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 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1532 {
1533         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1534
1535         cfqg_stats_reset(&cfqg->stats);
1536         cfqg_stats_reset(&cfqg->dead_stats);
1537 }
1538
1539 /*
1540  * Search for the cfq group current task belongs to. request_queue lock must
1541  * be held.
1542  */
1543 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1544                                                 struct blkcg *blkcg)
1545 {
1546         struct request_queue *q = cfqd->queue;
1547         struct cfq_group *cfqg = NULL;
1548
1549         /* avoid lookup for the common case where there's no blkcg */
1550         if (blkcg == &blkcg_root) {
1551                 cfqg = cfqd->root_group;
1552         } else {
1553                 struct blkcg_gq *blkg;
1554
1555                 blkg = blkg_lookup_create(blkcg, q);
1556                 if (!IS_ERR(blkg))
1557                         cfqg = blkg_to_cfqg(blkg);
1558         }
1559
1560         return cfqg;
1561 }
1562
1563 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1564 {
1565         /* Currently, all async queues are mapped to root group */
1566         if (!cfq_cfqq_sync(cfqq))
1567                 cfqg = cfqq->cfqd->root_group;
1568
1569         cfqq->cfqg = cfqg;
1570         /* cfqq reference on cfqg */
1571         cfqg_get(cfqg);
1572 }
1573
1574 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1575                                      struct blkg_policy_data *pd, int off)
1576 {
1577         struct cfq_group *cfqg = pd_to_cfqg(pd);
1578
1579         if (!cfqg->dev_weight)
1580                 return 0;
1581         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1582 }
1583
1584 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1585                                     struct seq_file *sf)
1586 {
1587         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1588                           cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1589                           false);
1590         return 0;
1591 }
1592
1593 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1594                                           struct blkg_policy_data *pd, int off)
1595 {
1596         struct cfq_group *cfqg = pd_to_cfqg(pd);
1597
1598         if (!cfqg->dev_leaf_weight)
1599                 return 0;
1600         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1601 }
1602
1603 static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1604                                          struct cftype *cft,
1605                                          struct seq_file *sf)
1606 {
1607         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1608                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1609                           false);
1610         return 0;
1611 }
1612
1613 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1614                             struct seq_file *sf)
1615 {
1616         seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1617         return 0;
1618 }
1619
1620 static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1621                                  struct seq_file *sf)
1622 {
1623         seq_printf(sf, "%u\n",
1624                    cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1625         return 0;
1626 }
1627
1628 static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1629                                     const char *buf, bool is_leaf_weight)
1630 {
1631         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1632         struct blkg_conf_ctx ctx;
1633         struct cfq_group *cfqg;
1634         int ret;
1635
1636         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1637         if (ret)
1638                 return ret;
1639
1640         ret = -EINVAL;
1641         cfqg = blkg_to_cfqg(ctx.blkg);
1642         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1643                 if (!is_leaf_weight) {
1644                         cfqg->dev_weight = ctx.v;
1645                         cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1646                 } else {
1647                         cfqg->dev_leaf_weight = ctx.v;
1648                         cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1649                 }
1650                 ret = 0;
1651         }
1652
1653         blkg_conf_finish(&ctx);
1654         return ret;
1655 }
1656
1657 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1658                                   const char *buf)
1659 {
1660         return __cfqg_set_weight_device(cgrp, cft, buf, false);
1661 }
1662
1663 static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1664                                        const char *buf)
1665 {
1666         return __cfqg_set_weight_device(cgrp, cft, buf, true);
1667 }
1668
1669 static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1670                             bool is_leaf_weight)
1671 {
1672         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1673         struct blkcg_gq *blkg;
1674         struct hlist_node *n;
1675
1676         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1677                 return -EINVAL;
1678
1679         spin_lock_irq(&blkcg->lock);
1680
1681         if (!is_leaf_weight)
1682                 blkcg->cfq_weight = val;
1683         else
1684                 blkcg->cfq_leaf_weight = val;
1685
1686         hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1687                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1688
1689                 if (!cfqg)
1690                         continue;
1691
1692                 if (!is_leaf_weight) {
1693                         if (!cfqg->dev_weight)
1694                                 cfqg->new_weight = blkcg->cfq_weight;
1695                 } else {
1696                         if (!cfqg->dev_leaf_weight)
1697                                 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1698                 }
1699         }
1700
1701         spin_unlock_irq(&blkcg->lock);
1702         return 0;
1703 }
1704
1705 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1706 {
1707         return __cfq_set_weight(cgrp, cft, val, false);
1708 }
1709
1710 static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1711 {
1712         return __cfq_set_weight(cgrp, cft, val, true);
1713 }
1714
1715 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1716                            struct seq_file *sf)
1717 {
1718         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1719
1720         blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1721                           cft->private, false);
1722         return 0;
1723 }
1724
1725 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1726                              struct seq_file *sf)
1727 {
1728         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1729
1730         blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1731                           cft->private, true);
1732         return 0;
1733 }
1734
1735 #ifdef CONFIG_DEBUG_BLK_CGROUP
1736 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1737                                       struct blkg_policy_data *pd, int off)
1738 {
1739         struct cfq_group *cfqg = pd_to_cfqg(pd);
1740         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1741         u64 v = 0;
1742
1743         if (samples) {
1744                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1745                 do_div(v, samples);
1746         }
1747         __blkg_prfill_u64(sf, pd, v);
1748         return 0;
1749 }
1750
1751 /* print avg_queue_size */
1752 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1753                                      struct seq_file *sf)
1754 {
1755         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1756
1757         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1758                           &blkcg_policy_cfq, 0, false);
1759         return 0;
1760 }
1761 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1762
1763 static struct cftype cfq_blkcg_files[] = {
1764         /* on root, weight is mapped to leaf_weight */
1765         {
1766                 .name = "weight_device",
1767                 .flags = CFTYPE_ONLY_ON_ROOT,
1768                 .read_seq_string = cfqg_print_leaf_weight_device,
1769                 .write_string = cfqg_set_leaf_weight_device,
1770                 .max_write_len = 256,
1771         },
1772         {
1773                 .name = "weight",
1774                 .flags = CFTYPE_ONLY_ON_ROOT,
1775                 .read_seq_string = cfq_print_leaf_weight,
1776                 .write_u64 = cfq_set_leaf_weight,
1777         },
1778
1779         /* no such mapping necessary for !roots */
1780         {
1781                 .name = "weight_device",
1782                 .flags = CFTYPE_NOT_ON_ROOT,
1783                 .read_seq_string = cfqg_print_weight_device,
1784                 .write_string = cfqg_set_weight_device,
1785                 .max_write_len = 256,
1786         },
1787         {
1788                 .name = "weight",
1789                 .flags = CFTYPE_NOT_ON_ROOT,
1790                 .read_seq_string = cfq_print_weight,
1791                 .write_u64 = cfq_set_weight,
1792         },
1793
1794         {
1795                 .name = "leaf_weight_device",
1796                 .read_seq_string = cfqg_print_leaf_weight_device,
1797                 .write_string = cfqg_set_leaf_weight_device,
1798                 .max_write_len = 256,
1799         },
1800         {
1801                 .name = "leaf_weight",
1802                 .read_seq_string = cfq_print_leaf_weight,
1803                 .write_u64 = cfq_set_leaf_weight,
1804         },
1805
1806         {
1807                 .name = "time",
1808                 .private = offsetof(struct cfq_group, stats.time),
1809                 .read_seq_string = cfqg_print_stat,
1810         },
1811         {
1812                 .name = "sectors",
1813                 .private = offsetof(struct cfq_group, stats.sectors),
1814                 .read_seq_string = cfqg_print_stat,
1815         },
1816         {
1817                 .name = "io_service_bytes",
1818                 .private = offsetof(struct cfq_group, stats.service_bytes),
1819                 .read_seq_string = cfqg_print_rwstat,
1820         },
1821         {
1822                 .name = "io_serviced",
1823                 .private = offsetof(struct cfq_group, stats.serviced),
1824                 .read_seq_string = cfqg_print_rwstat,
1825         },
1826         {
1827                 .name = "io_service_time",
1828                 .private = offsetof(struct cfq_group, stats.service_time),
1829                 .read_seq_string = cfqg_print_rwstat,
1830         },
1831         {
1832                 .name = "io_wait_time",
1833                 .private = offsetof(struct cfq_group, stats.wait_time),
1834                 .read_seq_string = cfqg_print_rwstat,
1835         },
1836         {
1837                 .name = "io_merged",
1838                 .private = offsetof(struct cfq_group, stats.merged),
1839                 .read_seq_string = cfqg_print_rwstat,
1840         },
1841         {
1842                 .name = "io_queued",
1843                 .private = offsetof(struct cfq_group, stats.queued),
1844                 .read_seq_string = cfqg_print_rwstat,
1845         },
1846 #ifdef CONFIG_DEBUG_BLK_CGROUP
1847         {
1848                 .name = "avg_queue_size",
1849                 .read_seq_string = cfqg_print_avg_queue_size,
1850         },
1851         {
1852                 .name = "group_wait_time",
1853                 .private = offsetof(struct cfq_group, stats.group_wait_time),
1854                 .read_seq_string = cfqg_print_stat,
1855         },
1856         {
1857                 .name = "idle_time",
1858                 .private = offsetof(struct cfq_group, stats.idle_time),
1859                 .read_seq_string = cfqg_print_stat,
1860         },
1861         {
1862                 .name = "empty_time",
1863                 .private = offsetof(struct cfq_group, stats.empty_time),
1864                 .read_seq_string = cfqg_print_stat,
1865         },
1866         {
1867                 .name = "dequeue",
1868                 .private = offsetof(struct cfq_group, stats.dequeue),
1869                 .read_seq_string = cfqg_print_stat,
1870         },
1871         {
1872                 .name = "unaccounted_time",
1873                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1874                 .read_seq_string = cfqg_print_stat,
1875         },
1876 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1877         { }     /* terminate */
1878 };
1879 #else /* GROUP_IOSCHED */
1880 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1881                                                 struct blkcg *blkcg)
1882 {
1883         return cfqd->root_group;
1884 }
1885
1886 static inline void
1887 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1888         cfqq->cfqg = cfqg;
1889 }
1890
1891 #endif /* GROUP_IOSCHED */
1892
1893 /*
1894  * The cfqd->service_trees holds all pending cfq_queue's that have
1895  * requests waiting to be processed. It is sorted in the order that
1896  * we will service the queues.
