioprio = 7;
3462                 cfq_clear_cfqq_idle_window(cfqq);
3463                 break;
3464         }
3465
3466         /*
3467          * keep track of original prio settings in case we have to temporarily
3468          * elevate the priority of this queue
3469          */
3470         cfqq->org_ioprio = cfqq->ioprio;
3471         cfq_clear_cfqq_prio_changed(cfqq);
3472 }
3473
3474 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3475 {
3476         int ioprio = cic->icq.ioc->ioprio;
3477         struct cfq_data *cfqd = cic_to_cfqd(cic);
3478         struct cfq_queue *cfqq;
3479
3480         /*
3481          * Check whether ioprio has changed.  The condition may trigger
3482          * spuriously on a newly created cic but there's no harm.
3483          */
3484         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3485                 return;
3486
3487         cfqq = cic->cfqq[BLK_RW_ASYNC];
3488         if (cfqq) {
3489                 struct cfq_queue *new_cfqq;
3490                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3491                                          GFP_ATOMIC);
3492                 if (new_cfqq) {
3493                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3494                         cfq_put_queue(cfqq);
3495                 }
3496         }
3497
3498         cfqq = cic->cfqq[BLK_RW_SYNC];
3499         if (cfqq)
3500                 cfq_mark_cfqq_prio_changed(cfqq);
3501
3502         cic->ioprio = ioprio;
3503 }
3504
3505 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3506                           pid_t pid, bool is_sync)
3507 {
3508         RB_CLEAR_NODE(&cfqq->rb_node);
3509         RB_CLEAR_NODE(&cfqq->p_node);
3510         INIT_LIST_HEAD(&cfqq->fifo);
3511
3512         cfqq->ref = 0;
3513         cfqq->cfqd = cfqd;
3514
3515         cfq_mark_cfqq_prio_changed(cfqq);
3516
3517         if (is_sync) {
3518                 if (!cfq_class_idle(cfqq))
3519                         cfq_mark_cfqq_idle_window(cfqq);
3520                 cfq_mark_cfqq_sync(cfqq);
3521         }
3522         cfqq->pid = pid;
3523 }
3524
3525 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3526 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3527 {
3528         struct cfq_data *cfqd = cic_to_cfqd(cic);
3529         struct cfq_queue *sync_cfqq;
3530         uint64_t id;
3531
3532         rcu_read_lock();
3533         id = bio_blkcg(bio)->id;
3534         rcu_read_unlock();
3535
3536         /*
3537          * Check whether blkcg has changed.  The condition may trigger
3538          * spuriously on a newly created cic but there's no harm.
3539          */
3540         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3541                 return;
3542
3543         sync_cfqq = cic_to_cfqq(cic, 1);
3544         if (sync_cfqq) {
3545                 /*
3546                  * Drop reference to sync queue. A new sync queue will be
3547                  * assigned in new group upon arrival of a fresh request.
3548                  */
3549                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3550                 cic_set_cfqq(cic, NULL, 1);
3551                 cfq_put_queue(sync_cfqq);
3552         }
3553
3554         cic->blkcg_id = id;
3555 }
3556 #else
3557 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3558 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3559
3560 static struct cfq_queue *
3561 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3562                      struct bio *bio, gfp_t gfp_mask)
3563 {
3564         struct blkcg *blkcg;
3565         struct cfq_queue *cfqq, *new_cfqq = NULL;
3566         struct cfq_group *cfqg;
3567
3568 retry:
3569         rcu_read_lock();
3570
3571         blkcg = bio_blkcg(bio);
3572         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3573         cfqq = cic_to_cfqq(cic, is_sync);
3574
3575         /*
3576          * Always try a new alloc if we fell back to the OOM cfqq
3577          * originally, since it should just be a temporary situation.
