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 <linux/blk-cgroup.h>
  18 #include "blk.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 /* blkio-related constants */
  71 #define CFQ_WEIGHT_LEGACY_MIN   10
  72 #define CFQ_WEIGHT_LEGACY_DFL   500
  73 #define CFQ_WEIGHT_LEGACY_MAX   1000
  74
  75 struct cfq_ttime {
  76         unsigned long last_end_request;
  77
  78         unsigned long ttime_total;
  79         unsigned long ttime_samples;
  80         unsigned long ttime_mean;
  81 };
  82
  83 /*
  84  * Most of our rbtree usage is for sorting with min extraction, so
  85  * if we cache the leftmost node we don't have to walk down the tree
  86  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  87  * move this into the elevator for the rq sorting as well.
  88  */
  89 struct cfq_rb_root {
  90         struct rb_root rb;
  91         struct rb_node *left;
  92         unsigned count;
  93         u64 min_vdisktime;
  94         struct cfq_ttime ttime;
  95 };
  96 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  97                         .ttime = {.last_end_request = jiffies,},}
  98
  99 /*
 100  * Per process-grouping structure
 101  */
 102 struct cfq_queue {
 103         /* reference count */
 104         int ref;
 105         /* various state flags, see below */
 106         unsigned int flags;
 107         /* parent cfq_data */
 108         struct cfq_data *cfqd;
 109         /* service_tree member */
 110         struct rb_node rb_node;
 111         /* service_tree key */
 112         unsigned long rb_key;
 113         /* prio tree member */
 114         struct rb_node p_node;
 115         /* prio tree root we belong to, if any */
 116         struct rb_root *p_root;
 117         /* sorted list of pending requests */
 118         struct rb_root sort_list;
 119         /* if fifo isn't expired, next request to serve */
 120         struct request *next_rq;
 121         /* requests queued in sort_list */
 122         int queued[2];
 123         /* currently allocated requests */
 124         int allocated[2];
 125         /* fifo list of requests in sort_list */
 126         struct list_head fifo;
 127
 128         /* time when queue got scheduled in to dispatch first request. */
 129         unsigned long dispatch_start;
 130         unsigned int allocated_slice;
 131         unsigned int slice_dispatch;
 132         /* time when first request from queue completed and slice started. */
 133         unsigned long slice_start;
 134         unsigned long slice_end;
 135         long slice_resid;
 136
 137         /* pending priority requests */
 138         int prio_pending;
 139         /* number of requests that are on the dispatch list or inside driver */
 140         int dispatched;
 141
 142         /* io prio of this group */
 143         unsigned short ioprio, org_ioprio;
 144         unsigned short ioprio_class;
 145
 146         pid_t pid;
 147
 148         u32 seek_history;
 149         sector_t last_request_pos;
 150
 151         struct cfq_rb_root *service_tree;
 152         struct cfq_queue *new_cfqq;
 153         struct cfq_group *cfqg;
 154         /* Number of sectors dispatched from queue in single dispatch round */
 155         unsigned long nr_sectors;
 156 };
 157
 158 /*
 159  * First index in the service_trees.
 160  * IDLE is handled separately, so it has negative index
 161  */
 162 enum wl_class_t {
 163         BE_WORKLOAD = 0,
 164         RT_WORKLOAD = 1,
 165         IDLE_WORKLOAD = 2,
 166         CFQ_PRIO_NR,
 167 };
 168
 169 /*
 170  * Second index in the service_trees.