1897  */
1898 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1899                                  bool add_front)
1900 {
1901         struct rb_node **p, *parent;
1902         struct cfq_queue *__cfqq;
1903         unsigned long rb_key;
1904         struct cfq_rb_root *st;
1905         int left;
1906         int new_cfqq = 1;
1907
1908         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
1909         if (cfq_class_idle(cfqq)) {
1910                 rb_key = CFQ_IDLE_DELAY;
1911                 parent = rb_last(&st->rb);
1912                 if (parent && parent != &cfqq->rb_node) {
1913                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1914                         rb_key += __cfqq->rb_key;
1915                 } else
1916                         rb_key += jiffies;
1917         } else if (!add_front) {
1918                 /*
1919                  * Get our rb key offset. Subtract any residual slice
1920                  * value carried from last service. A negative resid
1921                  * count indicates slice overrun, and this should position
1922                  * the next service time further away in the tree.
1923                  */
1924                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1925                 rb_key -= cfqq->slice_resid;
1926                 cfqq->slice_resid = 0;
1927         } else {
1928                 rb_key = -HZ;
1929                 __cfqq = cfq_rb_first(st);
1930                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1931         }
1932
1933         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1934                 new_cfqq = 0;
1935                 /*
1936                  * same position, nothing more to do
1937                  */
1938                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
1939                         return;
1940
1941                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1942                 cfqq->service_tree = NULL;
1943         }
1944
1945         left = 1;
1946         parent = NULL;
1947         cfqq->service_tree = st;
1948         p = &st->rb.rb_node;
1949         while (*p) {
1950                 parent = *p;
1951                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1952
1953                 /*
1954                  * sort by key, that represents service time.
1955                  */
1956                 if (time_before(rb_key, __cfqq->rb_key))
1957                         p = &parent->rb_left;
1958                 else {
1959                         p = &parent->rb_right;
1960                         left = 0;
1961                 }
1962         }
1963
1964         if (left)
1965                 st->left = &cfqq->rb_node;
1966
1967         cfqq->rb_key = rb_key;
1968         rb_link_node(&cfqq->rb_node, parent, p);
1969         rb_insert_color(&cfqq->rb_node, &st->rb);
1970         st->count++;
1971         if (add_front || !new_cfqq)
1972                 return;
1973         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1974 }
1975
1976 static struct cfq_queue *
1977 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1978                      sector_t sector, struct rb_node **ret_parent,
1979                      struct rb_node ***rb_link)
1980 {
1981         struct rb_node **p, *parent;
1982         struct cfq_queue *cfqq = NULL;
1983
1984         parent = NULL;
1985         p = &root->rb_node;
1986         while (*p) {
1987                 struct rb_node **n;
1988
1989                 parent = *p;
1990                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1991
1992                 /*
1993                  * Sort strictly based on sector.  Smallest to the left,
1994                  * largest to the right.
1995                  */
1996                 if (sector > blk_rq_pos(cfqq->next_rq))
1997                         n = &(*p)->rb_right;
1998                 else if (sector < blk_rq_pos(cfqq->next_rq))
1999                         n = &(*p)->rb_left;
2000                 else
2001                         break;
2002                 p = n;
2003                 cfqq = NULL;
2004         }
2005
2006         *ret_parent = parent;
2007         if (rb_link)
2008                 *rb_link = p;
2009         return cfqq;
2010 }
2011
2012 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2013 {
2014         struct rb_node **p, *parent;
2015         struct cfq_queue *__cfqq;
2016
2017         if (cfqq->p_root) {
2018                 rb_erase(&cfqq->p_node, cfqq->p_root);
2019                 cfqq->p_root = NULL;
2020         }
2021
2022         if (cfq_class_idle(cfqq))
2023                 return;
2024         if (!cfqq->next_rq)
2025                 return;
2026
2027         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2028         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2029                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2030         if (!__cfqq) {
2031                 rb_link_node(&cfqq->p_node, parent, p);
2032                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2033         } else
2034                 cfqq->p_root = NULL;
2035 }
2036
2037 /*
2038  * Update cfqq's position in the service tree.
2039  */
2040 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2041 {
2042         /*
2043          * Resorting requires the cfqq to be on the RR list already.
2044          */
2045         if (cfq_cfqq_on_rr(cfqq)) {
2046                 cfq_service_tree_add(cfqd, cfqq, 0);
2047                 cfq_prio_tree_add(cfqd, cfqq);
2048         }
2049 }
2050
2051 /*
2052  * add to busy list of queues for service, trying to be fair in ordering
2053  * the pending list according to last request service
2054  */
2055 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2056 {
2057         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2058         BUG_ON(cfq_cfqq_on_rr(cfqq));
2059         cfq_mark_cfqq_on_rr(cfqq);
2060         cfqd->busy_queues++;
2061         if (cfq_cfqq_sync(cfqq))
2062                 cfqd->busy_sync_queues++;
2063
2064         cfq_resort_rr_list(cfqd, cfqq);
2065 }
2066
2067 /*
2068  * Called when the cfqq no longer has requests pending, remove it from
2069  * the service tree.
2070  */
2071 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2072 {
2073         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2074         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2075         cfq_clear_cfqq_on_rr(cfqq);
2076
2077         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2078                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2079                 cfqq->service_tree = NULL;
2080         }
2081         if (cfqq->p_root) {
2082                 rb_erase(&cfqq->p_node, cfqq->p_root);
2083                 cfqq->p_root = NULL;
2084         }
2085
2086         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2087         BUG_ON(!cfqd->busy_queues);
2088         cfqd->busy_queues--;
2089         if (cfq_cfqq_sync(cfqq))
2090                 cfqd->busy_sync_queues--;
2091 }
2092
2093 /*
2094  * rb tree support functions
2095  */
2096 static void cfq_del_rq_rb(struct request *rq)
2097 {
2098         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2099         const int sync = rq_is_sync(rq);
2100
2101         BUG_ON(!cfqq->queued[sync]);
2102         cfqq->queued[sync]--;
2103
2104         elv_rb_del(&cfqq->sort_list, rq);
2105
2106         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2107                 /*
2108                  * Queue will be deleted from service tree when we actually
2109                  * expire it later. Right now just remove it from prio tree
2110                  * as it is empty.
2111                  */
2112                 if (cfqq->p_root) {
2113                         rb_erase(&cfqq->p_node, cfqq->p_root);
2114                         cfqq->p_root = NULL;
2115                 }
2116         }
2117 }
2118
2119 static void cfq_add_rq_rb(struct request *rq)
2120 {
2121         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2122         struct cfq_data *cfqd = cfqq->cfqd;
2123         struct request *prev;
2124
2125         cfqq->queued[rq_is_sync(rq)]++;
2126
2127         elv_rb_add(&cfqq->sort_list, rq);
2128
2129         if (!cfq_cfqq_on_rr(cfqq))
2130                 cfq_add_cfqq_rr(cfqd, cfqq);
2131
2132         /*
2133          * check if this request is a better next-serve candidate
2134          */
2135         prev = cfqq->next_rq;
2136         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2137
2138         /*
2139          * adjust priority tree position, if ->next_rq changes
2140          */
2141         if (prev != cfqq->next_rq)
2142                 cfq_prio_tree_add(cfqd, cfqq);
2143
2144         BUG_ON(!cfqq->next_rq);
2145 }
2146
2147 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2148 {
2149         elv_rb_del(&cfqq->sort_list, rq);
2150         cfqq->queued[rq_is_sync(rq)]--;
2151         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2152         cfq_add_rq_rb(rq);
2153         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2154                                  rq->cmd_flags);
2155 }
2156
2157 static struct request *
2158 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2159 {
2160         struct task_struct *tsk = current;
2161         struct cfq_io_cq *cic;
2162         struct cfq_queue *cfqq;
2163
2164         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2165         if (!cic)
2166                 return NULL;
2167
2168         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2169         if (cfqq) {
2170                 sector_t sector = bio->bi_sector + bio_sectors(bio);
2171
2172                 return elv_rb_find(&cfqq->sort_list, sector);
2173         }
2174
2175         return NULL;
2176 }
2177
2178 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2179 {
2180         struct cfq_data *cfqd = q->elevator->elevator_data;
2181
2182         cfqd->rq_in_driver++;
2183         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2184                                                 cfqd->rq_in_driver);
2185
2186         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2187 }
2188
2189 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2190 {
2191         struct cfq_data *cfqd = q->elevator->elevator_data;
2192
2193         WARN_ON(!cfqd->rq_in_driver);
2194         cfqd->rq_in_driver--;
2195         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2196                                                 cfqd->rq_in_driver);
2197 }
2198
2199 static void cfq_remove_request(struct request *rq)
2200 {
2201         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2202
2203         if (cfqq->next_rq == rq)
2204                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2205
2206         list_del_init(&rq->queuelist);
2207         cfq_del_rq_rb(rq);
2208
2209         cfqq->cfqd->rq_queued--;
2210         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2211         if (rq->cmd_flags & REQ_PRIO) {
2212                 WARN_ON(!cfqq->prio_pending);
2213                 cfqq->prio_pending--;
2214         }
2215 }
2216
2217 static int cfq_merge(struct request_queue *q, struct request **req,
2218                      struct bio *bio)
2219 {
2220         struct cfq_data *cfqd = q->elevator->elevator_data;
2221         struct request *__rq;
2222
2223         __rq = cfq_find_rq_fmerge(cfqd, bio);
2224         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2225                 *req = __rq;
2226                 return ELEVATOR_FRONT_MERGE;
2227         }
2228
2229         return ELEVATOR_NO_MERGE;
2230 }
2231
2232 static void cfq_merged_request(struct request_queue *q, struct request *req,
2233                                int type)
2234 {
2235         if (type == ELEVATOR_FRONT_MERGE) {
2236                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2237
2238                 cfq_reposition_rq_rb(cfqq, req);
2239         }
2240 }
2241
2242 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2243                                 struct bio *bio)
2244 {
2245         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2246 }
2247
2248 static void
2249 cfq_merged_requests(struct request_queue *q, struct request *rq,
2250                     struct request *next)
2251 {
2252         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2253         struct cfq_data *cfqd = q->elevator->elevator_data;
2254
2255         /*
2256          * reposition in fifo if next is older than rq
2257          */
2258         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2259             time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2260             cfqq == RQ_CFQQ(next)) {
2261                 list_move(&rq->queuelist, &next->queuelist);
2262                 rq_set_fifo_time(rq, rq_fifo_time(next));
2263         }
2264
2265         if (cfqq->next_rq == next)
2266                 cfqq->next_rq = rq;
2267         cfq_remove_request(next);
2268         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2269
2270         cfqq = RQ_CFQQ(next);
2271         /*
2272          * all requests of this queue are merged to other queues, delete it
2273          * from the service tree. If it's the active_queue,
2274          * cfq_dispatch_requests() will choose to expire it or do idle
2275          */
2276         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2277             cfqq != cfqd->active_queue)
2278                 cfq_del_cfqq_rr(cfqd, cfqq);
2279 }
2280
2281 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2282                            struct bio *bio)
2283 {
2284         struct cfq_data *cfqd = q->elevator->elevator_data;
2285         struct cfq_io_cq *cic;
2286         struct cfq_queue *cfqq;
2287
2288         /*
2289          * Disallow merge of a sync bio into an async request.