3578          */
3579         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3580                 cfqq = NULL;
3581                 if (new_cfqq) {
3582                         cfqq = new_cfqq;
3583                         new_cfqq = NULL;
3584                 } else if (gfp_mask & __GFP_WAIT) {
3585                         rcu_read_unlock();
3586                         spin_unlock_irq(cfqd->queue->queue_lock);
3587                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3588                                         gfp_mask | __GFP_ZERO,
3589                                         cfqd->queue->node);
3590                         spin_lock_irq(cfqd->queue->queue_lock);
3591                         if (new_cfqq)
3592                                 goto retry;
3593                         else
3594                                 return &cfqd->oom_cfqq;
3595                 } else {
3596                         cfqq = kmem_cache_alloc_node(cfq_pool,
3597                                         gfp_mask | __GFP_ZERO,
3598                                         cfqd->queue->node);
3599                 }
3600
3601                 if (cfqq) {
3602                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3603                         cfq_init_prio_data(cfqq, cic);
3604                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3605                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3606                 } else
3607                         cfqq = &cfqd->oom_cfqq;
3608         }
3609
3610         if (new_cfqq)
3611                 kmem_cache_free(cfq_pool, new_cfqq);
3612
3613         rcu_read_unlock();
3614         return cfqq;
3615 }
3616
3617 static struct cfq_queue **
3618 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3619 {
3620         switch (ioprio_class) {
3621         case IOPRIO_CLASS_RT:
3622                 return &cfqd->async_cfqq[0][ioprio];
3623         case IOPRIO_CLASS_NONE:
3624                 ioprio = IOPRIO_NORM;
3625                 /* fall through */
3626         case IOPRIO_CLASS_BE:
3627                 return &cfqd->async_cfqq[1][ioprio];
3628         case IOPRIO_CLASS_IDLE:
3629                 return &cfqd->async_idle_cfqq;
3630         default:
3631                 BUG();
3632         }
3633 }
3634
3635 static struct cfq_queue *
3636 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3637               struct bio *bio, gfp_t gfp_mask)
3638 {
3639         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3640         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3641         struct cfq_queue **async_cfqq = NULL;
3642         struct cfq_queue *cfqq = NULL;
3643
3644         if (!is_sync) {
3645                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3646                 cfqq = *async_cfqq;
3647         }
3648
3649         if (!cfqq)
3650                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3651
3652         /*
3653          * pin the queue now that it's allocated, scheduler exit will prune it
3654          */
3655         if (!is_sync && !(*async_cfqq)) {
3656                 cfqq->ref++;
3657                 *async_cfqq = cfqq;
3658         }
3659
3660         cfqq->ref++;
3661         return cfqq;
3662 }
3663
3664 static void
3665 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3666 {
3667         unsigned long elapsed = jiffies - ttime->last_end_request;
3668         elapsed = min(elapsed, 2UL * slice_idle);
3669
3670         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3671         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3672         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3673 }
3674
3675 static void
3676 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3677                         struct cfq_io_cq *cic)
3678 {
3679         if (cfq_cfqq_sync(cfqq)) {
3680                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3681                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3682                         cfqd->cfq_slice_idle);
3683         }
3684 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3685         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3686 #endif
3687 }
3688
3689 static void
3690 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3691                        struct request *rq)
3692 {
3693         sector_t sdist = 0;
3694         sector_t n_sec = blk_rq_sectors(rq);
3695         if (cfqq->last_request_pos) {
3696                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3697                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3698                 else
3699                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3700         }
3701
3702         cfqq->seek_history <<= 1;
3703         if (blk_queue_nonrot(cfqd->queue))
3704                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3705         else
3706                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3707 }
3708
3709 /*
3710  * Disable idle window if the process thinks too long or seeks so much that
3711  * it doesn't matter
3712  */
3713 static void
3714 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3715                        struct cfq_io_cq *cic)
3716 {
3717         int old_idle, enable_idle;
3718
3719         /*
3720          * Don't idle for async or idle io prio class
3721          */
3722         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3723                 return;
3724
3725         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3726
3727         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3728                 cfq_mark_cfqq_deep(cfqq);
3729
3730         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3731                 enable_idle = 0;
3732         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3733                  !cfqd->cfq_slice_idle ||
3734                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3735                 enable_idle = 0;
3736         else if (sample_valid(cic->ttime.ttime_samples)) {
3737                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3738                         enable_idle = 0;
3739                 else
3740                         enable_idle = 1;
3741         }
3742
3743         if (old_idle != enable_idle) {
3744                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3745                 if (enable_idle)
3746                         cfq_mark_cfqq_idle_window(cfqq);
3747                 else
3748                         cfq_clear_cfqq_idle_window(cfqq);
3749         }
3750 }
3751
3752 /*
3753  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3754  * no or if we aren't sure, a 1 will cause a preempt.
3755  */
3756 static bool
3757 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3758                    struct request *rq)
3759 {
3760         struct cfq_queue *cfqq;
3761
3762         cfqq = cfqd->active_queue;
3763         if (!cfqq)
3764                 return false;
3765
3766         if (cfq_class_idle(new_cfqq))
3767                 return false;
3768
3769         if (cfq_class_idle(cfqq))
3770                 return true;
3771
3772         /*
3773          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3774          */
3775         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3776                 return false;
3777
3778         /*
3779          * if the new request is sync, but the currently running queue is
3780          * not, let the sync request have priority.
