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         if (cfqg->new_weight) {
1279                 cfqg->weight = cfqg->new_weight;
1280                 cfqg->new_weight = 0;
1281         }
1282 }
1283
1284 static void
1285 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1286 {
1287         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1288
1289         if (cfqg->new_leaf_weight) {
1290                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1291                 cfqg->new_leaf_weight = 0;
1292         }
1293 }
1294
1295 static void
1296 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1297 {
1298         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1299         struct cfq_group *pos = cfqg;
1300         struct cfq_group *parent;
1301         bool propagate;
1302
1303         /* add to the service tree */
1304         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1305
1306         cfq_update_group_leaf_weight(cfqg);
1307         __cfq_group_service_tree_add(st, cfqg);
1308
1309         /*
1310          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1311          * entitled to.  vfraction is calculated by walking the tree
1312          * towards the root calculating the fraction it has at each level.
1313          * The compounded ratio is how much vfraction @cfqg owns.
1314          *
1315          * Start with the proportion tasks in this cfqg has against active
1316          * children cfqgs - its leaf_weight against children_weight.
1317          */
1318         propagate = !pos->nr_active++;
1319         pos->children_weight += pos->leaf_weight;
1320         vfr = vfr * pos->leaf_weight / pos->children_weight;
1321
1322         /*
1323          * Compound ->weight walking up the tree.  Both activation and
1324          * vfraction calculation are done in the same loop.  Propagation
1325          * stops once an already activated node is met.  vfraction
1326          * calculation should always continue to the root.
1327          */
1328         while ((parent = cfqg_parent(pos))) {
1329                 if (propagate) {
1330                         cfq_update_group_weight(pos);
1331                         propagate = !parent->nr_active++;
1332                         parent->children_weight += pos->weight;
1333                 }
1334                 vfr = vfr * pos->weight / parent->children_weight;
1335                 pos = parent;
1336         }
1337
1338         cfqg->vfraction = max_t(unsigned, vfr, 1);
1339 }
1340
1341 static void
1342 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1343 {
1344         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1345         struct cfq_group *__cfqg;
1346         struct rb_node *n;
1347
1348         cfqg->nr_cfqq++;
1349         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1350                 return;
1351
1352         /*
1353          * Currently put the group at the end. Later implement something
1354          * so that groups get lesser vtime based on their weights, so that
1355          * if group does not loose all if it was not continuously backlogged.
1356          */
1357         n = rb_last(&st->rb);
1358         if (n) {
1359                 __cfqg = rb_entry_cfqg(n);
1360                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1361         } else
1362                 cfqg->vdisktime = st->min_vdisktime;
1363         cfq_group_service_tree_add(st, cfqg);
1364 }
1365
1366 static void
1367 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1368 {
1369         struct cfq_group *pos = cfqg;
1370         bool propagate;
1371
1372         /*
1373          * Undo activation from cfq_group_service_tree_add().  Deactivate
1374          * @cfqg and propagate deactivation upwards.
1375          */
1376         propagate = !--pos->nr_active;
1377         pos->children_weight -= pos->leaf_weight;
1378
1379         while (propagate) {
1380                 struct cfq_group *parent = cfqg_parent(pos);
1381
1382                 /* @pos has 0 nr_active at this point */
1383                 WARN_ON_ONCE(pos->children_weight);
1384                 pos->vfraction = 0;
1385
1386                 if (!parent)
1387                         break;
1388
1389                 propagate = !--parent->nr_active;
1390                 parent->children_weight -= pos->weight;
1391                 pos = parent;
1392         }
1393
1394         /* remove from the service tree */
1395         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1396                 cfq_rb_erase(&cfqg->rb_node, st);
1397 }
1398
1399 static void
1400 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1401 {
1402         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1403
1404         BUG_ON(cfqg->nr_cfqq < 1);
1405         cfqg->nr_cfqq--;
1406
1407         /* If there are other cfq queues under this group, don't delete it */
1408         if (cfqg->nr_cfqq)
1409                 return;
1410
1411         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1412         cfq_group_service_tree_del(st, cfqg);
1413         cfqg->saved_wl_slice = 0;
1414         cfqg_stats_update_dequeue(cfqg);
1415 }
1416
1417 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1418                                                 unsigned int *unaccounted_time)
1419 {
1420         unsigned int slice_used;
1421
1422         /*
1423          * Queue got expired before even a single request completed or
1424          * got expired immediately after first request completion.
1425          */
1426         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1427                 /*
1428                  * Also charge the seek time incurred to the group, otherwise
1429                  * if there are mutiple queues in the group, each can dispatch
1430                  * a single request on seeky media and cause lots of seek time
1431                  * and group will never know it.
1432                  */
1433                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1434                                         1);
1435         } else {
1436                 slice_used = jiffies - cfqq->slice_start;
1437                 if (slice_used > cfqq->allocated_slice) {
1438                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1439                         slice_used = cfqq->allocated_slice;
1440                 }
1441                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1442                         *unaccounted_time += cfqq->slice_start -
1443                                         cfqq->dispatch_start;
1444         }
1445
1446         return slice_used;
1447 }
1448
1449 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1450                                 struct cfq_queue *cfqq)
1451 {
1452         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1453         unsigned int used_sl, charge, unaccounted_sl = 0;
1454         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1455                         - cfqg->service_tree_idle.count;
1456         unsigned int vfr;
1457
1458         BUG_ON(nr_sync < 0);
1459         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1460
1461         if (iops_mode(cfqd))
1462                 charge = cfqq->slice_dispatch;
1463         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1464                 charge = cfqq->allocated_slice;
1465
1466         /*
1467          * Can't update vdisktime while on service tree and cfqg->vfraction
1468          * is valid only while on it.  Cache vfr, leave the service tree,
1469          * update vdisktime and go back on.  The re-addition to the tree
1470          * will also update the weights as necessary.
1471          */
1472         vfr = cfqg->vfraction;
1473         cfq_group_service_tree_del(st, cfqg);
1474         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1475         cfq_group_service_tree_add(st, cfqg);
1476
1477         /* This group is being expired. Save the context */
1478         if (time_after(cfqd->workload_expires, jiffies)) {
1479                 cfqg->saved_wl_slice = cfqd->workload_expires
1480                                                 - jiffies;
1481                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1482                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1483         } else
1484                 cfqg->saved_wl_slice = 0;
1485
1486         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1487                                         st->min_vdisktime);
1488         cfq_log_cfqq(cfqq->cfqd, cfqq,
1489                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1490                      used_sl, cfqq->slice_dispatch, charge,
1491                      iops_mode(cfqd), cfqq->nr_sectors);
1492         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1493         cfqg_stats_set_start_empty_time(cfqg);
1494 }
1495
1496 /**
1497  * cfq_init_cfqg_base - initialize base part of a cfq_group
1498  * @cfqg: cfq_group to initialize
1499  *
1500  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1501  * is enabled or not.
1502  */
1503 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1504 {
1505         struct cfq_rb_root *st;
1506         int i, j;
1507
1508         for_each_cfqg_st(cfqg, i, j, st)
1509                 *st = CFQ_RB_ROOT;
1510         RB_CLEAR_NODE(&cfqg->rb_node);
1511
1512         cfqg->ttime.last_end_request = jiffies;
1513 }
1514
1515 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1516 static void cfq_pd_init(struct blkcg_gq *blkg)
1517 {
1518         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1519
1520         cfq_init_cfqg_base(cfqg);
1521         cfqg->weight = blkg->blkcg->cfq_weight;
1522         cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
1523 }
1524
1525 static void cfq_pd_offline(struct blkcg_gq *blkg)
1526 {
1527         /*
1528          * @blkg is going offline and will be ignored by
1529          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1530          * that they don't get lost.  If IOs complete after this point, the
1531          * stats for them will be lost.  Oh well...
1532          */
1533         cfqg_stats_xfer_dead(blkg_to_cfqg(blkg));
1534 }
1535
1536 /* offset delta from cfqg->stats to cfqg->dead_stats */
1537 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1538                                         offsetof(struct cfq_group, stats);
1539
1540 /* to be used by recursive prfill, sums live and dead stats recursively */
1541 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1542 {
1543         u64 sum = 0;
1544
1545         sum += blkg_stat_recursive_sum(pd, off);
1546         sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1547         return sum;
1548 }
1549
1550 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1551 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1552                                                        int off)
1553 {
1554         struct blkg_rwstat a, b;
1555
1556         a = blkg_rwstat_recursive_sum(pd, off);
1557         b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1558         blkg_rwstat_merge(&a, &b);
1559         return a;
1560 }
1561
1562 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1563 {
1564         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1565
1566         cfqg_stats_reset(&cfqg->stats);
1567         cfqg_stats_reset(&cfqg->dead_stats);
1568 }
1569
1570 /*
1571  * Search for the cfq group current task belongs to. request_queue lock must
1572  * be held.