2290          */
2291         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2292                 return false;
2293
2294         /*
2295          * Lookup the cfqq that this bio will be queued with and allow
2296          * merge only if rq is queued there.
2297          */
2298         cic = cfq_cic_lookup(cfqd, current->io_context);
2299         if (!cic)
2300                 return false;
2301
2302         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2303         return cfqq == RQ_CFQQ(rq);
2304 }
2305
2306 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2307 {
2308         del_timer(&cfqd->idle_slice_timer);
2309         cfqg_stats_update_idle_time(cfqq->cfqg);
2310 }
2311
2312 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2313                                    struct cfq_queue *cfqq)
2314 {
2315         if (cfqq) {
2316                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2317                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2318                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2319                 cfqq->slice_start = 0;
2320                 cfqq->dispatch_start = jiffies;
2321                 cfqq->allocated_slice = 0;
2322                 cfqq->slice_end = 0;
2323                 cfqq->slice_dispatch = 0;
2324                 cfqq->nr_sectors = 0;
2325
2326                 cfq_clear_cfqq_wait_request(cfqq);
2327                 cfq_clear_cfqq_must_dispatch(cfqq);
2328                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2329                 cfq_clear_cfqq_fifo_expire(cfqq);
2330                 cfq_mark_cfqq_slice_new(cfqq);
2331
2332                 cfq_del_timer(cfqd, cfqq);
2333         }
2334
2335         cfqd->active_queue = cfqq;
2336 }
2337
2338 /*
2339  * current cfqq expired its slice (or was too idle), select new one
2340  */
2341 static void
2342 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2343                     bool timed_out)
2344 {
2345         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2346
2347         if (cfq_cfqq_wait_request(cfqq))
2348                 cfq_del_timer(cfqd, cfqq);
2349
2350         cfq_clear_cfqq_wait_request(cfqq);
2351         cfq_clear_cfqq_wait_busy(cfqq);
2352
2353         /*
2354          * If this cfqq is shared between multiple processes, check to
2355          * make sure that those processes are still issuing I/Os within
2356          * the mean seek distance.  If not, it may be time to break the
2357          * queues apart again.
2358          */
2359         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2360                 cfq_mark_cfqq_split_coop(cfqq);
2361
2362         /*
2363          * store what was left of this slice, if the queue idled/timed out
2364          */
2365         if (timed_out) {
2366                 if (cfq_cfqq_slice_new(cfqq))
2367                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2368                 else
2369                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2370                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2371         }
2372
2373         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2374
2375         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2376                 cfq_del_cfqq_rr(cfqd, cfqq);
2377
2378         cfq_resort_rr_list(cfqd, cfqq);
2379
2380         if (cfqq == cfqd->active_queue)
2381                 cfqd->active_queue = NULL;
2382
2383         if (cfqd->active_cic) {
2384                 put_io_context(cfqd->active_cic->icq.ioc);
2385                 cfqd->active_cic = NULL;
2386         }
2387 }
2388
2389 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2390 {
2391         struct cfq_queue *cfqq = cfqd->active_queue;
2392
2393         if (cfqq)
2394                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2395 }
2396
2397 /*
2398  * Get next queue for service. Unless we have a queue preemption,
2399  * we'll simply select the first cfqq in the service tree.
2400  */
2401 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2402 {
2403         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2404                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2405
2406         if (!cfqd->rq_queued)
2407                 return NULL;
2408
2409         /* There is nothing to dispatch */
2410         if (!st)
2411                 return NULL;
2412         if (RB_EMPTY_ROOT(&st->rb))
2413                 return NULL;
2414         return cfq_rb_first(st);
2415 }
2416
2417 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2418 {
2419         struct cfq_group *cfqg;
2420         struct cfq_queue *cfqq;
2421         int i, j;
2422         struct cfq_rb_root *st;
2423
2424         if (!cfqd->rq_queued)
2425                 return NULL;
2426
2427         cfqg = cfq_get_next_cfqg(cfqd);
2428         if (!cfqg)
2429                 return NULL;
2430
2431         for_each_cfqg_st(cfqg, i, j, st)
2432                 if ((cfqq = cfq_rb_first(st)) != NULL)
2433                         return cfqq;
2434         return NULL;
2435 }
2436
2437 /*
2438  * Get and set a new active queue for service.
2439  */
2440 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2441                                               struct cfq_queue *cfqq)
2442 {
2443         if (!cfqq)
2444                 cfqq = cfq_get_next_queue(cfqd);
2445
2446         __cfq_set_active_queue(cfqd, cfqq);
2447         return cfqq;
2448 }
2449
2450 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2451                                           struct request *rq)
2452 {
2453         if (blk_rq_pos(rq) >= cfqd->last_position)
2454                 return blk_rq_pos(rq) - cfqd->last_position;
2455         else
2456                 return cfqd->last_position - blk_rq_pos(rq);
2457 }
2458
2459 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2460                                struct request *rq)
2461 {
2462         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2463 }
2464
2465 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2466                                     struct cfq_queue *cur_cfqq)
2467 {
2468         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2469         struct rb_node *parent, *node;
2470         struct cfq_queue *__cfqq;
2471         sector_t sector = cfqd->last_position;
2472
2473         if (RB_EMPTY_ROOT(root))
2474                 return NULL;
2475
2476         /*
2477          * First, if we find a request starting at the end of the last
2478          * request, choose it.
2479          */
2480         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2481         if (__cfqq)
2482                 return __cfqq;
2483
2484         /*
2485          * If the exact sector wasn't found, the parent of the NULL leaf
2486          * will contain the closest sector.
2487          */
2488         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2489         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2490                 return __cfqq;
2491
2492         if (blk_rq_pos(__cfqq->next_rq) < sector)
2493                 node = rb_next(&__cfqq->p_node);
2494         else
2495                 node = rb_prev(&__cfqq->p_node);
2496         if (!node)
2497                 return NULL;
2498
2499         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2500         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2501                 return __cfqq;
2502
2503         return NULL;
2504 }
2505
2506 /*
2507  * cfqd - obvious
2508  * cur_cfqq - passed in so that we don't decide that the current queue is
2509  *            closely cooperating with itself.
2510  *
2511  * So, basically we're assuming that that cur_cfqq has dispatched at least
2512  * one request, and that cfqd->last_position reflects a position on the disk
2513  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2514  * assumption.
2515  */
2516 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2517                                               struct cfq_queue *cur_cfqq)
2518 {
2519         struct cfq_queue *cfqq;
2520
2521         if (cfq_class_idle(cur_cfqq))
2522                 return NULL;
2523         if (!cfq_cfqq_sync(cur_cfqq))
2524                 return NULL;
2525         if (CFQQ_SEEKY(cur_cfqq))
2526                 return NULL;
2527
2528         /*
2529          * Don't search priority tree if it's the only queue in the group.
2530          */
2531         if (cur_cfqq->cfqg->nr_cfqq == 1)
2532                 return NULL;
2533
2534         /*
2535          * We should notice if some of the queues are cooperating, eg
2536          * working closely on the same area of the disk. In that case,
2537          * we can group them together and don't waste time idling.
2538          */
2539         cfqq = cfqq_close(cfqd, cur_cfqq);
2540         if (!cfqq)
2541                 return NULL;
2542
2543         /* If new queue belongs to different cfq_group, don't choose it */
2544         if (cur_cfqq->cfqg != cfqq->cfqg)
2545                 return NULL;
2546
2547         /*
2548          * It only makes sense to merge sync queues.
2549          */
2550         if (!cfq_cfqq_sync(cfqq))
2551                 return NULL;
2552         if (CFQQ_SEEKY(cfqq))
2553                 return NULL;
2554
2555         /*
2556          * Do not merge queues of different priority classes
2557          */
2558         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2559                 return NULL;
2560
2561         return cfqq;
2562 }
2563
2564 /*
2565  * Determine whether we should enforce idle window for this queue.
2566  */
2567
2568 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2569 {
2570         enum wl_class_t wl_class = cfqq_class(cfqq);
2571         struct cfq_rb_root *st = cfqq->service_tree;
2572
2573         BUG_ON(!st);
2574         BUG_ON(!st->count);
2575
2576         if (!cfqd->cfq_slice_idle)
2577                 return false;
2578
2579         /* We never do for idle class queues. */
2580         if (wl_class == IDLE_WORKLOAD)
2581                 return false;
2582
2583         /* We do for queues that were marked with idle window flag. */
2584         if (cfq_cfqq_idle_window(cfqq) &&
2585            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2586                 return true;
2587
2588         /*
2589          * Otherwise, we do only if they are the last ones
2590          * in their service tree.
2591          */
2592         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2593            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2594                 return true;
2595         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2596         return false;
2597 }
2598
2599 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2600 {
2601         struct cfq_queue *cfqq = cfqd->active_queue;
2602         struct cfq_io_cq *cic;
2603         unsigned long sl, group_idle = 0;
2604
2605         /*
2606          * SSD device without seek penalty, disable idling. But only do so
2607          * for devices that support queuing, otherwise we still have a problem
2608          * with sync vs async workloads.
2609          */
2610         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2611                 return;
2612
2613         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2614         WARN_ON(cfq_cfqq_slice_new(cfqq));
2615
2616         /*
2617          * idle is disabled, either manually or by past process history
2618          */
2619         if (!cfq_should_idle(cfqd, cfqq)) {
2620                 /* no queue idling. Check for group idling */
2621                 if (cfqd->cfq_group_idle)
2622                         group_idle = cfqd->cfq_group_idle;
2623                 else
2624                         return;
2625         }
2626
2627         /*
2628          * still active requests from this queue, don't idle
2629          */
2630         if (cfqq->dispatched)
2631                 return;
2632
2633         /*
2634          * task has exited, don't wait
2635          */
2636         cic = cfqd->active_cic;
2637         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2638                 return;
2639
2640         /*
2641          * If our average think time is larger than the remaining time
2642          * slice, then don't idle. This avoids overrunning the allotted
2643          * time slice.
2644          */
2645         if (sample_valid(cic->ttime.ttime_samples) &&
2646             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2647                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2648                              cic->ttime.ttime_mean);
2649                 return;
2650         }
2651
2652         /* There are other queues in the group, don't do group idle */
2653         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2654                 return;
2655
2656         cfq_mark_cfqq_wait_request(cfqq);
2657
2658         if (group_idle)
2659                 sl = cfqd->cfq_group_idle;
2660         else
2661                 sl = cfqd->cfq_slice_idle;
2662
2663         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2664         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2665         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2666                         group_idle ? 1 : 0);
2667 }
2668
2669 /*
2670  * Move request from internal lists to the request queue dispatch list.