3781          */
3782         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3783                 return true;
3784
3785         if (new_cfqq->cfqg != cfqq->cfqg)
3786                 return false;
3787
3788         if (cfq_slice_used(cfqq))
3789                 return true;
3790
3791         /* Allow preemption only if we are idling on sync-noidle tree */
3792         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3793             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3794             new_cfqq->service_tree->count == 2 &&
3795             RB_EMPTY_ROOT(&cfqq->sort_list))
3796                 return true;
3797
3798         /*
3799          * So both queues are sync. Let the new request get disk time if
3800          * it's a metadata request and the current queue is doing regular IO.
3801          */
3802         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3803                 return true;
3804
3805         /*
3806          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3807          */
3808         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3809                 return true;
3810
3811         /* An idle queue should not be idle now for some reason */
3812         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3813                 return true;
3814
3815         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3816                 return false;
3817
3818         /*
3819          * if this request is as-good as one we would expect from the
3820          * current cfqq, let it preempt
3821          */
3822         if (cfq_rq_close(cfqd, cfqq, rq))
3823                 return true;
3824
3825         return false;
3826 }
3827
3828 /*
3829  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3830  * let it have half of its nominal slice.
3831  */
3832 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3833 {
3834         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3835
3836         cfq_log_cfqq(cfqd, cfqq, "preempt");
3837         cfq_slice_expired(cfqd, 1);
3838
3839         /*
3840          * workload type is changed, don't save slice, otherwise preempt
3841          * doesn't happen
3842          */
3843         if (old_type != cfqq_type(cfqq))
3844                 cfqq->cfqg->saved_wl_slice = 0;
3845
3846         /*
3847          * Put the new queue at the front of the of the current list,
3848          * so we know that it will be selected next.
3849          */
3850         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3851
3852         cfq_service_tree_add(cfqd, cfqq, 1);
3853
3854         cfqq->slice_end = 0;
3855         cfq_mark_cfqq_slice_new(cfqq);
3856 }
3857
3858 /*
3859  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3860  * something we should do about it
3861  */
3862 static void
3863 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3864                 struct request *rq)
3865 {
3866         struct cfq_io_cq *cic = RQ_CIC(rq);
3867
3868         cfqd->rq_queued++;
3869         if (rq->cmd_flags & REQ_PRIO)
3870                 cfqq->prio_pending++;
3871
3872         cfq_update_io_thinktime(cfqd, cfqq, cic);
3873         cfq_update_io_seektime(cfqd, cfqq, rq);
3874         cfq_update_idle_window(cfqd, cfqq, cic);
3875
3876         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3877
3878         if (cfqq == cfqd->active_queue) {
3879                 /*
3880                  * Remember that we saw a request from this process, but
3881                  * don't start queuing just yet. Otherwise we risk seeing lots
3882                  * of tiny requests, because we disrupt the normal plugging
3883                  * and merging. If the request is already larger than a single
3884                  * page, let it rip immediately. For that case we assume that
3885                  * merging is already done. Ditto for a busy system that
3886                  * has other work pending, don't risk delaying until the
3887                  * idle timer unplug to continue working.
3888                  */
3889                 if (cfq_cfqq_wait_request(cfqq)) {
3890                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3891                             cfqd->busy_queues > 1) {
3892                                 cfq_del_timer(cfqd, cfqq);
3893                                 cfq_clear_cfqq_wait_request(cfqq);
3894                                 __blk_run_queue(cfqd->queue);
3895                         } else {
3896                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3897                                 cfq_mark_cfqq_must_dispatch(cfqq);
3898                         }
3899                 }
3900         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3901                 /*
3902                  * not the active queue - expire current slice if it is
3903                  * idle and has expired it's mean thinktime or this new queue
3904                  * has some old slice time left and is of higher priority or
3905                  * this new queue is RT and the current one is BE
3906                  */
3907                 cfq_preempt_queue(cfqd, cfqq);
3908                 __blk_run_queue(cfqd->queue);
3909         }
3910 }
3911
3912 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3913 {
3914         struct cfq_data *cfqd = q->elevator->elevator_data;
3915         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3916
3917         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3918         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3919
3920         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3921         list_add_tail(&rq->queuelist, &cfqq->fifo);
3922         cfq_add_rq_rb(rq);
3923         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3924                                  rq->cmd_flags);
3925         cfq_rq_enqueued(cfqd, cfqq, rq);
3926 }
3927
3928 /*
3929  * Update hw_tag based on peak queue depth over 50 samples under
3930  * sufficient load.