1573  */
1574 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1575                                                 struct blkcg *blkcg)
1576 {
1577         struct request_queue *q = cfqd->queue;
1578         struct cfq_group *cfqg = NULL;
1579
1580         /* avoid lookup for the common case where there's no blkcg */
1581         if (blkcg == &blkcg_root) {
1582                 cfqg = cfqd->root_group;
1583         } else {
1584                 struct blkcg_gq *blkg;
1585
1586                 blkg = blkg_lookup_create(blkcg, q);
1587                 if (!IS_ERR(blkg))
1588                         cfqg = blkg_to_cfqg(blkg);
1589         }
1590
1591         return cfqg;
1592 }
1593
1594 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1595 {
1596         /* Currently, all async queues are mapped to root group */
1597         if (!cfq_cfqq_sync(cfqq))
1598                 cfqg = cfqq->cfqd->root_group;
1599
1600         cfqq->cfqg = cfqg;
1601         /* cfqq reference on cfqg */
1602         cfqg_get(cfqg);
1603 }
1604
1605 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1606                                      struct blkg_policy_data *pd, int off)
1607 {
1608         struct cfq_group *cfqg = pd_to_cfqg(pd);
1609
1610         if (!cfqg->dev_weight)
1611                 return 0;
1612         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1613 }
1614
1615 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1616                                     struct seq_file *sf)
1617 {
1618         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1619                           cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1620                           false);
1621         return 0;
1622 }
1623
1624 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1625                                           struct blkg_policy_data *pd, int off)
1626 {
1627         struct cfq_group *cfqg = pd_to_cfqg(pd);
1628
1629         if (!cfqg->dev_leaf_weight)
1630                 return 0;
1631         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1632 }
1633
1634 static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1635                                          struct cftype *cft,
1636                                          struct seq_file *sf)
1637 {
1638         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1639                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1640                           false);
1641         return 0;
1642 }
1643
1644 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1645                             struct seq_file *sf)
1646 {
1647         seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1648         return 0;
1649 }
1650
1651 static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1652                                  struct seq_file *sf)
1653 {
1654         seq_printf(sf, "%u\n",
1655                    cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1656         return 0;
1657 }
1658
1659 static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1660                                     const char *buf, bool is_leaf_weight)
1661 {
1662         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1663         struct blkg_conf_ctx ctx;
1664         struct cfq_group *cfqg;
1665         int ret;
1666
1667         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1668         if (ret)
1669                 return ret;
1670
1671         ret = -EINVAL;
1672         cfqg = blkg_to_cfqg(ctx.blkg);
1673         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1674                 if (!is_leaf_weight) {
1675                         cfqg->dev_weight = ctx.v;
1676                         cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1677                 } else {
1678                         cfqg->dev_leaf_weight = ctx.v;
1679                         cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1680                 }
1681                 ret = 0;
1682         }
1683
1684         blkg_conf_finish(&ctx);
1685         return ret;
1686 }
1687
1688 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1689                                   const char *buf)
1690 {
1691         return __cfqg_set_weight_device(cgrp, cft, buf, false);
1692 }
1693
1694 static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1695                                        const char *buf)
1696 {
1697         return __cfqg_set_weight_device(cgrp, cft, buf, true);
1698 }
1699
1700 static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1701                             bool is_leaf_weight)
1702 {
1703         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1704         struct blkcg_gq *blkg;
1705
1706         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1707                 return -EINVAL;
1708
1709         spin_lock_irq(&blkcg->lock);
1710
1711         if (!is_leaf_weight)
1712                 blkcg->cfq_weight = val;
1713         else
1714                 blkcg->cfq_leaf_weight = val;
1715
1716         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1717                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1718
1719                 if (!cfqg)
1720                         continue;
1721
1722                 if (!is_leaf_weight) {
1723                         if (!cfqg->dev_weight)
1724                                 cfqg->new_weight = blkcg->cfq_weight;
1725                 } else {
1726                         if (!cfqg->dev_leaf_weight)
1727                                 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1728                 }
1729         }
1730
1731         spin_unlock_irq(&blkcg->lock);
1732         return 0;
1733 }
1734
1735 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1736 {
1737         return __cfq_set_weight(cgrp, cft, val, false);
1738 }
1739
1740 static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1741 {
1742         return __cfq_set_weight(cgrp, cft, val, true);
1743 }
1744
1745 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1746                            struct seq_file *sf)
1747 {
1748         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1749
1750         blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1751                           cft->private, false);
1752         return 0;
1753 }
1754
1755 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1756                              struct seq_file *sf)
1757 {
1758         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1759
1760         blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1761                           cft->private, true);
1762         return 0;
1763 }
1764
1765 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1766                                       struct blkg_policy_data *pd, int off)
1767 {
1768         u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1769
1770         return __blkg_prfill_u64(sf, pd, sum);
1771 }
1772
1773 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1774                                         struct blkg_policy_data *pd, int off)
1775 {
1776         struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1777
1778         return __blkg_prfill_rwstat(sf, pd, &sum);
1779 }
1780
1781 static int cfqg_print_stat_recursive(struct cgroup *cgrp, struct cftype *cft,
1782                                      struct seq_file *sf)
1783 {
1784         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1785
1786         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_stat_recursive,
1787                           &blkcg_policy_cfq, cft->private, false);
1788         return 0;
1789 }
1790
1791 static int cfqg_print_rwstat_recursive(struct cgroup *cgrp, struct cftype *cft,
1792                                        struct seq_file *sf)
1793 {
1794         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1795
1796         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_rwstat_recursive,
1797                           &blkcg_policy_cfq, cft->private, true);
1798         return 0;
1799 }
1800
1801 #ifdef CONFIG_DEBUG_BLK_CGROUP
1802 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1803                                       struct blkg_policy_data *pd, int off)
1804 {
1805         struct cfq_group *cfqg = pd_to_cfqg(pd);
1806         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1807         u64 v = 0;
1808
1809         if (samples) {
1810                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1811                 v = div64_u64(v, samples);
1812         }
1813         __blkg_prfill_u64(sf, pd, v);
1814         return 0;
1815 }
1816
1817 /* print avg_queue_size */
1818 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1819                                      struct seq_file *sf)
1820 {
1821         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1822
1823         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1824                           &blkcg_policy_cfq, 0, false);
1825         return 0;
1826 }
1827 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1828
1829 static struct cftype cfq_blkcg_files[] = {
1830         /* on root, weight is mapped to leaf_weight */
1831         {
1832                 .name = "weight_device",
1833                 .flags = CFTYPE_ONLY_ON_ROOT,
1834                 .read_seq_string = cfqg_print_leaf_weight_device,
1835                 .write_string = cfqg_set_leaf_weight_device,
1836                 .max_write_len = 256,
1837         },
1838         {
1839                 .name = "weight",
1840                 .flags = CFTYPE_ONLY_ON_ROOT,
1841                 .read_seq_string = cfq_print_leaf_weight,
1842                 .write_u64 = cfq_set_leaf_weight,
1843         },
1844
1845         /* no such mapping necessary for !roots */
1846         {
1847                 .name = "weight_device",
1848                 .flags = CFTYPE_NOT_ON_ROOT,
1849                 .read_seq_string = cfqg_print_weight_device,
1850                 .write_string = cfqg_set_weight_device,
1851                 .max_write_len = 256,
1852         },
1853         {
1854                 .name = "weight",
1855                 .flags = CFTYPE_NOT_ON_ROOT,
1856                 .read_seq_string = cfq_print_weight,
1857                 .write_u64 = cfq_set_weight,
1858         },
1859
1860         {
1861                 .name = "leaf_weight_device",
1862                 .read_seq_string = cfqg_print_leaf_weight_device,
1863                 .write_string = cfqg_set_leaf_weight_device,
1864                 .max_write_len = 256,
1865         },
1866         {
1867                 .name = "leaf_weight",
1868                 .read_seq_string = cfq_print_leaf_weight,
1869                 .write_u64 = cfq_set_leaf_weight,
1870         },
1871
1872         /* statistics, covers only the tasks in the cfqg */
1873         {
1874                 .name = "time",
1875                 .private = offsetof(struct cfq_group, stats.time),
1876                 .read_seq_string = cfqg_print_stat,
1877         },
1878         {
1879                 .name = "sectors",
1880                 .private = offsetof(struct cfq_group, stats.sectors),
1881                 .read_seq_string = cfqg_print_stat,
1882         },
1883         {
1884                 .name = "io_service_bytes",
1885                 .private = offsetof(struct cfq_group, stats.service_bytes),
1886                 .read_seq_string = cfqg_print_rwstat,
1887         },
1888         {
1889                 .name = "io_serviced",
1890                 .private = offsetof(struct cfq_group, stats.serviced),
1891                 .read_seq_string = cfqg_print_rwstat,
1892         },
1893         {
1894                 .name = "io_service_time",
1895                 .private = offsetof(struct cfq_group, stats.service_time),
1896                 .read_seq_string = cfqg_print_rwstat,
1897         },
1898         {
1899                 .name = "io_wait_time",
1900                 .private = offsetof(struct cfq_group, stats.wait_time),
1901                 .read_seq_string = cfqg_print_rwstat,
1902         },
1903         {
1904                 .name = "io_merged",
1905                 .private = offsetof(struct cfq_group, stats.merged),
1906                 .read_seq_string = cfqg_print_rwstat,
1907         },
1908         {
1909                 .name = "io_queued",
1910                 .private = offsetof(struct cfq_group, stats.queued),
1911                 .read_seq_string = cfqg_print_rwstat,
1912         },
1913
1914         /* the same statictics which cover the cfqg and its descendants */
1915         {
1916                 .name = "time_recursive",
1917                 .private = offsetof(struct cfq_group, stats.time),
1918                 .read_seq_string = cfqg_print_stat_recursive,
1919         },
1920         {
1921                 .name = "sectors_recursive",
1922                 .private = offsetof(struct cfq_group, stats.sectors),
1923                 .read_seq_string = cfqg_print_stat_recursive,
1924         },
1925         {
1926                 .name = "io_service_bytes_recursive",
1927                 .private = offsetof(struct cfq_group, stats.service_bytes),
1928                 .read_seq_string = cfqg_print_rwstat_recursive,
1929         },
1930         {
1931                 .name = "io_serviced_recursive",
1932                 .private = offsetof(struct cfq_group, stats.serviced),
1933                 .read_seq_string = cfqg_print_rwstat_recursive,
1934         },
1935         {
1936                 .name = "io_service_time_recursive",
1937                 .private = offsetof(struct cfq_group, stats.service_time),
1938                 .read_seq_string = cfqg_print_rwstat_recursive,
1939         },
1940         {
1941                 .name = "io_wait_time_recursive",
1942                 .private = offsetof(struct cfq_group, stats.wait_time),
1943                 .read_seq_string = cfqg_print_rwstat_recursive,
1944         },
1945         {
1946                 .name = "io_merged_recursive",
1947                 .private = offsetof(struct cfq_group, stats.merged),
1948                 .read_seq_string = cfqg_print_rwstat_recursive,
1949         },
1950         {
1951                 .name = "io_queued_recursive",
1952                 .private = offsetof(struct cfq_group, stats.queued),
1953                 .read_seq_string = cfqg_print_rwstat_recursive,
1954         },
1955 #ifdef CONFIG_DEBUG_BLK_CGROUP
1956         {
1957                 .name = "avg_queue_size",
1958                 .read_seq_string = cfqg_print_avg_queue_size,
1959         },
1960         {
1961                 .name = "group_wait_time",
1962                 .private = offsetof(struct cfq_group, stats.group_wait_time),
1963                 .read_seq_string = cfqg_print_stat,
1964         },
1965         {
1966                 .name = "idle_time",
1967                 .private = offsetof(struct cfq_group, stats.idle_time),
1968                 .read_seq_string = cfqg_print_stat,
1969         },
1970         {
1971                 .name = "empty_time",
1972                 .private = offsetof(struct cfq_group, stats.empty_time),
1973                 .read_seq_string = cfqg_print_stat,
1974         },
1975         {
1976                 .name = "dequeue",
1977                 .private = offsetof(struct cfq_group, stats.dequeue),
1978                 .read_seq_string = cfqg_print_stat,
1979         },
1980         {
1981                 .name = "unaccounted_time",
1982                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1983                 .read_seq_string = cfqg_print_stat,
1984         },
1985 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1986         { }     /* terminate */
1987 };
1988 #else /* GROUP_IOSCHED */
1989 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1990                                                 struct blkcg *blkcg)
1991 {
1992         return cfqd->root_group;
1993 }
1994
1995 static inline void
1996 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1997         cfqq->cfqg = cfqg;
1998 }
1999
2000 #endif /* GROUP_IOSCHED */
2001
2002 /*
2003  * The cfqd->service_trees holds all pending cfq_queue's that have
2004  * requests waiting to be processed. It is sorted in the order that
2005  * we will service the queues.
2006  */
2007 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2008                                  bool add_front)
2009 {
2010         struct rb_node **p, *parent;
2011         struct cfq_queue *__cfqq;
2012         unsigned long rb_key;
2013         struct cfq_rb_root *st;
2014         int left;
2015         int new_cfqq = 1;
2016
2017         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2018         if (cfq_class_idle(cfqq)) {
2019                 rb_key = CFQ_IDLE_DELAY;
2020                 parent = rb_last(&st->rb);
2021                 if (parent && parent != &cfqq->rb_node) {
2022                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2023                         rb_key += __cfqq->rb_key;
2024                 } else
2025                         rb_key += jiffies;
2026         } else if (!add_front) {
2027                 /*
2028                  * Get our rb key offset. Subtract any residual slice
2029                  * value carried from last service. A negative resid
2030                  * count indicates slice overrun, and this should position
2031                  * the next service time further away in the tree.
2032                  */
2033                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2034                 rb_key -= cfqq->slice_resid;
2035                 cfqq->slice_resid = 0;
2036         } else {
2037                 rb_key = -HZ;
2038                 __cfqq = cfq_rb_first(st);
2039                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2040         }
2041
2042         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2043                 new_cfqq = 0;
2044                 /*
2045                  * same position, nothing more to do
2046                  */
2047                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2048                         return;
2049
2050                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2051                 cfqq->service_tree = NULL;
2052         }
2053
2054         left = 1;
2055         parent = NULL;
2056         cfqq->service_tree = st;
2057         p = &st->rb.rb_node;
2058         while (*p) {
2059                 parent = *p;
2060                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2061
2062                 /*
2063                  * sort by key, that represents service time.
2064                  */
2065                 if (time_before(rb_key, __cfqq->rb_key))
2066                         p = &parent->rb_left;
2067                 else {
2068                         p = &parent->rb_right;
2069                         left = 0;
2070                 }
2071         }
2072
2073         if (left)
2074                 st->left = &cfqq->rb_node;
2075
2076         cfqq->rb_key = rb_key;
2077         rb_link_node(&cfqq->rb_node, parent, p);
2078         rb_insert_color(&cfqq->rb_node, &st->rb);
2079         st->count++;
2080         if (add_front || !new_cfqq)
2081                 return;
2082         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2083 }
2084
2085 static struct cfq_queue *
2086 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2087                      sector_t sector, struct rb_node **ret_parent,
2088                      struct rb_node ***rb_link)
2089 {
2090         struct rb_node **p, *parent;
2091         struct cfq_queue *cfqq = NULL;
2092
2093         parent = NULL;
2094         p = &root->rb_node;
2095         while (*p) {
2096                 struct rb_node **n;
2097
2098                 parent = *p;
2099                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2100
2101                 /*
2102                  * Sort strictly based on sector.  Smallest to the left,
2103                  * largest to the right.