2671  */
2672 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2673 {
2674         struct cfq_data *cfqd = q->elevator->elevator_data;
2675         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2676
2677         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2678
2679         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2680         cfq_remove_request(rq);
2681         cfqq->dispatched++;
2682         (RQ_CFQG(rq))->dispatched++;
2683         elv_dispatch_sort(q, rq);
2684
2685         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2686         cfqq->nr_sectors += blk_rq_sectors(rq);
2687         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2688 }
2689
2690 /*
2691  * return expired entry, or NULL to just start from scratch in rbtree
2692  */
2693 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2694 {
2695         struct request *rq = NULL;
2696
2697         if (cfq_cfqq_fifo_expire(cfqq))
2698                 return NULL;
2699
2700         cfq_mark_cfqq_fifo_expire(cfqq);
2701
2702         if (list_empty(&cfqq->fifo))
2703                 return NULL;
2704
2705         rq = rq_entry_fifo(cfqq->fifo.next);
2706         if (time_before(jiffies, rq_fifo_time(rq)))
2707                 rq = NULL;
2708
2709         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2710         return rq;
2711 }
2712
2713 static inline int
2714 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2715 {
2716         const int base_rq = cfqd->cfq_slice_async_rq;
2717
2718         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2719
2720         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2721 }
2722
2723 /*
2724  * Must be called with the queue_lock held.
2725  */
2726 static int cfqq_process_refs(struct cfq_queue *cfqq)
2727 {
2728         int process_refs, io_refs;
2729
2730         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2731         process_refs = cfqq->ref - io_refs;
2732         BUG_ON(process_refs < 0);
2733         return process_refs;
2734 }
2735
2736 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2737 {
2738         int process_refs, new_process_refs;
2739         struct cfq_queue *__cfqq;
2740
2741         /*
2742          * If there are no process references on the new_cfqq, then it is
2743          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2744          * chain may have dropped their last reference (not just their
2745          * last process reference).
2746          */
2747         if (!cfqq_process_refs(new_cfqq))
2748                 return;
2749
2750         /* Avoid a circular list and skip interim queue merges */
2751         while ((__cfqq = new_cfqq->new_cfqq)) {
2752                 if (__cfqq == cfqq)
2753                         return;
2754                 new_cfqq = __cfqq;
2755         }
2756
2757         process_refs = cfqq_process_refs(cfqq);
2758         new_process_refs = cfqq_process_refs(new_cfqq);
2759         /*
2760          * If the process for the cfqq has gone away, there is no
2761          * sense in merging the queues.
2762          */
2763         if (process_refs == 0 || new_process_refs == 0)
2764                 return;
2765
2766         /*
2767          * Merge in the direction of the lesser amount of work.
2768          */
2769         if (new_process_refs >= process_refs) {
2770                 cfqq->new_cfqq = new_cfqq;
2771                 new_cfqq->ref += process_refs;
2772         } else {
2773                 new_cfqq->new_cfqq = cfqq;
2774                 cfqq->ref += new_process_refs;
2775         }
2776 }
2777
2778 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2779                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2780 {
2781         struct cfq_queue *queue;
2782         int i;
2783         bool key_valid = false;
2784         unsigned long lowest_key = 0;
2785         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2786
2787         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2788                 /* select the one with lowest rb_key */
2789                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2790                 if (queue &&
2791                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2792                         lowest_key = queue->rb_key;
2793                         cur_best = i;
2794                         key_valid = true;
2795                 }
2796         }
2797
2798         return cur_best;
2799 }
2800
2801 static void
2802 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2803 {
2804         unsigned slice;
2805         unsigned count;
2806         struct cfq_rb_root *st;
2807         unsigned group_slice;
2808         enum wl_class_t original_class = cfqd->serving_wl_class;
2809
2810         /* Choose next priority. RT > BE > IDLE */
2811         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2812                 cfqd->serving_wl_class = RT_WORKLOAD;
2813         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2814                 cfqd->serving_wl_class = BE_WORKLOAD;
2815         else {
2816                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2817                 cfqd->workload_expires = jiffies + 1;
2818                 return;
2819         }
2820
2821         if (original_class != cfqd->serving_wl_class)
2822                 goto new_workload;
2823
2824         /*
2825          * For RT and BE, we have to choose also the type
2826          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2827          * expiration time
2828          */
2829         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2830         count = st->count;
2831
2832         /*
2833          * check workload expiration, and that we still have other queues ready
2834          */
2835         if (count && !time_after(jiffies, cfqd->workload_expires))
2836                 return;
2837
2838 new_workload:
2839         /* otherwise select new workload type */
2840         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2841                                         cfqd->serving_wl_class);
2842         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2843         count = st->count;
2844
2845         /*
2846          * the workload slice is computed as a fraction of target latency
2847          * proportional to the number of queues in that workload, over
2848          * all the queues in the same priority class
2849          */
2850         group_slice = cfq_group_slice(cfqd, cfqg);
2851
2852         slice = group_slice * count /
2853                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2854                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2855                                         cfqg));
2856
2857         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2858                 unsigned int tmp;
2859
2860                 /*
2861                  * Async queues are currently system wide. Just taking
2862                  * proportion of queues with-in same group will lead to higher
2863                  * async ratio system wide as generally root group is going
2864                  * to have higher weight. A more accurate thing would be to
2865                  * calculate system wide asnc/sync ratio.
2866                  */
2867                 tmp = cfqd->cfq_target_latency *
2868                         cfqg_busy_async_queues(cfqd, cfqg);
2869                 tmp = tmp/cfqd->busy_queues;
2870                 slice = min_t(unsigned, slice, tmp);
2871
2872                 /* async workload slice is scaled down according to
2873                  * the sync/async slice ratio. */
2874                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2875         } else
2876                 /* sync workload slice is at least 2 * cfq_slice_idle */
2877                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2878
2879         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2880         cfq_log(cfqd, "workload slice:%d", slice);
2881         cfqd->workload_expires = jiffies + slice;
2882 }
2883
2884 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2885 {
2886         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2887         struct cfq_group *cfqg;
2888
2889         if (RB_EMPTY_ROOT(&st->rb))
2890                 return NULL;
2891         cfqg = cfq_rb_first_group(st);
2892         update_min_vdisktime(st);
2893         return cfqg;
2894 }
2895
2896 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2897 {
2898         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2899
2900         cfqd->serving_group = cfqg;
2901
2902         /* Restore the workload type data */
2903         if (cfqg->saved_wl_slice) {
2904                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
2905                 cfqd->serving_wl_type = cfqg->saved_wl_type;
2906                 cfqd->serving_wl_class = cfqg->saved_wl_class;
2907         } else
2908                 cfqd->workload_expires = jiffies - 1;
2909
2910         choose_wl_class_and_type(cfqd, cfqg);
2911 }
2912
2913 /*
2914  * Select a queue for service. If we have a current active queue,
2915  * check whether to continue servicing it, or retrieve and set a new one.
2916  */
2917 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2918 {
2919         struct cfq_queue *cfqq, *new_cfqq = NULL;
2920
2921         cfqq = cfqd->active_queue;
2922         if (!cfqq)
2923                 goto new_queue;
2924
2925         if (!cfqd->rq_queued)
2926                 return NULL;
2927
2928         /*
2929          * We were waiting for group to get backlogged. Expire the queue
2930          */
2931         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2932                 goto expire;
2933
2934         /*
2935          * The active queue has run out of time, expire it and select new.
2936          */
2937         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2938                 /*
2939                  * If slice had not expired at the completion of last request
2940                  * we might not have turned on wait_busy flag. Don't expire
2941                  * the queue yet. Allow the group to get backlogged.
2942                  *
2943                  * The very fact that we have used the slice, that means we
2944                  * have been idling all along on this queue and it should be
2945                  * ok to wait for this request to complete.
2946                  */
2947                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2948                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2949                         cfqq = NULL;
2950                         goto keep_queue;
2951                 } else
2952                         goto check_group_idle;
2953         }
2954
2955         /*
2956          * The active queue has requests and isn't expired, allow it to
2957          * dispatch.
2958          */
2959         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2960                 goto keep_queue;
2961
2962         /*
2963          * If another queue has a request waiting within our mean seek
2964          * distance, let it run.  The expire code will check for close
2965          * cooperators and put the close queue at the front of the service
2966          * tree.  If possible, merge the expiring queue with the new cfqq.
2967          */
2968         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2969         if (new_cfqq) {
2970                 if (!cfqq->new_cfqq)
2971                         cfq_setup_merge(cfqq, new_cfqq);
2972                 goto expire;
2973         }
2974
2975         /*
2976          * No requests pending. If the active queue still has requests in
2977          * flight or is idling for a new request, allow either of these
2978          * conditions to happen (or time out) before selecting a new queue.
2979          */
2980         if (timer_pending(&cfqd->idle_slice_timer)) {
2981                 cfqq = NULL;
2982                 goto keep_queue;
2983         }
2984
2985         /*
2986          * This is a deep seek queue, but the device is much faster than
2987          * the queue can deliver, don't idle
2988          **/
2989         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2990             (cfq_cfqq_slice_new(cfqq) ||
2991             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2992                 cfq_clear_cfqq_deep(cfqq);
2993                 cfq_clear_cfqq_idle_window(cfqq);
2994         }
2995
2996         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2997                 cfqq = NULL;
2998                 goto keep_queue;
2999         }
3000
3001         /*
3002          * If group idle is enabled and there are requests dispatched from
3003          * this group, wait for requests to complete.
3004          */
3005 check_group_idle:
3006         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3007             cfqq->cfqg->dispatched &&
3008             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3009                 cfqq = NULL;
3010                 goto keep_queue;
3011         }
3012
3013 expire:
3014         cfq_slice_expired(cfqd, 0);
3015 new_queue:
3016         /*
3017          * Current queue expired. Check if we have to switch to a new
3018          * service tree
3019          */
3020         if (!new_cfqq)
3021                 cfq_choose_cfqg(cfqd);
3022
3023         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3024 keep_queue:
3025         return cfqq;
3026 }
3027
3028 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3029 {
3030         int dispatched = 0;
3031
3032         while (cfqq->next_rq) {
3033                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3034                 dispatched++;
3035         }
3036
3037         BUG_ON(!list_empty(&cfqq->fifo));
3038
3039         /* By default cfqq is not expired if it is empty. Do it explicitly */
3040         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3041         return dispatched;
3042 }
3043
3044 /*
3045  * Drain our current requests. Used for barriers and when switching
3046  * io schedulers on-the-fly.