3931  */
3932 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3933 {
3934         struct cfq_queue *cfqq = cfqd->active_queue;
3935
3936         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3937                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3938
3939         if (cfqd->hw_tag == 1)
3940                 return;
3941
3942         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3943             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3944                 return;
3945
3946         /*
3947          * If active queue hasn't enough requests and can idle, cfq might not
3948          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3949          * case
3950          */
3951         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3952             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3953             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3954                 return;
3955
3956         if (cfqd->hw_tag_samples++ < 50)
3957                 return;
3958
3959         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3960                 cfqd->hw_tag = 1;
3961         else
3962                 cfqd->hw_tag = 0;
3963 }
3964
3965 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3966 {
3967         struct cfq_io_cq *cic = cfqd->active_cic;
3968
3969         /* If the queue already has requests, don't wait */
3970         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3971                 return false;
3972
3973         /* If there are other queues in the group, don't wait */
3974         if (cfqq->cfqg->nr_cfqq > 1)
3975                 return false;
3976
3977         /* the only queue in the group, but think time is big */
3978         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3979                 return false;
3980
3981         if (cfq_slice_used(cfqq))
3982                 return true;
3983
3984         /* if slice left is less than think time, wait busy */
3985         if (cic && sample_valid(cic->ttime.ttime_samples)
3986             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3987                 return true;
3988
3989         /*
3990          * If think times is less than a jiffy than ttime_mean=0 and above
3991          * will not be true. It might happen that slice has not expired yet
3992          * but will expire soon (4-5 ns) during select_queue(). To cover the
3993          * case where think time is less than a jiffy, mark the queue wait
3994          * busy if only 1 jiffy is left in the slice.
3995          */
3996         if (cfqq->slice_end - jiffies == 1)
3997                 return true;
3998
3999         return false;
4000 }
4001
4002 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4003 {
4004         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4005         struct cfq_data *cfqd = cfqq->cfqd;
4006         const int sync = rq_is_sync(rq);
4007         unsigned long now;
4008
4009         now = jiffies;
4010         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4011                      !!(rq->cmd_flags & REQ_NOIDLE));
4012
4013         cfq_update_hw_tag(cfqd);
4014
4015         WARN_ON(!cfqd->rq_in_driver);
4016         WARN_ON(!cfqq->dispatched);
4017         cfqd->rq_in_driver--;
4018         cfqq->dispatched--;
4019         (RQ_CFQG(rq))->dispatched--;
4020         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4021                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4022
4023         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4024
4025         if (sync) {
4026                 struct cfq_rb_root *st;
4027
4028                 RQ_CIC(rq)->ttime.last_end_request = now;
4029
4030                 if (cfq_cfqq_on_rr(cfqq))
4031                         st = cfqq->service_tree;
4032                 else
4033                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4034                                         cfqq_type(cfqq));
4035
4036                 st->ttime.last_end_request = now;
4037                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4038                         cfqd->last_delayed_sync = now;
4039         }
4040
4041 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4042         cfqq->cfqg->ttime.last_end_request = now;
4043 #endif
4044
4045         /*
4046          * If this is the active queue, check if it needs to be expired,
4047          * or if we want to idle in case it has no pending requests.
4048          */
4049         if (cfqd->active_queue == cfqq) {
4050                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4051
4052                 if (cfq_cfqq_slice_new(cfqq)) {
4053                         cfq_set_prio_slice(cfqd, cfqq);
4054                         cfq_clear_cfqq_slice_new(cfqq);
4055                 }
4056
4057                 /*
4058                  * Should we wait for next request to come in before we expire
4059                  * the queue.