2104                  */
2105                 if (sector > blk_rq_pos(cfqq->next_rq))
2106                         n = &(*p)->rb_right;
2107                 else if (sector < blk_rq_pos(cfqq->next_rq))
2108                         n = &(*p)->rb_left;
2109                 else
2110                         break;
2111                 p = n;
2112                 cfqq = NULL;
2113         }
2114
2115         *ret_parent = parent;
2116         if (rb_link)
2117                 *rb_link = p;
2118         return cfqq;
2119 }
2120
2121 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2122 {
2123         struct rb_node **p, *parent;
2124         struct cfq_queue *__cfqq;
2125
2126         if (cfqq->p_root) {
2127                 rb_erase(&cfqq->p_node, cfqq->p_root);
2128                 cfqq->p_root = NULL;
2129         }
2130
2131         if (cfq_class_idle(cfqq))
2132                 return;
2133         if (!cfqq->next_rq)
2134                 return;
2135
2136         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2137         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2138                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2139         if (!__cfqq) {
2140                 rb_link_node(&cfqq->p_node, parent, p);
2141                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2142         } else
2143                 cfqq->p_root = NULL;
2144 }
2145
2146 /*
2147  * Update cfqq's position in the service tree.
2148  */
2149 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2150 {
2151         /*
2152          * Resorting requires the cfqq to be on the RR list already.
2153          */
2154         if (cfq_cfqq_on_rr(cfqq)) {
2155                 cfq_service_tree_add(cfqd, cfqq, 0);
2156                 cfq_prio_tree_add(cfqd, cfqq);
2157         }
2158 }
2159
2160 /*
2161  * add to busy list of queues for service, trying to be fair in ordering
2162  * the pending list according to last request service
2163  */
2164 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2165 {
2166         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2167         BUG_ON(cfq_cfqq_on_rr(cfqq));
2168         cfq_mark_cfqq_on_rr(cfqq);
2169         cfqd->busy_queues++;
2170         if (cfq_cfqq_sync(cfqq))
2171                 cfqd->busy_sync_queues++;
2172
2173         cfq_resort_rr_list(cfqd, cfqq);
2174 }
2175
2176 /*
2177  * Called when the cfqq no longer has requests pending, remove it from
2178  * the service tree.
2179  */
2180 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2181 {
2182         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2183         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2184         cfq_clear_cfqq_on_rr(cfqq);
2185
2186         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2187                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2188                 cfqq->service_tree = NULL;
2189         }
2190         if (cfqq->p_root) {
2191                 rb_erase(&cfqq->p_node, cfqq->p_root);
2192                 cfqq->p_root = NULL;
2193         }
2194
2195         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2196         BUG_ON(!cfqd->busy_queues);
2197         cfqd->busy_queues--;
2198         if (cfq_cfqq_sync(cfqq))
2199                 cfqd->busy_sync_queues--;
2200 }
2201
2202 /*
2203  * rb tree support functions
2204  */
2205 static void cfq_del_rq_rb(struct request *rq)
2206 {
2207         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2208         const int sync = rq_is_sync(rq);
2209
2210         BUG_ON(!cfqq->queued[sync]);
2211         cfqq->queued[sync]--;
2212
2213         elv_rb_del(&cfqq->sort_list, rq);
2214
2215         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2216                 /*
2217                  * Queue will be deleted from service tree when we actually
2218                  * expire it later. Right now just remove it from prio tree
2219                  * as it is empty.
2220                  */
2221                 if (cfqq->p_root) {
2222                         rb_erase(&cfqq->p_node, cfqq->p_root);
2223                         cfqq->p_root = NULL;
2224                 }
2225         }
2226 }
2227
2228 static void cfq_add_rq_rb(struct request *rq)
2229 {
2230         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2231         struct cfq_data *cfqd = cfqq->cfqd;
2232         struct request *prev;
2233
2234         cfqq->queued[rq_is_sync(rq)]++;
2235
2236         elv_rb_add(&cfqq->sort_list, rq);
2237
2238         if (!cfq_cfqq_on_rr(cfqq))
2239                 cfq_add_cfqq_rr(cfqd, cfqq);
2240
2241         /*
2242          * check if this request is a better next-serve candidate
2243          */
2244         prev = cfqq->next_rq;
2245         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2246
2247         /*
2248          * adjust priority tree position, if ->next_rq changes
2249          */
2250         if (prev != cfqq->next_rq)
2251                 cfq_prio_tree_add(cfqd, cfqq);
2252
2253         BUG_ON(!cfqq->next_rq);
2254 }
2255
2256 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2257 {
2258         elv_rb_del(&cfqq->sort_list, rq);
2259         cfqq->queued[rq_is_sync(rq)]--;
2260         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2261         cfq_add_rq_rb(rq);
2262         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2263                                  rq->cmd_flags);
2264 }
2265
2266 static struct request *
2267 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2268 {
2269         struct task_struct *tsk = current;
2270         struct cfq_io_cq *cic;
2271         struct cfq_queue *cfqq;
2272
2273         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2274         if (!cic)
2275                 return NULL;
2276
2277         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2278         if (cfqq)
2279                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2280
2281         return NULL;
2282 }
2283
2284 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2285 {
2286         struct cfq_data *cfqd = q->elevator->elevator_data;
2287
2288         cfqd->rq_in_driver++;
2289         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2290                                                 cfqd->rq_in_driver);
2291
2292         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2293 }
2294
2295 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2296 {
2297         struct cfq_data *cfqd = q->elevator->elevator_data;
2298
2299         WARN_ON(!cfqd->rq_in_driver);
2300         cfqd->rq_in_driver--;
2301         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2302                                                 cfqd->rq_in_driver);
2303 }
2304
2305 static void cfq_remove_request(struct request *rq)
2306 {
2307         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2308
2309         if (cfqq->next_rq == rq)
2310                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2311
2312         list_del_init(&rq->queuelist);
2313         cfq_del_rq_rb(rq);
2314
2315         cfqq->cfqd->rq_queued--;
2316         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2317         if (rq->cmd_flags & REQ_PRIO) {
2318                 WARN_ON(!cfqq->prio_pending);
2319                 cfqq->prio_pending--;
2320         }
2321 }
2322
2323 static int cfq_merge(struct request_queue *q, struct request **req,
2324                      struct bio *bio)
2325 {
2326         struct cfq_data *cfqd = q->elevator->elevator_data;
2327         struct request *__rq;
2328
2329         __rq = cfq_find_rq_fmerge(cfqd, bio);
2330         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2331                 *req = __rq;
2332                 return ELEVATOR_FRONT_MERGE;
2333         }
2334
2335         return ELEVATOR_NO_MERGE;
2336 }
2337
2338 static void cfq_merged_request(struct request_queue *q, struct request *req,
2339                                int type)
2340 {
2341         if (type == ELEVATOR_FRONT_MERGE) {
2342                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2343
2344                 cfq_reposition_rq_rb(cfqq, req);
2345         }
2346 }
2347
2348 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2349                                 struct bio *bio)
2350 {
2351         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2352 }
2353
2354 static void
2355 cfq_merged_requests(struct request_queue *q, struct request *rq,
2356                     struct request *next)
2357 {
2358         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2359         struct cfq_data *cfqd = q->elevator->elevator_data;
2360
2361         /*
2362          * reposition in fifo if next is older than rq
2363          */
2364         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2365             time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2366             cfqq == RQ_CFQQ(next)) {
2367                 list_move(&rq->queuelist, &next->queuelist);
2368                 rq_set_fifo_time(rq, rq_fifo_time(next));
2369         }
2370
2371         if (cfqq->next_rq == next)
2372                 cfqq->next_rq = rq;
2373         cfq_remove_request(next);
2374         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2375
2376         cfqq = RQ_CFQQ(next);
2377         /*
2378          * all requests of this queue are merged to other queues, delete it
2379          * from the service tree. If it's the active_queue,
2380          * cfq_dispatch_requests() will choose to expire it or do idle
2381          */
2382         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2383             cfqq != cfqd->active_queue)
2384                 cfq_del_cfqq_rr(cfqd, cfqq);
2385 }
2386
2387 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2388                            struct bio *bio)
2389 {
2390         struct cfq_data *cfqd = q->elevator->elevator_data;
2391         struct cfq_io_cq *cic;
2392         struct cfq_queue *cfqq;
2393
2394         /*
2395          * Disallow merge of a sync bio into an async request.
2396          */
2397         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2398                 return false;
2399
2400         /*
2401          * Lookup the cfqq that this bio will be queued with and allow
2402          * merge only if rq is queued there.
2403          */
2404         cic = cfq_cic_lookup(cfqd, current->io_context);
2405         if (!cic)
2406                 return false;
2407
2408         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2409         return cfqq == RQ_CFQQ(rq);
2410 }
2411
2412 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2413 {
2414         del_timer(&cfqd->idle_slice_timer);
2415         cfqg_stats_update_idle_time(cfqq->cfqg);
2416 }
2417
2418 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2419                                    struct cfq_queue *cfqq)
2420 {
2421         if (cfqq) {
2422                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2423                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2424                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2425                 cfqq->slice_start = 0;
2426                 cfqq->dispatch_start = jiffies;
2427                 cfqq->allocated_slice = 0;
2428                 cfqq->slice_end = 0;
2429                 cfqq->slice_dispatch = 0;
2430                 cfqq->nr_sectors = 0;
2431
2432                 cfq_clear_cfqq_wait_request(cfqq);
2433                 cfq_clear_cfqq_must_dispatch(cfqq);
2434                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2435                 cfq_clear_cfqq_fifo_expire(cfqq);
2436                 cfq_mark_cfqq_slice_new(cfqq);
2437
2438                 cfq_del_timer(cfqd, cfqq);
2439         }
2440
2441         cfqd->active_queue = cfqq;
2442 }
2443
2444 /*
2445  * current cfqq expired its slice (or was too idle), select new one
2446  */
2447 static void
2448 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2449                     bool timed_out)
2450 {
2451         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2452
2453         if (cfq_cfqq_wait_request(cfqq))
2454                 cfq_del_timer(cfqd, cfqq);
2455
2456         cfq_clear_cfqq_wait_request(cfqq);
2457         cfq_clear_cfqq_wait_busy(cfqq);
2458
2459         /*
2460          * If this cfqq is shared between multiple processes, check to
2461          * make sure that those processes are still issuing I/Os within
2462          * the mean seek distance.  If not, it may be time to break the
2463          * queues apart again.
2464          */
2465         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2466                 cfq_mark_cfqq_split_coop(cfqq);
2467
2468         /*
2469          * store what was left of this slice, if the queue idled/timed out
2470          */
2471         if (timed_out) {
2472                 if (cfq_cfqq_slice_new(cfqq))
2473                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2474                 else
2475                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2476                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2477         }
2478
2479         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2480
2481         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2482                 cfq_del_cfqq_rr(cfqd, cfqq);
2483
2484         cfq_resort_rr_list(cfqd, cfqq);
2485
2486         if (cfqq == cfqd->active_queue)
2487                 cfqd->active_queue = NULL;
2488
2489         if (cfqd->active_cic) {
2490                 put_io_context(cfqd->active_cic->icq.ioc);
2491                 cfqd->active_cic = NULL;
2492         }
2493 }
2494
2495 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2496 {
2497         struct cfq_queue *cfqq = cfqd->active_queue;
2498
2499         if (cfqq)
2500                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2501 }
2502
2503 /*
2504  * Get next queue for service. Unless we have a queue preemption,
2505  * we'll simply select the first cfqq in the service tree.
2506  */
2507 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2508 {
2509         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2510                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2511
2512         if (!cfqd->rq_queued)
2513                 return NULL;
2514
2515         /* There is nothing to dispatch */
2516         if (!st)
2517                 return NULL;
2518         if (RB_EMPTY_ROOT(&st->rb))
2519                 return NULL;
2520         return cfq_rb_first(st);
2521 }
2522
2523 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2524 {
2525         struct cfq_group *cfqg;
2526         struct cfq_queue *cfqq;
2527         int i, j;
2528         struct cfq_rb_root *st;
2529
2530         if (!cfqd->rq_queued)
2531                 return NULL;
2532
2533         cfqg = cfq_get_next_cfqg(cfqd);
2534         if (!cfqg)
2535                 return NULL;
2536
2537         for_each_cfqg_st(cfqg, i, j, st)
2538                 if ((cfqq = cfq_rb_first(st)) != NULL)
2539                         return cfqq;
2540         return NULL;
2541 }
2542
2543 /*
2544  * Get and set a new active queue for service.