3047  */
3048 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3049 {
3050         struct cfq_queue *cfqq;
3051         int dispatched = 0;
3052
3053         /* Expire the timeslice of the current active queue first */
3054         cfq_slice_expired(cfqd, 0);
3055         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3056                 __cfq_set_active_queue(cfqd, cfqq);
3057                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3058         }
3059
3060         BUG_ON(cfqd->busy_queues);
3061
3062         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3063         return dispatched;
3064 }
3065
3066 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3067         struct cfq_queue *cfqq)
3068 {
3069         /* the queue hasn't finished any request, can't estimate */
3070         if (cfq_cfqq_slice_new(cfqq))
3071                 return true;
3072         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3073                 cfqq->slice_end))
3074                 return true;
3075
3076         return false;
3077 }
3078
3079 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3080 {
3081         unsigned int max_dispatch;
3082
3083         /*
3084          * Drain async requests before we start sync IO
3085          */
3086         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3087                 return false;
3088
3089         /*
3090          * If this is an async queue and we have sync IO in flight, let it wait
3091          */
3092         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3093                 return false;
3094
3095         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3096         if (cfq_class_idle(cfqq))
3097                 max_dispatch = 1;
3098
3099         /*
3100          * Does this cfqq already have too much IO in flight?
3101          */
3102         if (cfqq->dispatched >= max_dispatch) {
3103                 bool promote_sync = false;
3104                 /*
3105                  * idle queue must always only have a single IO in flight
3106                  */
3107                 if (cfq_class_idle(cfqq))
3108                         return false;
3109
3110                 /*
3111                  * If there is only one sync queue
3112                  * we can ignore async queue here and give the sync
3113                  * queue no dispatch limit. The reason is a sync queue can
3114                  * preempt async queue, limiting the sync queue doesn't make
3115                  * sense. This is useful for aiostress test.
3116                  */
3117                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3118                         promote_sync = true;
3119
3120                 /*
3121                  * We have other queues, don't allow more IO from this one
3122                  */
3123                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3124                                 !promote_sync)
3125                         return false;
3126
3127                 /*
3128                  * Sole queue user, no limit
3129                  */
3130                 if (cfqd->busy_queues == 1 || promote_sync)
3131                         max_dispatch = -1;
3132                 else
3133                         /*
3134                          * Normally we start throttling cfqq when cfq_quantum/2
3135                          * requests have been dispatched. But we can drive
3136                          * deeper queue depths at the beginning of slice
3137                          * subjected to upper limit of cfq_quantum.
3138                          * */
3139                         max_dispatch = cfqd->cfq_quantum;
3140         }
3141
3142         /*
3143          * Async queues must wait a bit before being allowed dispatch.
3144          * We also ramp up the dispatch depth gradually for async IO,
3145          * based on the last sync IO we serviced
3146          */
3147         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3148                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3149                 unsigned int depth;
3150
3151                 depth = last_sync / cfqd->cfq_slice[1];
3152                 if (!depth && !cfqq->dispatched)
3153                         depth = 1;
3154                 if (depth < max_dispatch)
3155                         max_dispatch = depth;
3156         }
3157
3158         /*
3159          * If we're below the current max, allow a dispatch
3160          */
3161         return cfqq->dispatched < max_dispatch;
3162 }
3163
3164 /*
3165  * Dispatch a request from cfqq, moving them to the request queue
3166  * dispatch list.
3167  */
3168 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3169 {
3170         struct request *rq;
3171
3172         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3173
3174         if (!cfq_may_dispatch(cfqd, cfqq))
3175                 return false;
3176
3177         /*
3178          * follow expired path, else get first next available
3179          */
3180         rq = cfq_check_fifo(cfqq);
3181         if (!rq)
3182                 rq = cfqq->next_rq;
3183
3184         /*
3185          * insert request into driver dispatch list
3186          */
3187         cfq_dispatch_insert(cfqd->queue, rq);
3188
3189         if (!cfqd->active_cic) {
3190                 struct cfq_io_cq *cic = RQ_CIC(rq);
3191
3192                 atomic_long_inc(&cic->icq.ioc->refcount);
3193                 cfqd->active_cic = cic;
3194         }
3195
3196         return true;
3197 }
3198
3199 /*
3200  * Find the cfqq that we need to service and move a request from that to the
3201  * dispatch list
3202  */
3203 static int cfq_dispatch_requests(struct request_queue *q, int force)
3204 {
3205         struct cfq_data *cfqd = q->elevator->elevator_data;
3206         struct cfq_queue *cfqq;
3207
3208         if (!cfqd->busy_queues)
3209                 return 0;
3210
3211         if (unlikely(force))
3212                 return cfq_forced_dispatch(cfqd);
3213
3214         cfqq = cfq_select_queue(cfqd);
3215         if (!cfqq)
3216                 return 0;
3217
3218         /*
3219          * Dispatch a request from this cfqq, if it is allowed
3220          */
3221         if (!cfq_dispatch_request(cfqd, cfqq))
3222                 return 0;
3223
3224         cfqq->slice_dispatch++;
3225         cfq_clear_cfqq_must_dispatch(cfqq);
3226
3227         /*
3228          * expire an async queue immediately if it has used up its slice. idle
3229          * queue always expire after 1 dispatch round.
3230          */
3231         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3232             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3233             cfq_class_idle(cfqq))) {
3234                 cfqq->slice_end = jiffies + 1;
3235                 cfq_slice_expired(cfqd, 0);
3236         }
3237
3238         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3239         return 1;
3240 }
3241
3242 /*
3243  * task holds one reference to the queue, dropped when task exits. each rq
3244  * in-flight on this queue also holds a reference, dropped when rq is freed.
3245  *
3246  * Each cfq queue took a reference on the parent group. Drop it now.
3247  * queue lock must be held here.
3248  */
3249 static void cfq_put_queue(struct cfq_queue *cfqq)
3250 {
3251         struct cfq_data *cfqd = cfqq->cfqd;
3252         struct cfq_group *cfqg;
3253
3254         BUG_ON(cfqq->ref <= 0);
3255
3256         cfqq->ref--;
3257         if (cfqq->ref)
3258                 return;
3259
3260         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3261         BUG_ON(rb_first(&cfqq->sort_list));
3262         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3263         cfqg = cfqq->cfqg;
3264
3265         if (unlikely(cfqd->active_queue == cfqq)) {
3266                 __cfq_slice_expired(cfqd, cfqq, 0);
3267                 cfq_schedule_dispatch(cfqd);
3268         }
3269
3270         BUG_ON(cfq_cfqq_on_rr(cfqq));
3271         kmem_cache_free(cfq_pool, cfqq);
3272         cfqg_put(cfqg);
3273 }
3274
3275 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3276 {
3277         struct cfq_queue *__cfqq, *next;
3278
3279         /*
3280          * If this queue was scheduled to merge with another queue, be
3281          * sure to drop the reference taken on that queue (and others in
3282          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3283          */
3284         __cfqq = cfqq->new_cfqq;
3285         while (__cfqq) {
3286                 if (__cfqq == cfqq) {
3287                         WARN(1, "cfqq->new_cfqq loop detected\n");
3288                         break;
3289                 }
3290                 next = __cfqq->new_cfqq;
3291                 cfq_put_queue(__cfqq);
3292                 __cfqq = next;
3293         }
3294 }
3295
3296 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3297 {
3298         if (unlikely(cfqq == cfqd->active_queue)) {
3299                 __cfq_slice_expired(cfqd, cfqq, 0);
3300                 cfq_schedule_dispatch(cfqd);
3301         }
3302
3303         cfq_put_cooperator(cfqq);
3304
3305         cfq_put_queue(cfqq);
3306 }
3307
3308 static void cfq_init_icq(struct io_cq *icq)
3309 {
3310         struct cfq_io_cq *cic = icq_to_cic(icq);
3311
3312         cic->ttime.last_end_request = jiffies;
3313 }
3314
3315 static void cfq_exit_icq(struct io_cq *icq)
3316 {
3317         struct cfq_io_cq *cic = icq_to_cic(icq);
3318         struct cfq_data *cfqd = cic_to_cfqd(cic);
3319
3320         if (cic->cfqq[BLK_RW_ASYNC]) {
3321                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3322                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3323         }
3324
3325         if (cic->cfqq[BLK_RW_SYNC]) {
3326                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3327                 cic->cfqq[BLK_RW_SYNC] = NULL;
3328         }
3329 }
3330
3331 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3332 {
3333         struct task_struct *tsk = current;
3334         int ioprio_class;
3335
3336         if (!cfq_cfqq_prio_changed(cfqq))
3337                 return;
3338
3339         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3340         switch (ioprio_class) {
3341         default:
3342                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3343         case IOPRIO_CLASS_NONE:
3344                 /*
3345                  * no prio set, inherit CPU scheduling settings
3346                  */
3347                 cfqq->ioprio = task_nice_ioprio(tsk);
3348                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3349                 break;
3350         case IOPRIO_CLASS_RT:
3351                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3352                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3353                 break;
3354         case IOPRIO_CLASS_BE:
3355                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3356                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3357                 break;
3358         case IOPRIO_CLASS_IDLE:
3359                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3360                 cfqq->ioprio = 7;
3361                 cfq_clear_cfqq_idle_window(cfqq);
3362                 break;
3363         }
3364
3365         /*
3366          * keep track of original prio settings in case we have to temporarily
3367          * elevate the priority of this queue
3368          */
3369         cfqq->org_ioprio = cfqq->ioprio;
3370         cfq_clear_cfqq_prio_changed(cfqq);
3371 }
3372
3373 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3374 {
3375         int ioprio = cic->icq.ioc->ioprio;
3376         struct cfq_data *cfqd = cic_to_cfqd(cic);
3377         struct cfq_queue *cfqq;
3378
3379         /*
3380          * Check whether ioprio has changed.  The condition may trigger
3381          * spuriously on a newly created cic but there's no harm.
3382          */
3383         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3384                 return;
3385
3386         cfqq = cic->cfqq[BLK_RW_ASYNC];
3387         if (cfqq) {
3388                 struct cfq_queue *new_cfqq;
3389                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3390                                          GFP_ATOMIC);
3391                 if (new_cfqq) {
3392                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3393                         cfq_put_queue(cfqq);
3394                 }
3395         }
3396
3397         cfqq = cic->cfqq[BLK_RW_SYNC];
3398         if (cfqq)
3399                 cfq_mark_cfqq_prio_changed(cfqq);
3400
3401         cic->ioprio = ioprio;
3402 }
3403
3404 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3405                           pid_t pid, bool is_sync)
3406 {
3407         RB_CLEAR_NODE(&cfqq->rb_node);
3408         RB_CLEAR_NODE(&cfqq->p_node);
3409         INIT_LIST_HEAD(&cfqq->fifo);
3410
3411         cfqq->ref = 0;
3412         cfqq->cfqd = cfqd;
3413
3414         cfq_mark_cfqq_prio_changed(cfqq);
3415
3416         if (is_sync) {
3417                 if (!cfq_class_idle(cfqq))
3418                         cfq_mark_cfqq_idle_window(cfqq);
3419                 cfq_mark_cfqq_sync(cfqq);
3420         }
3421         cfqq->pid = pid;
3422 }
3423
3424 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3425 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3426 {
3427         struct cfq_data *cfqd = cic_to_cfqd(cic);
3428         struct cfq_queue *sync_cfqq;
3429         uint64_t id;
3430
3431         rcu_read_lock();
3432         id = bio_blkcg(bio)->id;
3433         rcu_read_unlock();
3434
3435         /*
3436          * Check whether blkcg has changed.  The condition may trigger
3437          * spuriously on a newly created cic but there's no harm.