 171  */
 172 enum wl_type_t {
 173         ASYNC_WORKLOAD = 0,
 174         SYNC_NOIDLE_WORKLOAD = 1,
 175         SYNC_WORKLOAD = 2
 176 };
 177
 178 struct cfqg_stats {
 179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 180         /* number of ios merged */
 181         struct blkg_rwstat              merged;
 182         /* total time spent on device in ns, may not be accurate w/ queueing */
 183         struct blkg_rwstat              service_time;
 184         /* total time spent waiting in scheduler queue in ns */
 185         struct blkg_rwstat              wait_time;
 186         /* number of IOs queued up */
 187         struct blkg_rwstat              queued;
 188         /* total disk time and nr sectors dispatched by this group */
 189         struct blkg_stat                time;
 190 #ifdef CONFIG_DEBUG_BLK_CGROUP
 191         /* time not charged to this cgroup */
 192         struct blkg_stat                unaccounted_time;
 193         /* sum of number of ios queued across all samples */
 194         struct blkg_stat                avg_queue_size_sum;
 195         /* count of samples taken for average */
 196         struct blkg_stat                avg_queue_size_samples;
 197         /* how many times this group has been removed from service tree */
 198         struct blkg_stat                dequeue;
 199         /* total time spent waiting for it to be assigned a timeslice. */
 200         struct blkg_stat                group_wait_time;
 201         /* time spent idling for this blkcg_gq */
 202         struct blkg_stat                idle_time;
 203         /* total time with empty current active q with other requests queued */
 204         struct blkg_stat                empty_time;
 205         /* fields after this shouldn't be cleared on stat reset */
 206         uint64_t                        start_group_wait_time;
 207         uint64_t                        start_idle_time;
 208         uint64_t                        start_empty_time;
 209         uint16_t                        flags;
 210 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 211 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 212 };
 213
 214 /* Per-cgroup data */
 215 struct cfq_group_data {
 216         /* must be the first member */
 217         struct blkcg_policy_data cpd;
 218
 219         unsigned int weight;
 220         unsigned int leaf_weight;
 221 };
 222
 223 /* This is per cgroup per device grouping structure */
 224 struct cfq_group {
 225         /* must be the first member */
 226         struct blkg_policy_data pd;
 227
 228         /* group service_tree member */
 229         struct rb_node rb_node;
 230
 231         /* group service_tree key */
 232         u64 vdisktime;
 233
 234         /*
 235          * The number of active cfqgs and sum of their weights under this
 236          * cfqg.  This covers this cfqg's leaf_weight and all children's
 237          * weights, but does not cover weights of further descendants.
 238          *
 239          * If a cfqg is on the service tree, it's active.  An active cfqg
 240          * also activates its parent and contributes to the children_weight
 241          * of the parent.
 242          */
 243         int nr_active;
 244         unsigned int children_weight;
 245
 246         /*
 247          * vfraction is the fraction of vdisktime that the tasks in this
 248          * cfqg are entitled to.  This is determined by compounding the
 249          * ratios walking up from this cfqg to the root.
 250          *
 251          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 252          * vfractions on a service tree is approximately 1.  The sum may
 253          * deviate a bit due to rounding errors and fluctuations caused by
 254          * cfqgs entering and leaving the service tree.
 255          */
 256         unsigned int vfraction;
 257
 258         /*
 259          * There are two weights - (internal) weight is the weight of this
 260          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 261          * this cfqg against the child cfqgs.  For the root cfqg, both
 262          * weights are kept in sync for backward compatibility.
 263          */
 264         unsigned int weight;
 265         unsigned int new_weight;
 266         unsigned int dev_weight;
 267
 268         unsigned int leaf_weight;
 269         unsigned int new_leaf_weight;
 270         unsigned int dev_leaf_weight;
 271
 272         /* number of cfqq currently on this group */
 273         int nr_cfqq;
 274
 275         /*
 276          * Per group busy queues average. Useful for workload slice calc. We
 277          * create the array for each prio class but at run time it is used
 278          * only for RT and BE class and slot for IDLE class remains unused.
 279          * This is primarily done to avoid confusion and a gcc warning.
 280          */
 281         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 282         /*
 283          * rr lists of queues with requests. We maintain service trees for
 284          * RT and BE classes. These trees are subdivided in subclasses
 285          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 286          * class there is no subclassification and all the cfq queues go on
 287          * a single tree service_tree_idle.