2545  */
2546 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2547                                               struct cfq_queue *cfqq)
2548 {
2549         if (!cfqq)
2550                 cfqq = cfq_get_next_queue(cfqd);
2551
2552         __cfq_set_active_queue(cfqd, cfqq);
2553         return cfqq;
2554 }
2555
2556 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2557                                           struct request *rq)
2558 {
2559         if (blk_rq_pos(rq) >= cfqd->last_position)
2560                 return blk_rq_pos(rq) - cfqd->last_position;
2561         else
2562                 return cfqd->last_position - blk_rq_pos(rq);
2563 }
2564
2565 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2566                                struct request *rq)
2567 {
2568         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2569 }
2570
2571 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2572                                     struct cfq_queue *cur_cfqq)
2573 {
2574         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2575         struct rb_node *parent, *node;
2576         struct cfq_queue *__cfqq;
2577         sector_t sector = cfqd->last_position;
2578
2579         if (RB_EMPTY_ROOT(root))
2580                 return NULL;
2581
2582         /*
2583          * First, if we find a request starting at the end of the last
2584          * request, choose it.
2585          */
2586         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2587         if (__cfqq)
2588                 return __cfqq;
2589
2590         /*
2591          * If the exact sector wasn't found, the parent of the NULL leaf
2592          * will contain the closest sector.
2593          */
2594         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2595         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2596                 return __cfqq;
2597
2598         if (blk_rq_pos(__cfqq->next_rq) < sector)
2599                 node = rb_next(&__cfqq->p_node);
2600         else
2601                 node = rb_prev(&__cfqq->p_node);
2602         if (!node)
2603                 return NULL;
2604
2605         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2606         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2607                 return __cfqq;
2608
2609         return NULL;
2610 }
2611
2612 /*
2613  * cfqd - obvious
2614  * cur_cfqq - passed in so that we don't decide that the current queue is
2615  *            closely cooperating with itself.
2616  *
2617  * So, basically we're assuming that that cur_cfqq has dispatched at least
2618  * one request, and that cfqd->last_position reflects a position on the disk
2619  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2620  * assumption.
2621  */
2622 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2623                                               struct cfq_queue *cur_cfqq)
2624 {
2625         struct cfq_queue *cfqq;
2626
2627         if (cfq_class_idle(cur_cfqq))
2628                 return NULL;
2629         if (!cfq_cfqq_sync(cur_cfqq))
2630                 return NULL;
2631         if (CFQQ_SEEKY(cur_cfqq))
2632                 return NULL;
2633
2634         /*
2635          * Don't search priority tree if it's the only queue in the group.
2636          */
2637         if (cur_cfqq->cfqg->nr_cfqq == 1)
2638                 return NULL;
2639
2640         /*
2641          * We should notice if some of the queues are cooperating, eg
2642          * working closely on the same area of the disk. In that case,
2643          * we can group them together and don't waste time idling.
2644          */
2645         cfqq = cfqq_close(cfqd, cur_cfqq);
2646         if (!cfqq)
2647                 return NULL;
2648
2649         /* If new queue belongs to different cfq_group, don't choose it */
2650         if (cur_cfqq->cfqg != cfqq->cfqg)
2651                 return NULL;
2652
2653         /*
2654          * It only makes sense to merge sync queues.
2655          */
2656         if (!cfq_cfqq_sync(cfqq))
2657                 return NULL;
2658         if (CFQQ_SEEKY(cfqq))
2659                 return NULL;
2660
2661         /*
2662          * Do not merge queues of different priority classes
2663          */
2664         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2665                 return NULL;
2666
2667         return cfqq;
2668 }
2669
2670 /*
2671  * Determine whether we should enforce idle window for this queue.
2672  */
2673
2674 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2675 {
2676         enum wl_class_t wl_class = cfqq_class(cfqq);
2677         struct cfq_rb_root *st = cfqq->service_tree;
2678
2679         BUG_ON(!st);
2680         BUG_ON(!st->count);
2681
2682         if (!cfqd->cfq_slice_idle)
2683                 return false;
2684
2685         /* We never do for idle class queues. */
2686         if (wl_class == IDLE_WORKLOAD)
2687                 return false;
2688
2689         /* We do for queues that were marked with idle window flag. */
2690         if (cfq_cfqq_idle_window(cfqq) &&
2691            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2692                 return true;
2693
2694         /*
2695          * Otherwise, we do only if they are the last ones
2696          * in their service tree.
2697          */
2698         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2699            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2700                 return true;
2701         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2702         return false;
2703 }
2704
2705 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2706 {
2707         struct cfq_queue *cfqq = cfqd->active_queue;
2708         struct cfq_io_cq *cic;
2709         unsigned long sl, group_idle = 0;
2710
2711         /*
2712          * SSD device without seek penalty, disable idling. But only do so
2713          * for devices that support queuing, otherwise we still have a problem
2714          * with sync vs async workloads.
2715          */
2716         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2717                 return;
2718
2719         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2720         WARN_ON(cfq_cfqq_slice_new(cfqq));
2721
2722         /*
2723          * idle is disabled, either manually or by past process history
2724          */
2725         if (!cfq_should_idle(cfqd, cfqq)) {
2726                 /* no queue idling. Check for group idling */
2727                 if (cfqd->cfq_group_idle)
2728                         group_idle = cfqd->cfq_group_idle;
2729                 else
2730                         return;
2731         }
2732
2733         /*
2734          * still active requests from this queue, don't idle
2735          */
2736         if (cfqq->dispatched)
2737                 return;
2738
2739         /*
2740          * task has exited, don't wait
2741          */
2742         cic = cfqd->active_cic;
2743         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2744                 return;
2745
2746         /*
2747          * If our average think time is larger than the remaining time
2748          * slice, then don't idle. This avoids overrunning the allotted
2749          * time slice.
2750          */
2751         if (sample_valid(cic->ttime.ttime_samples) &&
2752             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2753                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2754                              cic->ttime.ttime_mean);
2755                 return;
2756         }
2757
2758         /* There are other queues in the group, don't do group idle */
2759         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2760                 return;
2761
2762         cfq_mark_cfqq_wait_request(cfqq);
2763
2764         if (group_idle)
2765                 sl = cfqd->cfq_group_idle;
2766         else
2767                 sl = cfqd->cfq_slice_idle;
2768
2769         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2770         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2771         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2772                         group_idle ? 1 : 0);
2773 }
2774
2775 /*
2776  * Move request from internal lists to the request queue dispatch list.
2777  */
2778 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2779 {
2780         struct cfq_data *cfqd = q->elevator->elevator_data;
2781         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2782
2783         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2784
2785         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2786         cfq_remove_request(rq);
2787         cfqq->dispatched++;
2788         (RQ_CFQG(rq))->dispatched++;
2789         elv_dispatch_sort(q, rq);
2790
2791         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2792         cfqq->nr_sectors += blk_rq_sectors(rq);
2793         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2794 }
2795
2796 /*
2797  * return expired entry, or NULL to just start from scratch in rbtree
2798  */
2799 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2800 {
2801         struct request *rq = NULL;
2802
2803         if (cfq_cfqq_fifo_expire(cfqq))
2804                 return NULL;
2805
2806         cfq_mark_cfqq_fifo_expire(cfqq);
2807
2808         if (list_empty(&cfqq->fifo))
2809                 return NULL;
2810
2811         rq = rq_entry_fifo(cfqq->fifo.next);
2812         if (time_before(jiffies, rq_fifo_time(rq)))
2813                 rq = NULL;
2814
2815         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2816         return rq;
2817 }
2818
2819 static inline int
2820 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2821 {
2822         const int base_rq = cfqd->cfq_slice_async_rq;
2823
2824         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2825
2826         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2827 }
2828
2829 /*
2830  * Must be called with the queue_lock held.
2831  */
2832 static int cfqq_process_refs(struct cfq_queue *cfqq)
2833 {
2834         int process_refs, io_refs;
2835
2836         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2837         process_refs = cfqq->ref - io_refs;
2838         BUG_ON(process_refs < 0);
2839         return process_refs;
2840 }
2841
2842 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2843 {
2844         int process_refs, new_process_refs;
2845         struct cfq_queue *__cfqq;
2846
2847         /*
2848          * If there are no process references on the new_cfqq, then it is
2849          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2850          * chain may have dropped their last reference (not just their
2851          * last process reference).
2852          */
2853         if (!cfqq_process_refs(new_cfqq))
2854                 return;
2855
2856         /* Avoid a circular list and skip interim queue merges */
2857         while ((__cfqq = new_cfqq->new_cfqq)) {
2858                 if (__cfqq == cfqq)
2859                         return;
2860                 new_cfqq = __cfqq;
2861         }
2862
2863         process_refs = cfqq_process_refs(cfqq);
2864         new_process_refs = cfqq_process_refs(new_cfqq);
2865         /*
2866          * If the process for the cfqq has gone away, there is no
2867          * sense in merging the queues.
2868          */
2869         if (process_refs == 0 || new_process_refs == 0)
2870                 return;
2871
2872         /*
2873          * Merge in the direction of the lesser amount of work.
2874          */
2875         if (new_process_refs >= process_refs) {
2876                 cfqq->new_cfqq = new_cfqq;
2877                 new_cfqq->ref += process_refs;
2878         } else {
2879                 new_cfqq->new_cfqq = cfqq;
2880                 cfqq->ref += new_process_refs;
2881         }
2882 }
2883
2884 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2885                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2886 {
2887         struct cfq_queue *queue;
2888         int i;
2889         bool key_valid = false;
2890         unsigned long lowest_key = 0;
2891         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2892
2893         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2894                 /* select the one with lowest rb_key */
2895                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2896                 if (queue &&
2897                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2898                         lowest_key = queue->rb_key;
2899                         cur_best = i;
2900                         key_valid = true;
2901                 }
2902         }
2903
2904         return cur_best;
2905 }
2906
2907 static void
2908 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2909 {
2910         unsigned slice;
2911         unsigned count;
2912         struct cfq_rb_root *st;
2913         unsigned group_slice;
2914         enum wl_class_t original_class = cfqd->serving_wl_class;
2915
2916         /* Choose next priority. RT > BE > IDLE */
2917         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2918                 cfqd->serving_wl_class = RT_WORKLOAD;
2919         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2920                 cfqd->serving_wl_class = BE_WORKLOAD;
2921         else {
2922                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2923                 cfqd->workload_expires = jiffies + 1;
2924                 return;
2925         }
2926
2927         if (original_class != cfqd->serving_wl_class)
2928                 goto new_workload;
2929
2930         /*
2931          * For RT and BE, we have to choose also the type
2932          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2933          * expiration time
2934          */
2935         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2936         count = st->count;
2937
2938         /*
2939          * check workload expiration, and that we still have other queues ready
2940          */
2941         if (count && !time_after(jiffies, cfqd->workload_expires))
2942                 return;
2943
2944 new_workload:
2945         /* otherwise select new workload type */
2946         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2947                                         cfqd->serving_wl_class);
2948         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2949         count = st->count;
2950
2951         /*
2952          * the workload slice is computed as a fraction of target latency
2953          * proportional to the number of queues in that workload, over
2954          * all the queues in the same priority class
2955          */
2956         group_slice = cfq_group_slice(cfqd, cfqg);
2957
2958         slice = group_slice * count /
2959                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2960                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2961                                         cfqg));
2962
2963         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2964                 unsigned int tmp;
2965
2966                 /*
2967                  * Async queues are currently system wide. Just taking
2968                  * proportion of queues with-in same group will lead to higher
2969                  * async ratio system wide as generally root group is going
2970                  * to have higher weight. A more accurate thing would be to
2971                  * calculate system wide asnc/sync ratio.