4060                  */
4061                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4062                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4063                         if (!cfqd->cfq_slice_idle)
4064                                 extend_sl = cfqd->cfq_group_idle;
4065                         cfqq->slice_end = jiffies + extend_sl;
4066                         cfq_mark_cfqq_wait_busy(cfqq);
4067                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4068                 }
4069
4070                 /*
4071                  * Idling is not enabled on:
4072                  * - expired queues
4073                  * - idle-priority queues
4074                  * - async queues
4075                  * - queues with still some requests queued
4076                  * - when there is a close cooperator
4077                  */
4078                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4079                         cfq_slice_expired(cfqd, 1);
4080                 else if (sync && cfqq_empty &&
4081                          !cfq_close_cooperator(cfqd, cfqq)) {
4082                         cfq_arm_slice_timer(cfqd);
4083                 }
4084         }
4085
4086         if (!cfqd->rq_in_driver)
4087                 cfq_schedule_dispatch(cfqd);
4088 }
4089
4090 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4091 {
4092         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4093                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4094                 return ELV_MQUEUE_MUST;
4095         }
4096
4097         return ELV_MQUEUE_MAY;
4098 }
4099
4100 static int cfq_may_queue(struct request_queue *q, int rw)
4101 {
4102         struct cfq_data *cfqd = q->elevator->elevator_data;
4103         struct task_struct *tsk = current;
4104         struct cfq_io_cq *cic;
4105         struct cfq_queue *cfqq;
4106
4107         /*
4108          * don't force setup of a queue from here, as a call to may_queue
4109          * does not necessarily imply that a request actually will be queued.
4110          * so just lookup a possibly existing queue, or return 'may queue'
4111          * if that fails
4112          */
4113         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4114         if (!cic)
4115                 return ELV_MQUEUE_MAY;
4116
4117         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4118         if (cfqq) {
4119                 cfq_init_prio_data(cfqq, cic);
4120
4121                 return __cfq_may_queue(cfqq);
4122         }
4123
4124         return ELV_MQUEUE_MAY;
4125 }
4126
4127 /*
4128  * queue lock held here
4129  */
4130 static void cfq_put_request(struct request *rq)
4131 {
4132         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4133
4134         if (cfqq) {
4135                 const int rw = rq_data_dir(rq);
4136
4137                 BUG_ON(!cfqq->allocated[rw]);
4138                 cfqq->allocated[rw]--;
4139
4140                 /* Put down rq reference on cfqg */
4141                 cfqg_put(RQ_CFQG(rq));
4142                 rq->elv.priv[0] = NULL;
4143                 rq->elv.priv[1] = NULL;
4144
4145                 cfq_put_queue(cfqq);
4146         }
4147 }
4148
4149 static struct cfq_queue *
4150 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4151                 struct cfq_queue *cfqq)
4152 {
4153         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4154         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4155         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4156         cfq_put_queue(cfqq);
4157         return cic_to_cfqq(cic, 1);
4158 }
4159
4160 /*
4161  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4162  * was the last process referring to said cfqq.
4163  */
4164 static struct cfq_queue *
4165 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4166 {
4167         if (cfqq_process_refs(cfqq) == 1) {
4168                 cfqq->pid = current->pid;
4169                 cfq_clear_cfqq_coop(cfqq);
4170                 cfq_clear_cfqq_split_coop(cfqq);
4171                 return cfqq;
4172         }
4173
4174         cic_set_cfqq(cic, NULL, 1);
4175
4176         cfq_put_cooperator(cfqq);
4177
4178         cfq_put_queue(cfqq);
4179         return NULL;
4180 }
4181 /*
4182  * Allocate cfq data structures associated with this request.
4183  */
4184 static int
4185 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4186                 gfp_t gfp_mask)
4187 {
4188         struct cfq_data *cfqd = q->elevator->elevator_data;
4189         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4190         const int rw = rq_data_dir(rq);
4191         const bool is_sync = rq_is_sync(rq);
4192         struct cfq_queue *cfqq;
4193
4194         might_sleep_if(gfp_mask & __GFP_WAIT);
4195
4196         spin_lock_irq(q->queue_lock);
4197
4198         check_ioprio_changed(cic, bio);
4199         check_blkcg_changed(cic, bio);
4200 new_queue:
4201         cfqq = cic_to_cfqq(cic, is_sync);
4202         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4203                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4204                 cic_set_cfqq(cic, cfqq, is_sync);
4205         } else {
4206                 /*
4207                  * If the queue was seeky for too long, break it apart.
4208                  */
4209                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4210                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4211                         cfqq = split_cfqq(cic, cfqq);
4212                         if (!cfqq)
4213                                 goto new_queue;
4214                 }
4215
4216                 /*
4217                  * Check to see if this queue is scheduled to merge with
4218                  * another, closely cooperating queue.  The merging of
4219                  * queues happens here as it must be done in process context.
4220                  * The reference on new_cfqq was taken in merge_cfqqs.