 288          * Counts are embedded in the cfq_rb_root
 289          */
 290         struct cfq_rb_root service_trees[2][3];
 291         struct cfq_rb_root service_tree_idle;
 292
 293         unsigned long saved_wl_slice;
 294         enum wl_type_t saved_wl_type;
 295         enum wl_class_t saved_wl_class;
 296
 297         /* number of requests that are on the dispatch list or inside driver */
 298         int dispatched;
 299         struct cfq_ttime ttime;
 300         struct cfqg_stats stats;        /* stats for this cfqg */
 301
 302         /* async queue for each priority case */
 303         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 304         struct cfq_queue *async_idle_cfqq;
 305
 306 };
 307
 308 struct cfq_io_cq {
 309         struct io_cq            icq;            /* must be the first member */
 310         struct cfq_queue        *cfqq[2];
 311         struct cfq_ttime        ttime;
 312         int                     ioprio;         /* the current ioprio */
 313 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 314         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 315 #endif
 316 };
 317
 318 /*
 319  * Per block device queue structure
 320  */
 321 struct cfq_data {
 322         struct request_queue *queue;
 323         /* Root service tree for cfq_groups */
 324         struct cfq_rb_root grp_service_tree;
 325         struct cfq_group *root_group;
 326
 327         /*
 328          * The priority currently being served
 329          */
 330         enum wl_class_t serving_wl_class;
 331         enum wl_type_t serving_wl_type;
 332         unsigned long workload_expires;
 333         struct cfq_group *serving_group;
 334
 335         /*
 336          * Each priority tree is sorted by next_request position.  These
 337          * trees are used when determining if two or more queues are
 338          * interleaving requests (see cfq_close_cooperator).
 339          */
 340         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 341
 342         unsigned int busy_queues;
 343         unsigned int busy_sync_queues;
 344
 345         int rq_in_driver;
 346         int rq_in_flight[2];
 347
 348         /*
 349          * queue-depth detection
 350          */
 351         int rq_queued;
 352         int hw_tag;
 353         /*
 354          * hw_tag can be
 355          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 356          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 357          *  0 => no NCQ
 358          */
 359         int hw_tag_est_depth;
 360         unsigned int hw_tag_samples;
 361
 362         /*
 363          * idle window management
 364          */
 365         struct timer_list idle_slice_timer;
 366         struct work_struct unplug_work;
 367
 368         struct cfq_queue *active_queue;
 369         struct cfq_io_cq *active_cic;
 370
 371         sector_t last_position;
 372
 373         /*
 374          * tunables, see top of file
 375          */
 376         unsigned int cfq_quantum;
 377         unsigned int cfq_fifo_expire[2];
 378         unsigned int cfq_back_penalty;
 379         unsigned int cfq_back_max;
 380         unsigned int cfq_slice[2];
 381         unsigned int cfq_slice_async_rq;
 382         unsigned int cfq_slice_idle;
 383         unsigned int cfq_group_idle;
 384         unsigned int cfq_latency;
 385         unsigned int cfq_target_latency;
 386
 387         /*
 388          * Fallback dummy cfqq for extreme OOM conditions
 389          */
 390         struct cfq_queue oom_cfqq;
 391
 392         unsigned long last_delayed_sync;
 393 };
 394
 395 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 396 static void cfq_put_queue(struct cfq_queue *cfqq);
 397
 398 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 399                                             enum wl_class_t class,
 400                                             enum wl_type_t type)
 401 {
 402         if (!cfqg)
 403                 return NULL;
 404
 405         if (class == IDLE_WORKLOAD)
 406                 return &cfqg->service_tree_idle;
 407
 408         return &cfqg->service_trees[class][type];
 409 }
 410
 411 enum cfqq_state_flags {
 412         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 413         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 414         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 415         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 416         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 417         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 418         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 419         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 420         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 421         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 422         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 423         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 424         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 425 };
 426
 427 #define CFQ_CFQQ_FNS(name)                                              \