3438          */
3439         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3440                 return;
3441
3442         sync_cfqq = cic_to_cfqq(cic, 1);
3443         if (sync_cfqq) {
3444                 /*
3445                  * Drop reference to sync queue. A new sync queue will be
3446                  * assigned in new group upon arrival of a fresh request.
3447                  */
3448                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3449                 cic_set_cfqq(cic, NULL, 1);
3450                 cfq_put_queue(sync_cfqq);
3451         }
3452
3453         cic->blkcg_id = id;
3454 }
3455 #else
3456 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3457 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3458
3459 static struct cfq_queue *
3460 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3461                      struct bio *bio, gfp_t gfp_mask)
3462 {
3463         struct blkcg *blkcg;
3464         struct cfq_queue *cfqq, *new_cfqq = NULL;
3465         struct cfq_group *cfqg;
3466
3467 retry:
3468         rcu_read_lock();
3469
3470         blkcg = bio_blkcg(bio);
3471         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3472         cfqq = cic_to_cfqq(cic, is_sync);
3473
3474         /*
3475          * Always try a new alloc if we fell back to the OOM cfqq
3476          * originally, since it should just be a temporary situation.
3477          */
3478         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3479                 cfqq = NULL;
3480                 if (new_cfqq) {
3481                         cfqq = new_cfqq;
3482                         new_cfqq = NULL;
3483                 } else if (gfp_mask & __GFP_WAIT) {
3484                         rcu_read_unlock();
3485                         spin_unlock_irq(cfqd->queue->queue_lock);
3486                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3487                                         gfp_mask | __GFP_ZERO,
3488                                         cfqd->queue->node);
3489                         spin_lock_irq(cfqd->queue->queue_lock);
3490                         if (new_cfqq)
3491                                 goto retry;
3492                 } else {
3493                         cfqq = kmem_cache_alloc_node(cfq_pool,
3494                                         gfp_mask | __GFP_ZERO,
3495                                         cfqd->queue->node);
3496                 }
3497
3498                 if (cfqq) {
3499                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3500                         cfq_init_prio_data(cfqq, cic);
3501                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3502                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3503                 } else
3504                         cfqq = &cfqd->oom_cfqq;
3505         }
3506
3507         if (new_cfqq)
3508                 kmem_cache_free(cfq_pool, new_cfqq);
3509
3510         rcu_read_unlock();
3511         return cfqq;
3512 }
3513
3514 static struct cfq_queue **
3515 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3516 {
3517         switch (ioprio_class) {
3518         case IOPRIO_CLASS_RT:
3519                 return &cfqd->async_cfqq[0][ioprio];
3520         case IOPRIO_CLASS_NONE:
3521                 ioprio = IOPRIO_NORM;
3522                 /* fall through */
3523         case IOPRIO_CLASS_BE:
3524                 return &cfqd->async_cfqq[1][ioprio];
3525         case IOPRIO_CLASS_IDLE:
3526                 return &cfqd->async_idle_cfqq;
3527         default:
3528                 BUG();
3529         }
3530 }
3531
3532 static struct cfq_queue *
3533 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3534               struct bio *bio, gfp_t gfp_mask)
3535 {
3536         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3537         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3538         struct cfq_queue **async_cfqq = NULL;
3539         struct cfq_queue *cfqq = NULL;
3540
3541         if (!is_sync) {
3542                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3543                 cfqq = *async_cfqq;
3544         }
3545
3546         if (!cfqq)
3547                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3548
3549         /*
3550          * pin the queue now that it's allocated, scheduler exit will prune it
3551          */
3552         if (!is_sync && !(*async_cfqq)) {
3553                 cfqq->ref++;
3554                 *async_cfqq = cfqq;
3555         }
3556
3557         cfqq->ref++;
3558         return cfqq;
3559 }
3560
3561 static void
3562 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3563 {
3564         unsigned long elapsed = jiffies - ttime->last_end_request;
3565         elapsed = min(elapsed, 2UL * slice_idle);
3566
3567         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3568         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3569         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3570 }
3571
3572 static void
3573 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3574                         struct cfq_io_cq *cic)
3575 {
3576         if (cfq_cfqq_sync(cfqq)) {
3577                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3578                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3579                         cfqd->cfq_slice_idle);
3580         }
3581 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3582         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3583 #endif
3584 }
3585
3586 static void
3587 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3588                        struct request *rq)
3589 {
3590         sector_t sdist = 0;
3591         sector_t n_sec = blk_rq_sectors(rq);
3592         if (cfqq->last_request_pos) {
3593                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3594                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3595                 else
3596                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3597         }
3598
3599         cfqq->seek_history <<= 1;
3600         if (blk_queue_nonrot(cfqd->queue))
3601                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3602         else
3603                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3604 }
3605
3606 /*
3607  * Disable idle window if the process thinks too long or seeks so much that
3608  * it doesn't matter
3609  */
3610 static void
3611 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3612                        struct cfq_io_cq *cic)
3613 {
3614         int old_idle, enable_idle;
3615
3616         /*
3617          * Don't idle for async or idle io prio class
3618          */
3619         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3620                 return;
3621
3622         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3623
3624         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3625                 cfq_mark_cfqq_deep(cfqq);
3626
3627         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3628                 enable_idle = 0;
3629         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3630                  !cfqd->cfq_slice_idle ||
3631                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3632                 enable_idle = 0;
3633         else if (sample_valid(cic->ttime.ttime_samples)) {
3634                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3635                         enable_idle = 0;
3636                 else
3637                         enable_idle = 1;
3638         }
3639
3640         if (old_idle != enable_idle) {
3641                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3642                 if (enable_idle)
3643                         cfq_mark_cfqq_idle_window(cfqq);
3644                 else
3645                         cfq_clear_cfqq_idle_window(cfqq);
3646         }
3647 }
3648
3649 /*
3650  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3651  * no or if we aren't sure, a 1 will cause a preempt.
3652  */
3653 static bool
3654 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3655                    struct request *rq)
3656 {
3657         struct cfq_queue *cfqq;
3658
3659         cfqq = cfqd->active_queue;
3660         if (!cfqq)
3661                 return false;
3662
3663         if (cfq_class_idle(new_cfqq))
3664                 return false;
3665
3666         if (cfq_class_idle(cfqq))
3667                 return true;
3668
3669         /*
3670          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3671          */
3672         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3673                 return false;
3674
3675         /*
3676          * if the new request is sync, but the currently running queue is
3677          * not, let the sync request have priority.
3678          */
3679         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3680                 return true;
3681
3682         if (new_cfqq->cfqg != cfqq->cfqg)
3683                 return false;
3684
3685         if (cfq_slice_used(cfqq))
3686                 return true;
3687
3688         /* Allow preemption only if we are idling on sync-noidle tree */
3689         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3690             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3691             new_cfqq->service_tree->count == 2 &&
3692             RB_EMPTY_ROOT(&cfqq->sort_list))
3693                 return true;
3694
3695         /*
3696          * So both queues are sync. Let the new request get disk time if
3697          * it's a metadata request and the current queue is doing regular IO.
3698          */
3699         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3700                 return true;
3701
3702         /*
3703          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3704          */
3705         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3706                 return true;
3707
3708         /* An idle queue should not be idle now for some reason */
3709         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3710                 return true;
3711
3712         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3713                 return false;
3714
3715         /*
3716          * if this request is as-good as one we would expect from the
3717          * current cfqq, let it preempt
3718          */
3719         if (cfq_rq_close(cfqd, cfqq, rq))
3720                 return true;
3721
3722         return false;
3723 }
3724
3725 /*
3726  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3727  * let it have half of its nominal slice.
3728  */
3729 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3730 {
3731         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3732
3733         cfq_log_cfqq(cfqd, cfqq, "preempt");
3734         cfq_slice_expired(cfqd, 1);
3735
3736         /*
3737          * workload type is changed, don't save slice, otherwise preempt
3738          * doesn't happen
3739          */
3740         if (old_type != cfqq_type(cfqq))
3741                 cfqq->cfqg->saved_wl_slice = 0;
3742
3743         /*
3744          * Put the new queue at the front of the of the current list,
3745          * so we know that it will be selected next.
3746          */
3747         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3748
3749         cfq_service_tree_add(cfqd, cfqq, 1);
3750
3751         cfqq->slice_end = 0;
3752         cfq_mark_cfqq_slice_new(cfqq);
3753 }
3754
3755 /*
3756  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3757  * something we should do about it
3758  */
3759 static void
3760 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3761                 struct request *rq)
3762 {
3763         struct cfq_io_cq *cic = RQ_CIC(rq);
3764
3765         cfqd->rq_queued++;
3766         if (rq->cmd_flags & REQ_PRIO)
3767                 cfqq->prio_pending++;
3768
3769         cfq_update_io_thinktime(cfqd, cfqq, cic);
3770         cfq_update_io_seektime(cfqd, cfqq, rq);
3771         cfq_update_idle_window(cfqd, cfqq, cic);
3772
3773         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3774
3775         if (cfqq == cfqd->active_queue) {
3776                 /*
3777                  * Remember that we saw a request from this process, but
3778                  * don't start queuing just yet. Otherwise we risk seeing lots
3779                  * of tiny requests, because we disrupt the normal plugging
3780                  * and merging. If the request is already larger than a single
3781                  * page, let it rip immediately. For that case we assume that
3782                  * merging is already done. Ditto for a busy system that
3783                  * has other work pending, don't risk delaying until the
3784                  * idle timer unplug to continue working.
3785                  */
3786                 if (cfq_cfqq_wait_request(cfqq)) {
3787                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3788                             cfqd->busy_queues > 1) {
3789                                 cfq_del_timer(cfqd, cfqq);
3790                                 cfq_clear_cfqq_wait_request(cfqq);
3791                                 __blk_run_queue(cfqd->queue);
3792                         } else {
3793                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3794                                 cfq_mark_cfqq_must_dispatch(cfqq);
3795                         }
3796                 }
3797         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3798                 /*
3799                  * not the active queue - expire current slice if it is
3800                  * idle and has expired it's mean thinktime or this new queue
3801                  * has some old slice time left and is of higher priority or
3802                  * this new queue is RT and the current one is BE
3803                  */
3804                 cfq_preempt_queue(cfqd, cfqq);
3805                 __blk_run_queue(cfqd->queue);
3806         }
3807 }
3808
3809 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3810 {
3811         struct cfq_data *cfqd = q->elevator->elevator_data;
3812         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3813
3814         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3815         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3816
3817         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3818         list_add_tail(&rq->queuelist, &cfqq->fifo);
3819         cfq_add_rq_rb(rq);
3820         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3821                                  rq->cmd_flags);
3822         cfq_rq_enqueued(cfqd, cfqq, rq);
3823 }
3824
3825 /*
3826  * Update hw_tag based on peak queue depth over 50 samples under
3827  * sufficient load.