4221                  */
4222                 if (cfqq->new_cfqq)
4223                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4224         }
4225
4226         cfqq->allocated[rw]++;
4227
4228         cfqq->ref++;
4229         cfqg_get(cfqq->cfqg);
4230         rq->elv.priv[0] = cfqq;
4231         rq->elv.priv[1] = cfqq->cfqg;
4232         spin_unlock_irq(q->queue_lock);
4233         return 0;
4234 }
4235
4236 static void cfq_kick_queue(struct work_struct *work)
4237 {
4238         struct cfq_data *cfqd =
4239                 container_of(work, struct cfq_data, unplug_work);
4240         struct request_queue *q = cfqd->queue;
4241
4242         spin_lock_irq(q->queue_lock);
4243         __blk_run_queue(cfqd->queue);
4244         spin_unlock_irq(q->queue_lock);
4245 }
4246
4247 /*
4248  * Timer running if the active_queue is currently idling inside its time slice
4249  */
4250 static void cfq_idle_slice_timer(unsigned long data)
4251 {
4252         struct cfq_data *cfqd = (struct cfq_data *) data;
4253         struct cfq_queue *cfqq;
4254         unsigned long flags;
4255         int timed_out = 1;
4256
4257         cfq_log(cfqd, "idle timer fired");
4258
4259         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4260
4261         cfqq = cfqd->active_queue;
4262         if (cfqq) {
4263                 timed_out = 0;
4264
4265                 /*
4266                  * We saw a request before the queue expired, let it through
4267                  */
4268                 if (cfq_cfqq_must_dispatch(cfqq))
4269                         goto out_kick;
4270
4271                 /*
4272                  * expired
4273                  */
4274                 if (cfq_slice_used(cfqq))
4275                         goto expire;
4276
4277                 /*
4278                  * only expire and reinvoke request handler, if there are
4279                  * other queues with pending requests
4280                  */
4281                 if (!cfqd->busy_queues)
4282                         goto out_cont;
4283
4284                 /*
4285                  * not expired and it has a request pending, let it dispatch
4286                  */
4287                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4288                         goto out_kick;
4289
4290                 /*
4291                  * Queue depth flag is reset only when the idle didn't succeed
4292                  */
4293                 cfq_clear_cfqq_deep(cfqq);
4294         }
4295 expire:
4296         cfq_slice_expired(cfqd, timed_out);
4297 out_kick:
4298         cfq_schedule_dispatch(cfqd);
4299 out_cont:
4300         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4301 }
4302
4303 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4304 {
4305         del_timer_sync(&cfqd->idle_slice_timer);
4306         cancel_work_sync(&cfqd->unplug_work);
4307 }
4308
4309 static void cfq_put_async_queues(struct cfq_data *cfqd)
4310 {
4311         int i;
4312
4313         for (i = 0; i < IOPRIO_BE_NR; i++) {
4314                 if (cfqd->async_cfqq[0][i])
4315                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4316                 if (cfqd->async_cfqq[1][i])
4317                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4318         }
4319
4320         if (cfqd->async_idle_cfqq)
4321                 cfq_put_queue(cfqd->async_idle_cfqq);
4322 }
4323
4324 static void cfq_exit_queue(struct elevator_queue *e)
4325 {
4326         struct cfq_data *cfqd = e->elevator_data;
4327         struct request_queue *q = cfqd->queue;
4328
4329         cfq_shutdown_timer_wq(cfqd);
4330
4331         spin_lock_irq(q->queue_lock);
4332
4333         if (cfqd->active_queue)
4334                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4335
4336         cfq_put_async_queues(cfqd);
4337
4338         spin_unlock_irq(q->queue_lock);
4339
4340         cfq_shutdown_timer_wq(cfqd);
4341
4342 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4343         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4344 #else
4345         kfree(cfqd->root_group);
4346 #endif
4347         kfree(cfqd);
4348 }
4349
4350 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4351 {
4352         struct cfq_data *cfqd;
4353         struct blkcg_gq *blkg __maybe_unused;
4354         int i, ret;
4355         struct elevator_queue *eq;
4356
4357         eq = elevator_alloc(q, e);
4358         if (!eq)
4359                 return -ENOMEM;
4360
4361         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4362         if (!cfqd) {
4363                 kobject_put(&eq->kobj);
4364                 return -ENOMEM;
4365         }
4366         eq->elevator_data = cfqd;
4367
4368         cfqd->queue = q;
4369         spin_lock_irq(q->queue_lock);
4370         q->elevator = eq;
4371         spin_unlock_irq(q->queue_lock);
4372
4373         /* Init root service tree */
4374         cfqd->grp_service_tree = CFQ_RB_ROOT;
4375
4376         /* Init root group and prefer root group over other groups by default */
4377 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4378         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4379         if (ret)
4380                 goto out_free;
4381
4382         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4383 #else
4384         ret = -ENOMEM;
4385         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4386                                         GFP_KERNEL, cfqd->queue->node);
4387         if (!cfqd->root_group)
4388                 goto out_free;
4389
4390         cfq_init_cfqg_base(cfqd->root_group);
4391 #endif
4392         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4393         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4394
4395         /*
4396          * Not strictly needed (since RB_ROOT just clears the node and we
4397          * zeroed cfqd on alloc), but better be safe in case someone decides
4398          * to add magic to the rb code
4399          */
4400         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4401                 cfqd->prio_trees[i] = RB_ROOT;
4402
4403         /*
4404          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4405          * Grab a permanent reference to it, so that the normal code flow
4406          * will not attempt to free it.  oom_cfqq is linked to root_group
4407          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4408          * the reference from linking right away.
