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