2972                  */
2973                 tmp = cfqd->cfq_target_latency *
2974                         cfqg_busy_async_queues(cfqd, cfqg);
2975                 tmp = tmp/cfqd->busy_queues;
2976                 slice = min_t(unsigned, slice, tmp);
2977
2978                 /* async workload slice is scaled down according to
2979                  * the sync/async slice ratio. */
2980                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2981         } else
2982                 /* sync workload slice is at least 2 * cfq_slice_idle */
2983                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2984
2985         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2986         cfq_log(cfqd, "workload slice:%d", slice);
2987         cfqd->workload_expires = jiffies + slice;
2988 }
2989
2990 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2991 {
2992         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2993         struct cfq_group *cfqg;
2994
2995         if (RB_EMPTY_ROOT(&st->rb))
2996                 return NULL;
2997         cfqg = cfq_rb_first_group(st);
2998         update_min_vdisktime(st);
2999         return cfqg;
3000 }
3001
3002 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3003 {
3004         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3005
3006         cfqd->serving_group = cfqg;
3007
3008         /* Restore the workload type data */
3009         if (cfqg->saved_wl_slice) {
3010                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3011                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3012                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3013         } else
3014                 cfqd->workload_expires = jiffies - 1;
3015
3016         choose_wl_class_and_type(cfqd, cfqg);
3017 }
3018
3019 /*
3020  * Select a queue for service. If we have a current active queue,
3021  * check whether to continue servicing it, or retrieve and set a new one.
3022  */
3023 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3024 {
3025         struct cfq_queue *cfqq, *new_cfqq = NULL;
3026
3027         cfqq = cfqd->active_queue;
3028         if (!cfqq)
3029                 goto new_queue;
3030
3031         if (!cfqd->rq_queued)
3032                 return NULL;
3033
3034         /*
3035          * We were waiting for group to get backlogged. Expire the queue
3036          */
3037         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3038                 goto expire;
3039
3040         /*
3041          * The active queue has run out of time, expire it and select new.
3042          */
3043         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3044                 /*
3045                  * If slice had not expired at the completion of last request
3046                  * we might not have turned on wait_busy flag. Don't expire
3047                  * the queue yet. Allow the group to get backlogged.
3048                  *
3049                  * The very fact that we have used the slice, that means we
3050                  * have been idling all along on this queue and it should be
3051                  * ok to wait for this request to complete.
3052                  */
3053                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3054                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3055                         cfqq = NULL;
3056                         goto keep_queue;
3057                 } else
3058                         goto check_group_idle;
3059         }
3060
3061         /*
3062          * The active queue has requests and isn't expired, allow it to
3063          * dispatch.
3064          */
3065         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3066                 goto keep_queue;
3067
3068         /*
3069          * If another queue has a request waiting within our mean seek
3070          * distance, let it run.  The expire code will check for close
3071          * cooperators and put the close queue at the front of the service
3072          * tree.  If possible, merge the expiring queue with the new cfqq.
3073          */
3074         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3075         if (new_cfqq) {
3076                 if (!cfqq->new_cfqq)
3077                         cfq_setup_merge(cfqq, new_cfqq);
3078                 goto expire;
3079         }
3080
3081         /*
3082          * No requests pending. If the active queue still has requests in
3083          * flight or is idling for a new request, allow either of these
3084          * conditions to happen (or time out) before selecting a new queue.
3085          */
3086         if (timer_pending(&cfqd->idle_slice_timer)) {
3087                 cfqq = NULL;
3088                 goto keep_queue;
3089         }
3090
3091         /*
3092          * This is a deep seek queue, but the device is much faster than
3093          * the queue can deliver, don't idle
3094          **/
3095         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3096             (cfq_cfqq_slice_new(cfqq) ||
3097             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3098                 cfq_clear_cfqq_deep(cfqq);
3099                 cfq_clear_cfqq_idle_window(cfqq);
3100         }
3101
3102         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3103                 cfqq = NULL;
3104                 goto keep_queue;
3105         }
3106
3107         /*
3108          * If group idle is enabled and there are requests dispatched from
3109          * this group, wait for requests to complete.
3110          */
3111 check_group_idle:
3112         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3113             cfqq->cfqg->dispatched &&
3114             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3115                 cfqq = NULL;
3116                 goto keep_queue;
3117         }
3118
3119 expire:
3120         cfq_slice_expired(cfqd, 0);
3121 new_queue:
3122         /*
3123          * Current queue expired. Check if we have to switch to a new
3124          * service tree
3125          */
3126         if (!new_cfqq)
3127                 cfq_choose_cfqg(cfqd);
3128
3129         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3130 keep_queue:
3131         return cfqq;
3132 }
3133
3134 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3135 {
3136         int dispatched = 0;
3137
3138         while (cfqq->next_rq) {
3139                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3140                 dispatched++;
3141         }
3142
3143         BUG_ON(!list_empty(&cfqq->fifo));
3144
3145         /* By default cfqq is not expired if it is empty. Do it explicitly */
3146         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3147         return dispatched;
3148 }
3149
3150 /*
3151  * Drain our current requests. Used for barriers and when switching
3152  * io schedulers on-the-fly.
3153  */
3154 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3155 {
3156         struct cfq_queue *cfqq;
3157         int dispatched = 0;
3158
3159         /* Expire the timeslice of the current active queue first */
3160         cfq_slice_expired(cfqd, 0);
3161         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3162                 __cfq_set_active_queue(cfqd, cfqq);
3163                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3164         }
3165
3166         BUG_ON(cfqd->busy_queues);
3167
3168         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3169         return dispatched;
3170 }
3171
3172 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3173         struct cfq_queue *cfqq)
3174 {
3175         /* the queue hasn't finished any request, can't estimate */
3176         if (cfq_cfqq_slice_new(cfqq))
3177                 return true;
3178         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3179                 cfqq->slice_end))
3180                 return true;
3181
3182         return false;
3183 }
3184
3185 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3186 {
3187         unsigned int max_dispatch;
3188
3189         /*
3190          * Drain async requests before we start sync IO
3191          */
3192         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3193                 return false;
3194
3195         /*
3196          * If this is an async queue and we have sync IO in flight, let it wait
3197          */
3198         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3199                 return false;
3200
3201         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3202         if (cfq_class_idle(cfqq))
3203                 max_dispatch = 1;
3204
3205         /*
3206          * Does this cfqq already have too much IO in flight?
3207          */
3208         if (cfqq->dispatched >= max_dispatch) {
3209                 bool promote_sync = false;
3210                 /*
3211                  * idle queue must always only have a single IO in flight
3212                  */
3213                 if (cfq_class_idle(cfqq))
3214                         return false;
3215
3216                 /*
3217                  * If there is only one sync queue
3218                  * we can ignore async queue here and give the sync
3219                  * queue no dispatch limit. The reason is a sync queue can
3220                  * preempt async queue, limiting the sync queue doesn't make
3221                  * sense. This is useful for aiostress test.
3222                  */
3223                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3224                         promote_sync = true;
3225
3226                 /*
3227                  * We have other queues, don't allow more IO from this one
3228                  */
3229                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3230                                 !promote_sync)
3231                         return false;
3232
3233                 /*
3234                  * Sole queue user, no limit
3235                  */
3236                 if (cfqd->busy_queues == 1 || promote_sync)
3237                         max_dispatch = -1;
3238                 else
3239                         /*
3240                          * Normally we start throttling cfqq when cfq_quantum/2
3241                          * requests have been dispatched. But we can drive
3242                          * deeper queue depths at the beginning of slice
3243                          * subjected to upper limit of cfq_quantum.
3244                          * */
3245                         max_dispatch = cfqd->cfq_quantum;
3246         }
3247
3248         /*
3249          * Async queues must wait a bit before being allowed dispatch.
3250          * We also ramp up the dispatch depth gradually for async IO,
3251          * based on the last sync IO we serviced
3252          */
3253         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3254                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3255                 unsigned int depth;
3256
3257                 depth = last_sync / cfqd->cfq_slice[1];
3258                 if (!depth && !cfqq->dispatched)
3259                         depth = 1;
3260                 if (depth < max_dispatch)
3261                         max_dispatch = depth;
3262         }
3263
3264         /*
3265          * If we're below the current max, allow a dispatch
3266          */
3267         return cfqq->dispatched < max_dispatch;
3268 }
3269
3270 /*
3271  * Dispatch a request from cfqq, moving them to the request queue
3272  * dispatch list.
3273  */
3274 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3275 {
3276         struct request *rq;
3277
3278         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3279
3280         if (!cfq_may_dispatch(cfqd, cfqq))
3281                 return false;
3282
3283         /*
3284          * follow expired path, else get first next available
3285          */
3286         rq = cfq_check_fifo(cfqq);
3287         if (!rq)
3288                 rq = cfqq->next_rq;
3289
3290         /*
3291          * insert request into driver dispatch list
3292          */
3293         cfq_dispatch_insert(cfqd->queue, rq);
3294
3295         if (!cfqd->active_cic) {
3296                 struct cfq_io_cq *cic = RQ_CIC(rq);
3297
3298                 atomic_long_inc(&cic->icq.ioc->refcount);
3299                 cfqd->active_cic = cic;
3300         }
3301
3302         return true;
3303 }
3304
3305 /*
3306  * Find the cfqq that we need to service and move a request from that to the
3307  * dispatch list
3308  */
3309 static int cfq_dispatch_requests(struct request_queue *q, int force)
3310 {
3311         struct cfq_data *cfqd = q->elevator->elevator_data;
3312         struct cfq_queue *cfqq;
3313
3314         if (!cfqd->busy_queues)
3315                 return 0;
3316
3317         if (unlikely(force))
3318                 return cfq_forced_dispatch(cfqd);
3319
3320         cfqq = cfq_select_queue(cfqd);
3321         if (!cfqq)
3322                 return 0;
3323
3324         /*
3325          * Dispatch a request from this cfqq, if it is allowed
3326          */
3327         if (!cfq_dispatch_request(cfqd, cfqq))
3328                 return 0;
3329
3330         cfqq->slice_dispatch++;
3331         cfq_clear_cfqq_must_dispatch(cfqq);
3332
3333         /*
3334          * expire an async queue immediately if it has used up its slice. idle
3335          * queue always expire after 1 dispatch round.
3336          */
3337         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3338             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3339             cfq_class_idle(cfqq))) {
3340                 cfqq->slice_end = jiffies + 1;
3341                 cfq_slice_expired(cfqd, 0);
3342         }
3343
3344         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3345         return 1;
3346 }
3347
3348 /*
3349  * task holds one reference to the queue, dropped when task exits. each rq
3350  * in-flight on this queue also holds a reference, dropped when rq is freed.
3351  *
3352  * Each cfq queue took a reference on the parent group. Drop it now.
3353  * queue lock must be held here.