4409          */
4410         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4411         cfqd->oom_cfqq.ref++;
4412
4413         spin_lock_irq(q->queue_lock);
4414         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4415         cfqg_put(cfqd->root_group);
4416         spin_unlock_irq(q->queue_lock);
4417
4418         init_timer(&cfqd->idle_slice_timer);
4419         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4420         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4421
4422         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4423
4424         cfqd->cfq_quantum = cfq_quantum;
4425         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4426         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4427         cfqd->cfq_back_max = cfq_back_max;
4428         cfqd->cfq_back_penalty = cfq_back_penalty;
4429         cfqd->cfq_slice[0] = cfq_slice_async;
4430         cfqd->cfq_slice[1] = cfq_slice_sync;
4431         cfqd->cfq_target_latency = cfq_target_latency;
4432         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4433         cfqd->cfq_slice_idle = cfq_slice_idle;
4434         cfqd->cfq_group_idle = cfq_group_idle;
4435         cfqd->cfq_latency = 1;
4436         cfqd->hw_tag = -1;
4437         /*
4438          * we optimistically start assuming sync ops weren't delayed in last
4439          * second, in order to have larger depth for async operations.
4440          */
4441         cfqd->last_delayed_sync = jiffies - HZ;
4442         return 0;
4443
4444 out_free:
4445         kfree(cfqd);
4446         kobject_put(&eq->kobj);
4447         return ret;
4448 }
4449
4450 /*
4451  * sysfs parts below -->
4452  */
4453 static ssize_t
4454 cfq_var_show(unsigned int var, char *page)
4455 {
4456         return sprintf(page, "%d\n", var);
4457 }
4458
4459 static ssize_t
4460 cfq_var_store(unsigned int *var, const char *page, size_t count)
4461 {
4462         char *p = (char *) page;
4463
4464         *var = simple_strtoul(p, &p, 10);
4465         return count;
4466 }
4467
4468 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4469 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4470 {                                                                       \
4471         struct cfq_data *cfqd = e->elevator_data;                       \
4472         unsigned int __data = __VAR;                                    \
4473         if (__CONV)                                                     \
4474                 __data = jiffies_to_msecs(__data);                      \
4475         return cfq_var_show(__data, (page));                            \
4476 }
4477 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4478 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4479 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4480 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4481 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4482 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4483 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4484 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4485 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4486 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4487 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4488 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4489 #undef SHOW_FUNCTION
4490
4491 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4492 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4493 {                                                                       \
4494         struct cfq_data *cfqd = e->elevator_data;                       \
4495         unsigned int __data;                                            \
4496         int ret = cfq_var_store(&__data, (page), count);                \
4497         if (__data < (MIN))                                             \
4498                 __data = (MIN);                                         \
4499         else if (__data > (MAX))                                        \
4500                 __data = (MAX);                                         \
4501         if (__CONV)                                                     \
4502                 *(__PTR) = msecs_to_jiffies(__data);                    \
4503         else                                                            \
4504                 *(__PTR) = __data;                                      \
4505         return ret;                                                     \
4506 }
4507 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4508 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4509                 UINT_MAX, 1);
4510 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4511                 UINT_MAX, 1);
4512 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4513 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4514                 UINT_MAX, 0);