3828  */
3829 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3830 {
3831         struct cfq_queue *cfqq = cfqd->active_queue;
3832
3833         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3834                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3835
3836         if (cfqd->hw_tag == 1)
3837                 return;
3838
3839         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3840             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3841                 return;
3842
3843         /*
3844          * If active queue hasn't enough requests and can idle, cfq might not
3845          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3846          * case
3847          */
3848         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3849             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3850             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3851                 return;
3852
3853         if (cfqd->hw_tag_samples++ < 50)
3854                 return;
3855
3856         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3857                 cfqd->hw_tag = 1;
3858         else
3859                 cfqd->hw_tag = 0;
3860 }
3861
3862 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3863 {
3864         struct cfq_io_cq *cic = cfqd->active_cic;
3865
3866         /* If the queue already has requests, don't wait */
3867         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3868                 return false;
3869
3870         /* If there are other queues in the group, don't wait */
3871         if (cfqq->cfqg->nr_cfqq > 1)
3872                 return false;
3873
3874         /* the only queue in the group, but think time is big */
3875         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3876                 return false;
3877
3878         if (cfq_slice_used(cfqq))
3879                 return true;
3880
3881         /* if slice left is less than think time, wait busy */
3882         if (cic && sample_valid(cic->ttime.ttime_samples)
3883             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3884                 return true;
3885
3886         /*
3887          * If think times is less than a jiffy than ttime_mean=0 and above
3888          * will not be true. It might happen that slice has not expired yet
3889          * but will expire soon (4-5 ns) during select_queue(). To cover the
3890          * case where think time is less than a jiffy, mark the queue wait
3891          * busy if only 1 jiffy is left in the slice.
3892          */
3893         if (cfqq->slice_end - jiffies == 1)
3894                 return true;
3895
3896         return false;
3897 }
3898
3899 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3900 {
3901         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3902         struct cfq_data *cfqd = cfqq->cfqd;
3903         const int sync = rq_is_sync(rq);
3904         unsigned long now;
3905
3906         now = jiffies;
3907         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3908                      !!(rq->cmd_flags & REQ_NOIDLE));
3909
3910         cfq_update_hw_tag(cfqd);
3911
3912         WARN_ON(!cfqd->rq_in_driver);
3913         WARN_ON(!cfqq->dispatched);
3914         cfqd->rq_in_driver--;
3915         cfqq->dispatched--;
3916         (RQ_CFQG(rq))->dispatched--;
3917         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3918                                      rq_io_start_time_ns(rq), rq->cmd_flags);
3919
3920         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3921
3922         if (sync) {
3923                 struct cfq_rb_root *st;
3924
3925                 RQ_CIC(rq)->ttime.last_end_request = now;
3926
3927                 if (cfq_cfqq_on_rr(cfqq))
3928                         st = cfqq->service_tree;
3929                 else
3930                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
3931                                         cfqq_type(cfqq));
3932
3933                 st->ttime.last_end_request = now;
3934                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3935                         cfqd->last_delayed_sync = now;
3936         }
3937
3938 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3939         cfqq->cfqg->ttime.last_end_request = now;
3940 #endif
3941
3942         /*
3943          * If this is the active queue, check if it needs to be expired,
3944          * or if we want to idle in case it has no pending requests.
3945          */
3946         if (cfqd->active_queue == cfqq) {
3947                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3948
3949                 if (cfq_cfqq_slice_new(cfqq)) {
3950                         cfq_set_prio_slice(cfqd, cfqq);
3951                         cfq_clear_cfqq_slice_new(cfqq);
3952                 }
3953
3954                 /*
3955                  * Should we wait for next request to come in before we expire
3956                  * the queue.
3957                  */
3958                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3959                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3960                         if (!cfqd->cfq_slice_idle)
3961                                 extend_sl = cfqd->cfq_group_idle;
3962                         cfqq->slice_end = jiffies + extend_sl;
3963                         cfq_mark_cfqq_wait_busy(cfqq);
3964                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3965                 }
3966
3967                 /*
3968                  * Idling is not enabled on:
3969                  * - expired queues
3970                  * - idle-priority queues
3971                  * - async queues
3972                  * - queues with still some requests queued
3973                  * - when there is a close cooperator
3974                  */
3975                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3976                         cfq_slice_expired(cfqd, 1);
3977                 else if (sync && cfqq_empty &&
3978                          !cfq_close_cooperator(cfqd, cfqq)) {
3979                         cfq_arm_slice_timer(cfqd);
3980                 }
3981         }
3982
3983         if (!cfqd->rq_in_driver)
3984                 cfq_schedule_dispatch(cfqd);
3985 }
3986
3987 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3988 {
3989         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3990                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3991                 return ELV_MQUEUE_MUST;
3992         }
3993
3994         return ELV_MQUEUE_MAY;
3995 }
3996
3997 static int cfq_may_queue(struct request_queue *q, int rw)
3998 {
3999         struct cfq_data *cfqd = q->elevator->elevator_data;
4000         struct task_struct *tsk = current;
4001         struct cfq_io_cq *cic;
4002         struct cfq_queue *cfqq;
4003
4004         /*
4005          * don't force setup of a queue from here, as a call to may_queue
4006          * does not necessarily imply that a request actually will be queued.
4007          * so just lookup a possibly existing queue, or return 'may queue'
4008          * if that fails
4009          */
4010         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4011         if (!cic)
4012                 return ELV_MQUEUE_MAY;
4013
4014         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4015         if (cfqq) {
4016                 cfq_init_prio_data(cfqq, cic);
4017
4018                 return __cfq_may_queue(cfqq);
4019         }
4020
4021         return ELV_MQUEUE_MAY;
4022 }
4023
4024 /*
4025  * queue lock held here
4026  */
4027 static void cfq_put_request(struct request *rq)
4028 {
4029         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4030
4031         if (cfqq) {
4032                 const int rw = rq_data_dir(rq);
4033
4034                 BUG_ON(!cfqq->allocated[rw]);
4035                 cfqq->allocated[rw]--;
4036
4037                 /* Put down rq reference on cfqg */
4038                 cfqg_put(RQ_CFQG(rq));
4039                 rq->elv.priv[0] = NULL;
4040                 rq->elv.priv[1] = NULL;
4041
4042                 cfq_put_queue(cfqq);
4043         }
4044 }
4045
4046 static struct cfq_queue *
4047 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4048                 struct cfq_queue *cfqq)
4049 {
4050         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4051         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4052         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4053         cfq_put_queue(cfqq);
4054         return cic_to_cfqq(cic, 1);
4055 }
4056
4057 /*
4058  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4059  * was the last process referring to said cfqq.
4060  */
4061 static struct cfq_queue *
4062 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4063 {
4064         if (cfqq_process_refs(cfqq) == 1) {
4065                 cfqq->pid = current->pid;
4066                 cfq_clear_cfqq_coop(cfqq);
4067                 cfq_clear_cfqq_split_coop(cfqq);
4068                 return cfqq;
4069         }
4070
4071         cic_set_cfqq(cic, NULL, 1);
4072
4073         cfq_put_cooperator(cfqq);
4074
4075         cfq_put_queue(cfqq);
4076         return NULL;
4077 }
4078 /*
4079  * Allocate cfq data structures associated with this request.
4080  */
4081 static int
4082 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4083                 gfp_t gfp_mask)
4084 {
4085         struct cfq_data *cfqd = q->elevator->elevator_data;
4086         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4087         const int rw = rq_data_dir(rq);
4088         const bool is_sync = rq_is_sync(rq);
4089         struct cfq_queue *cfqq;
4090
4091         might_sleep_if(gfp_mask & __GFP_WAIT);
4092
4093         spin_lock_irq(q->queue_lock);
4094
4095         check_ioprio_changed(cic, bio);
4096         check_blkcg_changed(cic, bio);
4097 new_queue:
4098         cfqq = cic_to_cfqq(cic, is_sync);
4099         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4100                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4101                 cic_set_cfqq(cic, cfqq, is_sync);
4102         } else {
4103                 /*
4104                  * If the queue was seeky for too long, break it apart.
4105                  */
4106                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4107                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4108                         cfqq = split_cfqq(cic, cfqq);
4109                         if (!cfqq)
4110                                 goto new_queue;
4111                 }
4112
4113                 /*
4114                  * Check to see if this queue is scheduled to merge with
4115                  * another, closely cooperating queue.  The merging of
4116                  * queues happens here as it must be done in process context.
4117                  * The reference on new_cfqq was taken in merge_cfqqs.