3354  */
3355 static void cfq_put_queue(struct cfq_queue *cfqq)
3356 {
3357         struct cfq_data *cfqd = cfqq->cfqd;
3358         struct cfq_group *cfqg;
3359
3360         BUG_ON(cfqq->ref <= 0);
3361
3362         cfqq->ref--;
3363         if (cfqq->ref)
3364                 return;
3365
3366         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3367         BUG_ON(rb_first(&cfqq->sort_list));
3368         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3369         cfqg = cfqq->cfqg;
3370
3371         if (unlikely(cfqd->active_queue == cfqq)) {
3372                 __cfq_slice_expired(cfqd, cfqq, 0);
3373                 cfq_schedule_dispatch(cfqd);
3374         }
3375
3376         BUG_ON(cfq_cfqq_on_rr(cfqq));
3377         kmem_cache_free(cfq_pool, cfqq);
3378         cfqg_put(cfqg);
3379 }
3380
3381 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3382 {
3383         struct cfq_queue *__cfqq, *next;
3384
3385         /*
3386          * If this queue was scheduled to merge with another queue, be
3387          * sure to drop the reference taken on that queue (and others in
3388          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3389          */
3390         __cfqq = cfqq->new_cfqq;
3391         while (__cfqq) {
3392                 if (__cfqq == cfqq) {
3393                         WARN(1, "cfqq->new_cfqq loop detected\n");
3394                         break;
3395                 }
3396                 next = __cfqq->new_cfqq;
3397                 cfq_put_queue(__cfqq);
3398                 __cfqq = next;
3399         }
3400 }
3401
3402 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3403 {
3404         if (unlikely(cfqq == cfqd->active_queue)) {
3405                 __cfq_slice_expired(cfqd, cfqq, 0);
3406                 cfq_schedule_dispatch(cfqd);
3407         }
3408
3409         cfq_put_cooperator(cfqq);
3410
3411         cfq_put_queue(cfqq);
3412 }
3413
3414 static void cfq_init_icq(struct io_cq *icq)
3415 {
3416         struct cfq_io_cq *cic = icq_to_cic(icq);
3417
3418         cic->ttime.last_end_request = jiffies;
3419 }
3420
3421 static void cfq_exit_icq(struct io_cq *icq)
3422 {
3423         struct cfq_io_cq *cic = icq_to_cic(icq);
3424         struct cfq_data *cfqd = cic_to_cfqd(cic);
3425
3426         if (cic->cfqq[BLK_RW_ASYNC]) {
3427                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3428                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3429         }
3430
3431         if (cic->cfqq[BLK_RW_SYNC]) {
3432                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3433                 cic->cfqq[BLK_RW_SYNC] = NULL;
3434         }
3435 }
3436
3437 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3438 {
3439         struct task_struct *tsk = current;
3440         int ioprio_class;
3441
3442         if (!cfq_cfqq_prio_changed(cfqq))
3443                 return;
3444
3445         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3446         switch (ioprio_class) {
3447         default:
3448                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3449         case IOPRIO_CLASS_NONE:
3450                 /*
3451                  * no prio set, inherit CPU scheduling settings
3452                  */
3453                 cfqq->ioprio = task_nice_ioprio(tsk);
3454                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3455                 break;
3456         case IOPRIO_CLASS_RT:
3457                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3458                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3459                 break;
3460         case IOPRIO_CLASS_BE:
3461                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3462                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3463                 break;
3464         case IOPRIO_CLASS_IDLE:
3465                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3466                 cfqq->ioprio = 7;
3467                 cfq_clear_cfqq_idle_window(cfqq);
3468                 break;
3469         }
3470
3471         /*
3472          * keep track of original prio settings in case we have to temporarily
3473          * elevate the priority of this queue
3474          */
3475         cfqq->org_ioprio = cfqq->ioprio;
3476         cfq_clear_cfqq_prio_changed(cfqq);
3477 }
3478
3479 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3480 {
3481         int ioprio = cic->icq.ioc->ioprio;
3482         struct cfq_data *cfqd = cic_to_cfqd(cic);
3483         struct cfq_queue *cfqq;
3484
3485         /*
3486          * Check whether ioprio has changed.  The condition may trigger
3487          * spuriously on a newly created cic but there's no harm.
3488          */
3489         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3490                 return;
3491
3492         cfqq = cic->cfqq[BLK_RW_ASYNC];
3493         if (cfqq) {
3494                 struct cfq_queue *new_cfqq;
3495                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3496                                          GFP_ATOMIC);
3497                 if (new_cfqq) {
3498                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3499                         cfq_put_queue(cfqq);
3500                 }
3501         }
3502
3503         cfqq = cic->cfqq[BLK_RW_SYNC];
3504         if (cfqq)
3505                 cfq_mark_cfqq_prio_changed(cfqq);
3506
3507         cic->ioprio = ioprio;
3508 }
3509
3510 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3511                           pid_t pid, bool is_sync)
3512 {
3513         RB_CLEAR_NODE(&cfqq->rb_node);
3514         RB_CLEAR_NODE(&cfqq->p_node);
3515         INIT_LIST_HEAD(&cfqq->fifo);
3516
3517         cfqq->ref = 0;
3518         cfqq->cfqd = cfqd;
3519
3520         cfq_mark_cfqq_prio_changed(cfqq);
3521
3522         if (is_sync) {
3523                 if (!cfq_class_idle(cfqq))
3524                         cfq_mark_cfqq_idle_window(cfqq);
3525                 cfq_mark_cfqq_sync(cfqq);
3526         }
3527         cfqq->pid = pid;
3528 }
3529
3530 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3531 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3532 {
3533         struct cfq_data *cfqd = cic_to_cfqd(cic);
3534         struct cfq_queue *sync_cfqq;
3535         uint64_t id;
3536
3537         rcu_read_lock();
3538         id = bio_blkcg(bio)->id;
3539         rcu_read_unlock();
3540
3541         /*
3542          * Check whether blkcg has changed.  The condition may trigger
3543          * spuriously on a newly created cic but there's no harm.
3544          */
3545         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3546                 return;
3547
3548         sync_cfqq = cic_to_cfqq(cic, 1);
3549         if (sync_cfqq) {
3550                 /*
3551                  * Drop reference to sync queue. A new sync queue will be
3552                  * assigned in new group upon arrival of a fresh request.
3553                  */
3554                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3555                 cic_set_cfqq(cic, NULL, 1);
3556                 cfq_put_queue(sync_cfqq);
3557         }
3558
3559         cic->blkcg_id = id;
3560 }
3561 #else
3562 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3563 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3564
3565 static struct cfq_queue *
3566 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3567                      struct bio *bio, gfp_t gfp_mask)
3568 {
3569         struct blkcg *blkcg;
3570         struct cfq_queue *cfqq, *new_cfqq = NULL;
3571         struct cfq_group *cfqg;
3572
3573 retry:
3574         rcu_read_lock();
3575
3576         blkcg = bio_blkcg(bio);
3577         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3578         if (!cfqg) {
3579                 cfqq = &cfqd->oom_cfqq;
3580                 goto out;
3581         }
3582
3583         cfqq = cic_to_cfqq(cic, is_sync);
3584
3585         /*
3586          * Always try a new alloc if we fell back to the OOM cfqq
3587          * originally, since it should just be a temporary situation.
3588          */
3589         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3590                 cfqq = NULL;
3591                 if (new_cfqq) {
3592                         cfqq = new_cfqq;
3593                         new_cfqq = NULL;
3594                 } else if (gfp_mask & __GFP_WAIT) {
3595                         rcu_read_unlock();
3596                         spin_unlock_irq(cfqd->queue->queue_lock);
3597                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3598                                         gfp_mask | __GFP_ZERO,
3599                                         cfqd->queue->node);
3600                         spin_lock_irq(cfqd->queue->queue_lock);
3601                         if (new_cfqq)
3602                                 goto retry;
3603                         else
3604                                 return &cfqd->oom_cfqq;
3605                 } else {
3606                         cfqq = kmem_cache_alloc_node(cfq_pool,
3607                                         gfp_mask | __GFP_ZERO,
3608                                         cfqd->queue->node);
3609                 }
3610
3611                 if (cfqq) {
3612                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3613                         cfq_init_prio_data(cfqq, cic);
3614                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3615                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3616                 } else
3617                         cfqq = &cfqd->oom_cfqq;
3618         }
3619 out:
3620         if (new_cfqq)
3621                 kmem_cache_free(cfq_pool, new_cfqq);
3622
3623         rcu_read_unlock();
3624         return cfqq;
3625 }
3626
3627 static struct cfq_queue **
3628 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3629 {
3630         switch (ioprio_class) {
3631         case IOPRIO_CLASS_RT:
3632                 return &cfqd->async_cfqq[0][ioprio];
3633         case IOPRIO_CLASS_NONE:
3634                 ioprio = IOPRIO_NORM;
3635                 /* fall through */
3636         case IOPRIO_CLASS_BE:
3637                 return &cfqd->async_cfqq[1][ioprio];
3638         case IOPRIO_CLASS_IDLE:
3639                 return &cfqd->async_idle_cfqq;
3640         default:
3641                 BUG();
3642         }
3643 }
3644
3645 static struct cfq_queue *
3646 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3647               struct bio *bio, gfp_t gfp_mask)
3648 {
3649         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3650         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3651         struct cfq_queue **async_cfqq = NULL;
3652         struct cfq_queue *cfqq = NULL;
3653
3654         if (!is_sync) {
3655                 if (!ioprio_valid(cic->ioprio)) {
3656                         struct task_struct *tsk = current;
3657                         ioprio = task_nice_ioprio(tsk);
3658                         ioprio_class = task_nice_ioclass(tsk);
3659                 }
3660                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3661                 cfqq = *async_cfqq;
3662         }
3663
3664         if (!cfqq)
3665                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3666
3667         /*
3668          * pin the queue now that it's allocated, scheduler exit will prune it
3669          */
3670         if (!is_sync && !(*async_cfqq)) {
3671                 cfqq->ref++;
3672                 *async_cfqq = cfqq;
3673         }
3674
3675         cfqq->ref++;
3676         return cfqq;
3677 }
3678
3679 static void
3680 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3681 {
3682         unsigned long elapsed = jiffies - ttime->last_end_request;
3683         elapsed = min(elapsed, 2UL * slice_idle);
3684
3685         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3686         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3687         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3688 }
3689
3690 static void
3691 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3692                         struct cfq_io_cq *cic)
3693 {
3694         if (cfq_cfqq_sync(cfqq)) {
3695                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3696                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3697                         cfqd->cfq_slice_idle);
3698         }
3699 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3700         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3701 #endif
3702 }
3703
3704 static void
3705 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3706                        struct request *rq)
3707 {
3708         sector_t sdist = 0;
3709         sector_t n_sec = blk_rq_sectors(rq);
3710         if (cfqq->last_request_pos) {
3711                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3712                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3713                 else
3714                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3715         }
3716
3717         cfqq->seek_history <<= 1;
3718         if (blk_queue_nonrot(cfqd->queue))
3719                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3720         else
3721                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3722 }
3723
3724 /*
3725  * Disable idle window if the process thinks too long or seeks so much that
3726  * it doesn't matter
3727  */
3728 static void
3729 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3730                        struct cfq_io_cq *cic)
3731 {
3732         int old_idle, enable_idle;
3733
3734         /*
3735          * Don't idle for async or idle io prio class
3736          */
3737         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3738                 return;
3739
3740         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3741
3742         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3743                 cfq_mark_cfqq_deep(cfqq);
3744
3745         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3746                 enable_idle = 0;
3747         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3748                  !cfqd->cfq_slice_idle ||
3749                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3750                 enable_idle = 0;
3751         else if (sample_valid(cic->ttime.ttime_samples)) {
3752                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3753                         enable_idle = 0;
3754                 else
3755                         enable_idle = 1;
3756         }
3757
3758         if (old_idle != enable_idle) {
3759                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3760                 if (enable_idle)
3761                         cfq_mark_cfqq_idle_window(cfqq);
3762                 else
3763                         cfq_clear_cfqq_idle_window(cfqq);
3764         }
3765 }
3766
3767 /*
3768  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3769  * no or if we aren't sure, a 1 will cause a preempt.
3770  */
3771 static bool
3772 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3773                    struct request *rq)
3774 {
3775         struct cfq_queue *cfqq;
3776
3777         cfqq = cfqd->active_queue;
3778         if (!cfqq)
3779                 return false;
3780
3781         if (cfq_class_idle(new_cfqq))
3782                 return false;
3783
3784         if (cfq_class_idle(cfqq))
3785                 return true;
3786
3787         /*
3788          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3789          */
3790         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3791                 return false;
3792
3793         /*
3794          * if the new request is sync, but the currently running queue is
3795          * not, let the sync request have priority.
3796          */
3797         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3798                 return true;
3799
3800         if (new_cfqq->cfqg != cfqq->cfqg)
3801                 return false;
3802
3803         if (cfq_slice_used(cfqq))
3804                 return true;
3805
3806         /* Allow preemption only if we are idling on sync-noidle tree */
3807         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3808             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3809             new_cfqq->service_tree->count == 2 &&
3810             RB_EMPTY_ROOT(&cfqq->sort_list))
3811                 return true;
3812
3813         /*
3814          * So both queues are sync. Let the new request get disk time if
3815          * it's a metadata request and the current queue is doing regular IO.
3816          */
3817         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3818                 return true;
3819
3820         /*
3821          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3822          */
3823         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3824                 return true;
3825
3826         /* An idle queue should not be idle now for some reason */
3827         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3828                 return true;
3829
3830         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3831                 return false;
3832
3833         /*
3834          * if this request is as-good as one we would expect from the
3835          * current cfqq, let it preempt
3836          */
3837         if (cfq_rq_close(cfqd, cfqq, rq))
3838                 return true;
3839
3840         return false;
3841 }
3842
3843 /*
3844  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3845  * let it have half of its nominal slice.