4515 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4516 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4517 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4518 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4519 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4520                 UINT_MAX, 0);
4521 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4522 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4523 #undef STORE_FUNCTION
4524
4525 #define CFQ_ATTR(name) \
4526         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4527
4528 static struct elv_fs_entry cfq_attrs[] = {
4529         CFQ_ATTR(quantum),
4530         CFQ_ATTR(fifo_expire_sync),
4531         CFQ_ATTR(fifo_expire_async),
4532         CFQ_ATTR(back_seek_max),
4533         CFQ_ATTR(back_seek_penalty),
4534         CFQ_ATTR(slice_sync),
4535         CFQ_ATTR(slice_async),
4536         CFQ_ATTR(slice_async_rq),
4537         CFQ_ATTR(slice_idle),
4538         CFQ_ATTR(group_idle),
4539         CFQ_ATTR(low_latency),
4540         CFQ_ATTR(target_latency),
4541         __ATTR_NULL
4542 };
4543
4544 static struct elevator_type iosched_cfq = {
4545         .ops = {
4546                 .elevator_merge_fn =            cfq_merge,
4547                 .elevator_merged_fn =           cfq_merged_request,
4548                 .elevator_merge_req_fn =        cfq_merged_requests,
4549                 .elevator_allow_merge_fn =      cfq_allow_merge,
4550                 .elevator_bio_merged_fn =       cfq_bio_merged,
4551                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4552                 .elevator_add_req_fn =          cfq_insert_request,
4553                 .elevator_activate_req_fn =     cfq_activate_request,
4554                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4555                 .elevator_completed_req_fn =    cfq_completed_request,
4556                 .elevator_former_req_fn =       elv_rb_former_request,
4557                 .elevator_latter_req_fn =       elv_rb_latter_request,
4558                 .elevator_init_icq_fn =         cfq_init_icq,
4559                 .elevator_exit_icq_fn =         cfq_exit_icq,
4560                 .elevator_set_req_fn =          cfq_set_request,
4561                 .elevator_put_req_fn =          cfq_put_request,
4562                 .elevator_may_queue_fn =        cfq_may_queue,
4563                 .elevator_init_fn =             cfq_init_queue,
4564                 .elevator_exit_fn =             cfq_exit_queue,
4565         },
4566         .icq_size       =       sizeof(struct cfq_io_cq),
4567         .icq_align      =       __alignof__(struct cfq_io_cq),
4568         .elevator_attrs =       cfq_attrs,
4569         .elevator_name  =       "cfq",
4570         .elevator_owner =       THIS_MODULE,
4571 };
4572
4573 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4574 static struct blkcg_policy blkcg_policy_cfq = {
4575         .pd_size                = sizeof(struct cfq_group),
4576         .cftypes                = cfq_blkcg_files,
4577
4578         .pd_init_fn             = cfq_pd_init,
4579         .pd_offline_fn          = cfq_pd_offline,
4580         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4581 };
4582 #endif
4583
4584 static int __init cfq_init(void)
4585 {
4586         int ret;
4587
4588         /*
4589          * could be 0 on HZ < 1000 setups
4590          */
4591         if (!cfq_slice_async)
4592                 cfq_slice_async = 1;
4593         if (!cfq_slice_idle)
4594                 cfq_slice_idle = 1;
4595
4596 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4597         if (!cfq_group_idle)
4598                 cfq_group_idle = 1;
4599
4600         ret = blkcg_policy_register(&blkcg_policy_cfq);
4601         if (ret)
4602                 return ret;
4603 #else
4604         cfq_group_idle = 0;
4605 #endif
4606
4607         ret = -ENOMEM;
4608         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4609         if (!cfq_pool)
4610                 goto err_pol_unreg;
4611
4612         ret = elv_register(&iosched_cfq);
4613         if (ret)
4614                 goto err_free_pool;
4615
4616         return 0;
4617
4618 err_free_pool:
4619         kmem_cache_destroy(cfq_pool);
4620 err_pol_unreg:
4621 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4622         blkcg_policy_unregister(&blkcg_policy_cfq);
4623 #endif
4624         return ret;
4625 }
4626
4627 static void __exit cfq_exit(void)
4628 {
4629 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4630         blkcg_policy_unregister(&blkcg_policy_cfq);
4631 #endif
4632         elv_unregister(&iosched_cfq);
4633         kmem_cache_destroy(cfq_pool);
4634 }
4635
4636 module_init(cfq_init);
4637 module_exit(cfq_exit);
4638
4639 MODULE_AUTHOR("Jens Axboe");
4640 MODULE_LICENSE("GPL");
4641 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");