4118                  */
4119                 if (cfqq->new_cfqq)
4120                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4121         }
4122
4123         cfqq->allocated[rw]++;
4124
4125         cfqq->ref++;
4126         cfqg_get(cfqq->cfqg);
4127         rq->elv.priv[0] = cfqq;
4128         rq->elv.priv[1] = cfqq->cfqg;
4129         spin_unlock_irq(q->queue_lock);
4130         return 0;
4131 }
4132
4133 static void cfq_kick_queue(struct work_struct *work)
4134 {
4135         struct cfq_data *cfqd =
4136                 container_of(work, struct cfq_data, unplug_work);
4137         struct request_queue *q = cfqd->queue;
4138
4139         spin_lock_irq(q->queue_lock);
4140         __blk_run_queue(cfqd->queue);
4141         spin_unlock_irq(q->queue_lock);
4142 }
4143
4144 /*
4145  * Timer running if the active_queue is currently idling inside its time slice
4146  */
4147 static void cfq_idle_slice_timer(unsigned long data)
4148 {
4149         struct cfq_data *cfqd = (struct cfq_data *) data;
4150         struct cfq_queue *cfqq;
4151         unsigned long flags;
4152         int timed_out = 1;
4153
4154         cfq_log(cfqd, "idle timer fired");
4155
4156         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4157
4158         cfqq = cfqd->active_queue;
4159         if (cfqq) {
4160                 timed_out = 0;
4161
4162                 /*
4163                  * We saw a request before the queue expired, let it through
4164                  */
4165                 if (cfq_cfqq_must_dispatch(cfqq))
4166                         goto out_kick;
4167
4168                 /*
4169                  * expired
4170                  */
4171                 if (cfq_slice_used(cfqq))
4172                         goto expire;
4173
4174                 /*
4175                  * only expire and reinvoke request handler, if there are
4176                  * other queues with pending requests
4177                  */
4178                 if (!cfqd->busy_queues)
4179                         goto out_cont;
4180
4181                 /*
4182                  * not expired and it has a request pending, let it dispatch
4183                  */
4184                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4185                         goto out_kick;
4186
4187                 /*
4188                  * Queue depth flag is reset only when the idle didn't succeed
4189                  */
4190                 cfq_clear_cfqq_deep(cfqq);
4191         }
4192 expire:
4193         cfq_slice_expired(cfqd, timed_out);
4194 out_kick:
4195         cfq_schedule_dispatch(cfqd);
4196 out_cont:
4197         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4198 }
4199
4200 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4201 {
4202         del_timer_sync(&cfqd->idle_slice_timer);
4203         cancel_work_sync(&cfqd->unplug_work);
4204 }
4205
4206 static void cfq_put_async_queues(struct cfq_data *cfqd)
4207 {
4208         int i;
4209
4210         for (i = 0; i < IOPRIO_BE_NR; i++) {
4211                 if (cfqd->async_cfqq[0][i])
4212                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4213                 if (cfqd->async_cfqq[1][i])
4214                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4215         }
4216
4217         if (cfqd->async_idle_cfqq)
4218                 cfq_put_queue(cfqd->async_idle_cfqq);
4219 }
4220
4221 static void cfq_exit_queue(struct elevator_queue *e)
4222 {
4223         struct cfq_data *cfqd = e->elevator_data;
4224         struct request_queue *q = cfqd->queue;
4225
4226         cfq_shutdown_timer_wq(cfqd);
4227
4228         spin_lock_irq(q->queue_lock);
4229
4230         if (cfqd->active_queue)
4231                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4232
4233         cfq_put_async_queues(cfqd);
4234
4235         spin_unlock_irq(q->queue_lock);
4236
4237         cfq_shutdown_timer_wq(cfqd);
4238
4239 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4240         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4241 #else
4242         kfree(cfqd->root_group);
4243 #endif
4244         kfree(cfqd);
4245 }
4246
4247 static int cfq_init_queue(struct request_queue *q)
4248 {
4249         struct cfq_data *cfqd;
4250         struct blkcg_gq *blkg __maybe_unused;
4251         int i, ret;
4252
4253         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4254         if (!cfqd)
4255                 return -ENOMEM;
4256
4257         cfqd->queue = q;
4258         q->elevator->elevator_data = cfqd;
4259
4260         /* Init root service tree */
4261         cfqd->grp_service_tree = CFQ_RB_ROOT;
4262
4263         /* Init root group and prefer root group over other groups by default */
4264 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4265         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4266         if (ret)
4267                 goto out_free;
4268
4269         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4270 #else
4271         ret = -ENOMEM;
4272         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4273                                         GFP_KERNEL, cfqd->queue->node);
4274         if (!cfqd->root_group)
4275                 goto out_free;
4276
4277         cfq_init_cfqg_base(cfqd->root_group);
4278 #endif
4279         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4280         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4281
4282         /*
4283          * Not strictly needed (since RB_ROOT just clears the node and we
4284          * zeroed cfqd on alloc), but better be safe in case someone decides
4285          * to add magic to the rb code
4286          */
4287         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4288                 cfqd->prio_trees[i] = RB_ROOT;
4289
4290         /*
4291          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4292          * Grab a permanent reference to it, so that the normal code flow
4293          * will not attempt to free it.  oom_cfqq is linked to root_group
4294          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4295          * the reference from linking right away.
4296          */
4297         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4298         cfqd->oom_cfqq.ref++;
4299
4300         spin_lock_irq(q->queue_lock);
4301         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4302         cfqg_put(cfqd->root_group);
4303         spin_unlock_irq(q->queue_lock);
4304
4305         init_timer(&cfqd->idle_slice_timer);
4306         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4307         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4308
4309         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4310
4311         cfqd->cfq_quantum = cfq_quantum;
4312         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4313         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4314         cfqd->cfq_back_max = cfq_back_max;
4315         cfqd->cfq_back_penalty = cfq_back_penalty;
4316         cfqd->cfq_slice[0] = cfq_slice_async;
4317         cfqd->cfq_slice[1] = cfq_slice_sync;
4318         cfqd->cfq_target_latency = cfq_target_latency;
4319         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4320         cfqd->cfq_slice_idle = cfq_slice_idle;
4321         cfqd->cfq_group_idle = cfq_group_idle;
4322         cfqd->cfq_latency = 1;
4323         cfqd->hw_tag = -1;
4324         /*
4325          * we optimistically start assuming sync ops weren't delayed in last
4326          * second, in order to have larger depth for async operations.
4327          */
4328         cfqd->last_delayed_sync = jiffies - HZ;
4329         return 0;
4330
4331 out_free:
4332         kfree(cfqd);
4333         return ret;
4334 }
4335
4336 /*
4337  * sysfs parts below -->
4338  */
4339 static ssize_t
4340 cfq_var_show(unsigned int var, char *page)
4341 {
4342         return sprintf(page, "%d\n", var);
4343 }
4344
4345 static ssize_t
4346 cfq_var_store(unsigned int *var, const char *page, size_t count)
4347 {
4348         char *p = (char *) page;
4349
4350         *var = simple_strtoul(p, &p, 10);
4351         return count;
4352 }
4353
4354 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4355 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4356 {                                                                       \
4357         struct cfq_data *cfqd = e->elevator_data;                       \
4358         unsigned int __data = __VAR;                                    \
4359         if (__CONV)                                                     \
4360                 __data = jiffies_to_msecs(__data);                      \
4361         return cfq_var_show(__data, (page));                            \
4362 }
4363 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4364 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4365 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4366 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4367 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4368 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4369 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4370 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4371 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4372 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4373 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4374 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4375 #undef SHOW_FUNCTION
4376
4377 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4378 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4379 {                                                                       \
4380         struct cfq_data *cfqd = e->elevator_data;                       \
4381         unsigned int __data;                                            \
4382         int ret = cfq_var_store(&__data, (page), count);                \
4383         if (__data < (MIN))                                             \
4384                 __data = (MIN);                                         \
4385         else if (__data > (MAX))                                        \
4386                 __data = (MAX);                                         \
4387         if (__CONV)                                                     \
4388                 *(__PTR) = msecs_to_jiffies(__data);                    \
4389         else                                                            \
4390                 *(__PTR) = __data;                                      \
4391         return ret;                                                     \
4392 }
4393 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4394 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4395                 UINT_MAX, 1);
4396 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4397                 UINT_MAX, 1);
4398 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4399 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4400                 UINT_MAX, 0);
4401 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4402 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4403 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4404 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4405 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4406                 UINT_MAX, 0);
4407 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4408 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4409 #undef STORE_FUNCTION
4410
4411 #define CFQ_ATTR(name) \
4412         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4413
4414 static struct elv_fs_entry cfq_attrs[] = {
4415         CFQ_ATTR(quantum),
4416         CFQ_ATTR(fifo_expire_sync),
4417         CFQ_ATTR(fifo_expire_async),
4418         CFQ_ATTR(back_seek_max),
4419         CFQ_ATTR(back_seek_penalty),
4420         CFQ_ATTR(slice_sync),
4421         CFQ_ATTR(slice_async),
4422         CFQ_ATTR(slice_async_rq),
4423         CFQ_ATTR(slice_idle),
4424         CFQ_ATTR(group_idle),
4425         CFQ_ATTR(low_latency),
4426         CFQ_ATTR(target_latency),
4427         __ATTR_NULL
4428 };
4429
4430 static struct elevator_type iosched_cfq = {
4431         .ops = {
4432                 .elevator_merge_fn =            cfq_merge,
4433                 .elevator_merged_fn =           cfq_merged_request,
4434                 .elevator_merge_req_fn =        cfq_merged_requests,
4435                 .elevator_allow_merge_fn =      cfq_allow_merge,
4436                 .elevator_bio_merged_fn =       cfq_bio_merged,
4437                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4438                 .elevator_add_req_fn =          cfq_insert_request,
4439                 .elevator_activate_req_fn =     cfq_activate_request,
4440                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4441                 .elevator_completed_req_fn =    cfq_completed_request,
4442                 .elevator_former_req_fn =       elv_rb_former_request,
4443                 .elevator_latter_req_fn =       elv_rb_latter_request,
4444                 .elevator_init_icq_fn =         cfq_init_icq,
4445                 .elevator_exit_icq_fn =         cfq_exit_icq,
4446                 .elevator_set_req_fn =          cfq_set_request,
4447                 .elevator_put_req_fn =          cfq_put_request,
4448                 .elevator_may_queue_fn =        cfq_may_queue,
4449                 .elevator_init_fn =             cfq_init_queue,
4450                 .elevator_exit_fn =             cfq_exit_queue,
4451         },
4452         .icq_size       =       sizeof(struct cfq_io_cq),
4453         .icq_align      =       __alignof__(struct cfq_io_cq),
4454         .elevator_attrs =       cfq_attrs,
4455         .elevator_name  =       "cfq",
4456         .elevator_owner =       THIS_MODULE,
4457 };
4458
4459 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4460 static struct blkcg_policy blkcg_policy_cfq = {
4461         .pd_size                = sizeof(struct cfq_group),
4462         .cftypes                = cfq_blkcg_files,
4463
4464         .pd_init_fn             = cfq_pd_init,
4465         .pd_offline_fn          = cfq_pd_offline,
4466         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4467 };
4468 #endif
4469
4470 static int __init cfq_init(void)
4471 {
4472         int ret;
4473
4474         /*
4475          * could be 0 on HZ < 1000 setups
4476          */
4477         if (!cfq_slice_async)
4478                 cfq_slice_async = 1;
4479         if (!cfq_slice_idle)
4480                 cfq_slice_idle = 1;
4481
4482 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4483         if (!cfq_group_idle)
4484                 cfq_group_idle = 1;
4485
4486         ret = blkcg_policy_register(&blkcg_policy_cfq);
4487         if (ret)
4488                 return ret;
4489 #else
4490         cfq_group_idle = 0;
4491 #endif
4492
4493         ret = -ENOMEM;
4494         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4495         if (!cfq_pool)
4496                 goto err_pol_unreg;
4497
4498         ret = elv_register(&iosched_cfq);
4499         if (ret)
4500                 goto err_free_pool;
4501
4502         return 0;
4503
4504 err_free_pool:
4505         kmem_cache_destroy(cfq_pool);
4506 err_pol_unreg:
4507 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4508         blkcg_policy_unregister(&blkcg_policy_cfq);
4509 #endif
4510         return ret;
4511 }
4512
4513 static void __exit cfq_exit(void)
4514 {
4515 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4516         blkcg_policy_unregister(&blkcg_policy_cfq);
4517 #endif
4518         elv_unregister(&iosched_cfq);
4519         kmem_cache_destroy(cfq_pool);
4520 }
4521
4522 module_init(cfq_init);
4523 module_exit(cfq_exit);
4524
4525 MODULE_AUTHOR("Jens Axboe");
4526 MODULE_LICENSE("GPL");
4527 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");