3846  */
3847 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3848 {
3849         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3850
3851         cfq_log_cfqq(cfqd, cfqq, "preempt");
3852         cfq_slice_expired(cfqd, 1);
3853
3854         /*
3855          * workload type is changed, don't save slice, otherwise preempt
3856          * doesn't happen
3857          */
3858         if (old_type != cfqq_type(cfqq))
3859                 cfqq->cfqg->saved_wl_slice = 0;
3860
3861         /*
3862          * Put the new queue at the front of the of the current list,
3863          * so we know that it will be selected next.
3864          */
3865         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3866
3867         cfq_service_tree_add(cfqd, cfqq, 1);
3868
3869         cfqq->slice_end = 0;
3870         cfq_mark_cfqq_slice_new(cfqq);
3871 }
3872
3873 /*
3874  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3875  * something we should do about it
3876  */
3877 static void
3878 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3879                 struct request *rq)
3880 {
3881         struct cfq_io_cq *cic = RQ_CIC(rq);
3882
3883         cfqd->rq_queued++;
3884         if (rq->cmd_flags & REQ_PRIO)
3885                 cfqq->prio_pending++;
3886
3887         cfq_update_io_thinktime(cfqd, cfqq, cic);
3888         cfq_update_io_seektime(cfqd, cfqq, rq);
3889         cfq_update_idle_window(cfqd, cfqq, cic);
3890
3891         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3892
3893         if (cfqq == cfqd->active_queue) {
3894                 /*
3895                  * Remember that we saw a request from this process, but
3896                  * don't start queuing just yet. Otherwise we risk seeing lots
3897                  * of tiny requests, because we disrupt the normal plugging
3898                  * and merging. If the request is already larger than a single
3899                  * page, let it rip immediately. For that case we assume that
3900                  * merging is already done. Ditto for a busy system that
3901                  * has other work pending, don't risk delaying until the
3902                  * idle timer unplug to continue working.
3903                  */
3904                 if (cfq_cfqq_wait_request(cfqq)) {
3905                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3906                             cfqd->busy_queues > 1) {
3907                                 cfq_del_timer(cfqd, cfqq);
3908                                 cfq_clear_cfqq_wait_request(cfqq);
3909                                 __blk_run_queue(cfqd->queue);
3910                         } else {
3911                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3912                                 cfq_mark_cfqq_must_dispatch(cfqq);
3913                         }
3914                 }
3915         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3916                 /*
3917                  * not the active queue - expire current slice if it is
3918                  * idle and has expired it's mean thinktime or this new queue
3919                  * has some old slice time left and is of higher priority or
3920                  * this new queue is RT and the current one is BE
3921                  */
3922                 cfq_preempt_queue(cfqd, cfqq);
3923                 __blk_run_queue(cfqd->queue);
3924         }
3925 }
3926
3927 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3928 {
3929         struct cfq_data *cfqd = q->elevator->elevator_data;
3930         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3931
3932         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3933         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3934
3935         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3936         list_add_tail(&rq->queuelist, &cfqq->fifo);
3937         cfq_add_rq_rb(rq);
3938         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3939                                  rq->cmd_flags);
3940         cfq_rq_enqueued(cfqd, cfqq, rq);
3941 }
3942
3943 /*
3944  * Update hw_tag based on peak queue depth over 50 samples under
3945  * sufficient load.
3946  */
3947 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3948 {
3949         struct cfq_queue *cfqq = cfqd->active_queue;
3950
3951         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3952                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3953
3954         if (cfqd->hw_tag == 1)
3955                 return;
3956
3957         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3958             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3959                 return;
3960
3961         /*
3962          * If active queue hasn't enough requests and can idle, cfq might not
3963          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3964          * case
3965          */
3966         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3967             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3968             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3969                 return;
3970
3971         if (cfqd->hw_tag_samples++ < 50)
3972                 return;
3973
3974         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3975                 cfqd->hw_tag = 1;
3976         else
3977                 cfqd->hw_tag = 0;
3978 }
3979
3980 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3981 {
3982         struct cfq_io_cq *cic = cfqd->active_cic;
3983
3984         /* If the queue already has requests, don't wait */
3985         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3986                 return false;
3987
3988         /* If there are other queues in the group, don't wait */
3989         if (cfqq->cfqg->nr_cfqq > 1)
3990                 return false;
3991
3992         /* the only queue in the group, but think time is big */
3993         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3994                 return false;
3995
3996         if (cfq_slice_used(cfqq))
3997                 return true;
3998
3999         /* if slice left is less than think time, wait busy */
4000         if (cic && sample_valid(cic->ttime.ttime_samples)
4001             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4002                 return true;
4003
4004         /*
4005          * If think times is less than a jiffy than ttime_mean=0 and above
4006          * will not be true. It might happen that slice has not expired yet
4007          * but will expire soon (4-5 ns) during select_queue(). To cover the
4008          * case where think time is less than a jiffy, mark the queue wait
4009          * busy if only 1 jiffy is left in the slice.
4010          */
4011         if (cfqq->slice_end - jiffies == 1)
4012                 return true;
4013
4014         return false;
4015 }
4016
4017 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4018 {
4019         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4020         struct cfq_data *cfqd = cfqq->cfqd;
4021         const int sync = rq_is_sync(rq);
4022         unsigned long now;
4023
4024         now = jiffies;
4025         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4026                      !!(rq->cmd_flags & REQ_NOIDLE));
4027
4028         cfq_update_hw_tag(cfqd);
4029
4030         WARN_ON(!cfqd->rq_in_driver);
4031         WARN_ON(!cfqq->dispatched);
4032         cfqd->rq_in_driver--;
4033         cfqq->dispatched--;
4034         (RQ_CFQG(rq))->dispatched--;
4035         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4036                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4037
4038         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4039
4040         if (sync) {
4041                 struct cfq_rb_root *st;
4042
4043                 RQ_CIC(rq)->ttime.last_end_request = now;
4044
4045                 if (cfq_cfqq_on_rr(cfqq))
4046                         st = cfqq->service_tree;
4047                 else
4048                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4049                                         cfqq_type(cfqq));
4050
4051                 st->ttime.last_end_request = now;
4052                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4053                         cfqd->last_delayed_sync = now;
4054         }
4055
4056 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4057         cfqq->cfqg->ttime.last_end_request = now;
4058 #endif
4059
4060         /*
4061          * If this is the active queue, check if it needs to be expired,
4062          * or if we want to idle in case it has no pending requests.
4063          */
4064         if (cfqd->active_queue == cfqq) {
4065                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4066
4067                 if (cfq_cfqq_slice_new(cfqq)) {
4068                         cfq_set_prio_slice(cfqd, cfqq);
4069                         cfq_clear_cfqq_slice_new(cfqq);
4070                 }
4071
4072                 /*
4073                  * Should we wait for next request to come in before we expire
4074                  * the queue.
4075                  */
4076                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4077                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4078                         if (!cfqd->cfq_slice_idle)
4079                                 extend_sl = cfqd->cfq_group_idle;
4080                         cfqq->slice_end = jiffies + extend_sl;
4081                         cfq_mark_cfqq_wait_busy(cfqq);
4082                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4083                 }
4084
4085                 /*
4086                  * Idling is not enabled on:
4087                  * - expired queues
4088                  * - idle-priority queues
4089                  * - async queues
4090                  * - queues with still some requests queued
4091                  * - when there is a close cooperator
4092                  */
4093                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4094                         cfq_slice_expired(cfqd, 1);
4095                 else if (sync && cfqq_empty &&
4096                          !cfq_close_cooperator(cfqd, cfqq)) {
4097                         cfq_arm_slice_timer(cfqd);
4098                 }
4099         }
4100
4101         if (!cfqd->rq_in_driver)
4102                 cfq_schedule_dispatch(cfqd);
4103 }
4104
4105 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4106 {
4107         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4108                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4109                 return ELV_MQUEUE_MUST;
4110         }
4111
4112         return ELV_MQUEUE_MAY;
4113 }
4114
4115 static int cfq_may_queue(struct request_queue *q, int rw)
4116 {
4117         struct cfq_data *cfqd = q->elevator->elevator_data;
4118         struct task_struct *tsk = current;
4119         struct cfq_io_cq *cic;
4120         struct cfq_queue *cfqq;
4121
4122         /*
4123          * don't force setup of a queue from here, as a call to may_queue
4124          * does not necessarily imply that a request actually will be queued.
4125          * so just lookup a possibly existing queue, or return 'may queue'
4126          * if that fails
4127          */
4128         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4129         if (!cic)
4130                 return ELV_MQUEUE_MAY;
4131
4132         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4133         if (cfqq) {
4134                 cfq_init_prio_data(cfqq, cic);
4135
4136                 return __cfq_may_queue(cfqq);
4137         }
4138
4139         return ELV_MQUEUE_MAY;
4140 }
4141
4142 /*
4143  * queue lock held here
4144  */
4145 static void cfq_put_request(struct request *rq)
4146 {
4147         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4148
4149         if (cfqq) {
4150                 const int rw = rq_data_dir(rq);
4151
4152                 BUG_ON(!cfqq->allocated[rw]);
4153                 cfqq->allocated[rw]--;
4154
4155                 /* Put down rq reference on cfqg */
4156                 cfqg_put(RQ_CFQG(rq));
4157                 rq->elv.priv[0] = NULL;
4158                 rq->elv.priv[1] = NULL;
4159
4160                 cfq_put_queue(cfqq);
4161         }
4162 }
4163
4164 static struct cfq_queue *
4165 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4166                 struct cfq_queue *cfqq)
4167 {
4168         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4169         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4170         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4171         cfq_put_queue(cfqq);
4172         return cic_to_cfqq(cic, 1);
4173 }
4174
4175 /*
4176  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4177  * was the last process referring to said cfqq.
4178  */
4179 static struct cfq_queue *
4180 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4181 {
4182         if (cfqq_process_refs(cfqq) == 1) {
4183                 cfqq->pid = current->pid;
4184                 cfq_clear_cfqq_coop(cfqq);
4185                 cfq_clear_cfqq_split_coop(cfqq);
4186                 return cfqq;
4187         }
4188
4189         cic_set_cfqq(cic, NULL, 1);
4190
4191         cfq_put_cooperator(cfqq);
4192
4193         cfq_put_queue(cfqq);
4194         return NULL;
4195 }
4196 /*
4197  * Allocate cfq data structures associated with this request.
4198  */
4199 static int
4200 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4201                 gfp_t gfp_mask)
4202 {
4203         struct cfq_data *cfqd = q->elevator->elevator_data;
4204         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4205         const int rw = rq_data_dir(rq);
4206         const bool is_sync = rq_is_sync(rq);
4207         struct cfq_queue *cfqq;
4208
4209         might_sleep_if(gfp_mask & __GFP_WAIT);
4210
4211         spin_lock_irq(q->queue_lock);
4212
4213         check_ioprio_changed(cic, bio);
4214         check_blkcg_changed(cic, bio);
4215 new_queue:
4216         cfqq = cic_to_cfqq(cic, is_sync);
4217         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4218                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4219                 cic_set_cfqq(cic, cfqq, is_sync);
4220         } else {
4221                 /*
4222                  * If the queue was seeky for too long, break it apart.
4223                  */
4224                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4225                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4226                         cfqq = split_cfqq(cic, cfqq);
4227                         if (!cfqq)
4228                                 goto new_queue;
4229                 }
4230
4231                 /*
4232                  * Check to see if this queue is scheduled to merge with
4233                  * another, closely cooperating queue.  The merging of
4234                  * queues happens here as it must be done in process context.
4235                  * The reference on new_cfqq was taken in merge_cfqqs.
4236                  */
4237                 if (cfqq->new_cfqq)
4238                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4239         }
4240
4241         cfqq->allocated[rw]++;
4242
4243         cfqq->ref++;
4244         cfqg_get(cfqq->cfqg);
4245         rq->elv.priv[0] = cfqq;
4246         rq->elv.priv[1] = cfqq->cfqg;
4247         spin_unlock_irq(q->queue_lock);
4248         return 0;
4249 }
4250
4251 static void cfq_kick_queue(struct work_struct *work)
4252 {
4253         struct cfq_data *cfqd =
4254                 container_of(work, struct cfq_data, unplug_work);
4255         struct request_queue *q = cfqd->queue;
4256
4257         spin_lock_irq(q->queue_lock);
4258         __blk_run_queue(cfqd->queue);
4259         spin_unlock_irq(q->queue_lock);
4260 }
4261
4262 /*
4263  * Timer running if the active_queue is currently idling inside its time slice
4264  */
4265 static void cfq_idle_slice_timer(unsigned long data)
4266 {
4267         struct cfq_data *cfqd = (struct cfq_data *) data;
4268         struct cfq_queue *cfqq;
4269         unsigned long flags;
4270         int timed_out = 1;
4271
4272         cfq_log(cfqd, "idle timer fired");
4273
4274         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4275
4276         cfqq = cfqd->active_queue;
4277         if (cfqq) {
4278                 timed_out = 0;
4279
4280                 /*
4281                  * We saw a request before the queue expired, let it through
4282                  */
4283                 if (cfq_cfqq_must_dispatch(cfqq))
4284                         goto out_kick;
4285
4286                 /*
4287                  * expired
4288                  */
4289                 if (cfq_slice_used(cfqq))
4290                         goto expire;
4291
4292                 /*
4293                  * only expire and reinvoke request handler, if there are
4294                  * other queues with pending requests
4295                  */
4296                 if (!cfqd->busy_queues)
4297                         goto out_cont;
4298
4299                 /*
4300                  * not expired and it has a request pending, let it dispatch
4301                  */
4302                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4303                         goto out_kick;
4304
4305                 /*
4306                  * Queue depth flag is reset only when the idle didn't succeed
4307                  */
4308                 cfq_clear_cfqq_deep(cfqq);
4309         }
4310 expire:
4311         cfq_slice_expired(cfqd, timed_out);
4312 out_kick:
4313         cfq_schedule_dispatch(cfqd);
4314 out_cont:
4315         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4316 }
4317
4318 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4319 {
4320         del_timer_sync(&cfqd->idle_slice_timer);
4321         cancel_work_sync(&cfqd->unplug_work);
4322 }
4323
4324 static void cfq_put_async_queues(struct cfq_data *cfqd)
4325 {
4326         int i;
4327
4328         for (i = 0; i < IOPRIO_BE_NR; i++) {
4329                 if (cfqd->async_cfqq[0][i])
4330                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4331                 if (cfqd->async_cfqq[1][i])
4332                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4333         }
4334
4335         if (cfqd->async_idle_cfqq)
4336                 cfq_put_queue(cfqd->async_idle_cfqq);
4337 }
4338
4339 static void cfq_exit_queue(struct elevator_queue *e)
4340 {
4341         struct cfq_data *cfqd = e->elevator_data;
4342         struct request_queue *q = cfqd->queue;
4343
4344         cfq_shutdown_timer_wq(cfqd);
4345
4346         spin_lock_irq(q->queue_lock);
4347
4348         if (cfqd->active_queue)
4349                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4350
4351         cfq_put_async_queues(cfqd);
4352
4353         spin_unlock_irq(q->queue_lock);
4354
4355         cfq_shutdown_timer_wq(cfqd);
4356
4357 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4358         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4359 #else
4360         kfree(cfqd->root_group);
4361 #endif
4362         kfree(cfqd);
4363 }
4364
4365 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4366 {
4367         struct cfq_data *cfqd;
4368         struct blkcg_gq *blkg __maybe_unused;
4369         int i, ret;
4370         struct elevator_queue *eq;
4371
4372         eq = elevator_alloc(q, e);
4373         if (!eq)
4374                 return -ENOMEM;
4375
4376         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4377         if (!cfqd) {
4378                 kobject_put(&eq->kobj);
4379                 return -ENOMEM;
4380         }
4381         eq->elevator_data = cfqd;
4382
4383         cfqd->queue = q;
4384         spin_lock_irq(q->queue_lock);
4385         q->elevator = eq;
4386         spin_unlock_irq(q->queue_lock);
4387
4388         /* Init root service tree */
4389         cfqd->grp_service_tree = CFQ_RB_ROOT;
4390
4391         /* Init root group and prefer root group over other groups by default */
4392 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4393         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4394         if (ret)
4395                 goto out_free;
4396
4397         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4398 #else
4399         ret = -ENOMEM;
4400         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4401                                         GFP_KERNEL, cfqd->queue->node);
4402         if (!cfqd->root_group)
4403                 goto out_free;
4404
4405         cfq_init_cfqg_base(cfqd->root_group);
4406 #endif
4407         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4408         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4409
4410         /*
4411          * Not strictly needed (since RB_ROOT just clears the node and we
4412          * zeroed cfqd on alloc), but better be safe in case someone decides
4413          * to add magic to the rb code
4414          */
4415         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4416                 cfqd->prio_trees[i] = RB_ROOT;
4417
4418         /*
4419          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4420          * Grab a permanent reference to it, so that the normal code flow
4421          * will not attempt to free it.  oom_cfqq is linked to root_group
4422          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4423          * the reference from linking right away.
4424          */
4425         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4426         cfqd->oom_cfqq.ref++;
4427
4428         spin_lock_irq(q->queue_lock);
4429         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4430         cfqg_put(cfqd->root_group);
4431         spin_unlock_irq(q->queue_lock);
4432
4433         init_timer(&cfqd->idle_slice_timer);
4434         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4435         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4436
4437         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4438
4439         cfqd->cfq_quantum = cfq_quantum;
4440         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4441         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4442         cfqd->cfq_back_max = cfq_back_max;
4443         cfqd->cfq_back_penalty = cfq_back_penalty;
4444         cfqd->cfq_slice[0] = cfq_slice_async;
4445         cfqd->cfq_slice[1] = cfq_slice_sync;
4446         cfqd->cfq_target_latency = cfq_target_latency;
4447         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4448         cfqd->cfq_slice_idle = cfq_slice_idle;
4449         cfqd->cfq_group_idle = cfq_group_idle;
4450         cfqd->cfq_latency = 1;
4451         cfqd->hw_tag = -1;
4452         /*
4453          * we optimistically start assuming sync ops weren't delayed in last
4454          * second, in order to have larger depth for async operations.
4455          */
4456         cfqd->last_delayed_sync = jiffies - HZ;
4457         return 0;
4458
4459 out_free:
4460         kfree(cfqd);
4461         kobject_put(&eq->kobj);
4462         return ret;
4463 }
4464
4465 /*
4466  * sysfs parts below -->
4467  */
4468 static ssize_t
4469 cfq_var_show(unsigned int var, char *page)
4470 {
4471         return sprintf(page, "%d\n", var);
4472 }
4473
4474 static ssize_t
4475 cfq_var_store(unsigned int *var, const char *page, size_t count)
4476 {
4477         char *p = (char *) page;
4478
4479         *var = simple_strtoul(p, &p, 10);
4480         return count;
4481 }
4482
4483 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4484 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4485 {                                                                       \
4486         struct cfq_data *cfqd = e->elevator_data;                       \
4487         unsigned int __data = __VAR;                                    \
4488         if (__CONV)                                                     \
4489                 __data = jiffies_to_msecs(__data);                      \
4490         return cfq_var_show(__data, (page));                            \
4491 }
4492 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4493 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4494 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4495 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4496 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4497 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4498 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4499 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4500 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4501 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4502 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4503 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4504 #undef SHOW_FUNCTION
4505
4506 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4507 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4508 {                                                                       \
4509         struct cfq_data *cfqd = e->elevator_data;                       \
4510         unsigned int __data;                                            \
4511         int ret = cfq_var_store(&__data, (page), count);                \
4512         if (__data < (MIN))                                             \
4513                 __data = (MIN);                                         \
4514         else if (__data > (MAX))                                        \
4515                 __data = (MAX);                                         \
4516         if (__CONV)                                                     \
4517                 *(__PTR) = msecs_to_jiffies(__data);                    \
4518         else                                                            \
4519                 *(__PTR) = __data;                                      \
4520         return ret;                                                     \
4521 }
4522 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4523 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4524                 UINT_MAX, 1);
4525 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4526                 UINT_MAX, 1);
4527 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4528 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4529                 UINT_MAX, 0);
4530 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4531 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4532 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4533 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4534 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4535                 UINT_MAX, 0);
4536 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4537 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4538 #undef STORE_FUNCTION
4539
4540 #define CFQ_ATTR(name) \
4541         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4542
4543 static struct elv_fs_entry cfq_attrs[] = {
4544         CFQ_ATTR(quantum),
4545         CFQ_ATTR(fifo_expire_sync),
4546         CFQ_ATTR(fifo_expire_async),
4547         CFQ_ATTR(back_seek_max),
4548         CFQ_ATTR(back_seek_penalty),
4549         CFQ_ATTR(slice_sync),
4550         CFQ_ATTR(slice_async),
4551         CFQ_ATTR(slice_async_rq),
4552         CFQ_ATTR(slice_idle),
4553         CFQ_ATTR(group_idle),
4554         CFQ_ATTR(low_latency),
4555         CFQ_ATTR(target_latency),
4556         __ATTR_NULL
4557 };
4558
4559 static struct elevator_type iosched_cfq = {
4560         .ops = {
4561                 .elevator_merge_fn =            cfq_merge,
4562                 .elevator_merged_fn =           cfq_merged_request,
4563                 .elevator_merge_req_fn =        cfq_merged_requests,
4564                 .elevator_allow_merge_fn =      cfq_allow_merge,
4565                 .elevator_bio_merged_fn =       cfq_bio_merged,
4566                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4567                 .elevator_add_req_fn =          cfq_insert_request,
4568                 .elevator_activate_req_fn =     cfq_activate_request,
4569                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4570                 .elevator_completed_req_fn =    cfq_completed_request,
4571                 .elevator_former_req_fn =       elv_rb_former_request,
4572                 .elevator_latter_req_fn =       elv_rb_latter_request,
4573                 .elevator_init_icq_fn =         cfq_init_icq,
4574                 .elevator_exit_icq_fn =         cfq_exit_icq,
4575                 .elevator_set_req_fn =          cfq_set_request,
4576                 .elevator_put_req_fn =          cfq_put_request,
4577                 .elevator_may_queue_fn =        cfq_may_queue,
4578                 .elevator_init_fn =             cfq_init_queue,
4579                 .elevator_exit_fn =             cfq_exit_queue,
4580         },
4581         .icq_size       =       sizeof(struct cfq_io_cq),
4582         .icq_align      =       __alignof__(struct cfq_io_cq),
4583         .elevator_attrs =       cfq_attrs,
4584         .elevator_name  =       "cfq",
4585         .elevator_owner =       THIS_MODULE,
4586 };
4587
4588 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4589 static struct blkcg_policy blkcg_policy_cfq = {
4590         .pd_size                = sizeof(struct cfq_group),
4591         .cftypes                = cfq_blkcg_files,
4592
4593         .pd_init_fn             = cfq_pd_init,
4594         .pd_offline_fn          = cfq_pd_offline,
4595         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4596 };
4597 #endif
4598
4599 static int __init cfq_init(void)
4600 {
4601         int ret;
4602
4603         /*
4604          * could be 0 on HZ < 1000 setups
4605          */
4606         if (!cfq_slice_async)
4607                 cfq_slice_async = 1;
4608         if (!cfq_slice_idle)
4609                 cfq_slice_idle = 1;
4610
4611 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4612         if (!cfq_group_idle)
4613                 cfq_group_idle = 1;
4614
4615         ret = blkcg_policy_register(&blkcg_policy_cfq);
4616         if (ret)
4617                 return ret;
4618 #else
4619         cfq_group_idle = 0;
4620 #endif
4621
4622         ret = -ENOMEM;
4623         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4624         if (!cfq_pool)
4625                 goto err_pol_unreg;
4626
4627         ret = elv_register(&iosched_cfq);
4628         if (ret)
4629                 goto err_free_pool;
4630
4631         return 0;
4632
4633 err_free_pool:
4634         kmem_cache_destroy(cfq_pool);
4635 err_pol_unreg:
4636 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4637         blkcg_policy_unregister(&blkcg_policy_cfq);
4638 #endif
4639         return ret;
4640 }
4641
4642 static void __exit cfq_exit(void)
4643 {
4644 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4645         blkcg_policy_unregister(&blkcg_policy_cfq);
4646 #endif
4647         elv_unregister(&iosched_cfq);
4648         kmem_cache_destroy(cfq_pool);
4649 }
4650
4651 module_init(cfq_init);
4652 module_exit(cfq_exit);
4653
4654 MODULE_AUTHOR("Jens Axboe");
4655 MODULE_LICENSE("GPL");
4656 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");