 428 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 429 {                                                                       \
 430         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 431 }                                                                       \
 432 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 433 {                                                                       \
 434         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 435 }                                                                       \
 436 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 437 {                                                                       \
 438         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 439 }
 440
 441 CFQ_CFQQ_FNS(on_rr);
 442 CFQ_CFQQ_FNS(wait_request);
 443 CFQ_CFQQ_FNS(must_dispatch);
 444 CFQ_CFQQ_FNS(must_alloc_slice);
 445 CFQ_CFQQ_FNS(fifo_expire);
 446 CFQ_CFQQ_FNS(idle_window);
 447 CFQ_CFQQ_FNS(prio_changed);
 448 CFQ_CFQQ_FNS(slice_new);
 449 CFQ_CFQQ_FNS(sync);
 450 CFQ_CFQQ_FNS(coop);
 451 CFQ_CFQQ_FNS(split_coop);
 452 CFQ_CFQQ_FNS(deep);
 453 CFQ_CFQQ_FNS(wait_busy);
 454 #undef CFQ_CFQQ_FNS
 455
 456 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 457
 458 /* cfqg stats flags */
 459 enum cfqg_stats_flags {
 460         CFQG_stats_waiting = 0,
 461         CFQG_stats_idling,
 462         CFQG_stats_empty,
 463 };
 464
 465 #define CFQG_FLAG_FNS(name)                                             \
 466 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 467 {                                                                       \
 468         stats->flags |= (1 << CFQG_stats_##name);                       \
 469 }                                                                       \
 470 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 471 {                                                                       \
 472         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 473 }                                                                       \
 474 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 475 {                                                                       \
 476         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 477 }                                                                       \
 478
 479 CFQG_FLAG_FNS(waiting)
 480 CFQG_FLAG_FNS(idling)
 481 CFQG_FLAG_FNS(empty)
 482 #undef CFQG_FLAG_FNS
 483
 484 /* This should be called with the queue_lock held. */
 485 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 486 {
 487         unsigned long long now;
 488
 489         if (!cfqg_stats_waiting(stats))
 490                 return;
 491
 492         now = sched_clock();
 493         if (time_after64(now, stats->start_group_wait_time))
 494                 blkg_stat_add(&stats->group_wait_time,
 495                               now - stats->start_group_wait_time);
 496         cfqg_stats_clear_waiting(stats);
 497 }
 498
 499 /* This should be called with the queue_lock held. */
 500 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 501                                                  struct cfq_group *curr_cfqg)
 502 {
 503         struct cfqg_stats *stats = &cfqg->stats;
 504
 505         if (cfqg_stats_waiting(stats))
 506                 return;
 507         if (cfqg == curr_cfqg)
 508                 return;
 509         stats->start_group_wait_time = sched_clock();
 510         cfqg_stats_mark_waiting(stats);
 511 }
 512
 513 /* This should be called with the queue_lock held. */
 514 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 515 {
 516         unsigned long long now;
 517
 518         if (!cfqg_stats_empty(stats))
 519                 return;
 520
 521         now = sched_clock();
 522         if (time_after64(now, stats->start_empty_time))
 523                 blkg_stat_add(&stats->empty_time,
 524                               now - stats->start_empty_time);
 525         cfqg_stats_clear_empty(stats);
 526 }
 527
 528 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 529 {
 530         blkg_stat_add(&cfqg->stats.dequeue, 1);
 531 }
 532
 533 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 534 {
 535         struct cfqg_stats *stats = &cfqg->stats;
 536
 537         if (blkg_rwstat_total(&stats->queued))
 538                 return;
 539
 540         /*
 541          * group is already marked empty. This can happen if cfqq got new
 542          * request in parent group and moved to this group while being added
 543          * to service tree. Just ignore the event and move on.
 544          */
 545         if (cfqg_stats_empty(stats))
 546                 return;
 547
 548         stats->start_empty_time = sched_clock();
 549         cfqg_stats_mark_empty(stats);
 550 }
 551
 552 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 553 {
 554         struct cfqg_stats *stats = &cfqg->stats;
 555
 556         if (cfqg_stats_idling(stats)) {
 557                 unsigned long long now = sched_clock();
 558
 559                 if (time_after64(now, stats->start_idle_time))
 560                         blkg_stat_add(&stats->idle_time,
 561                                       now - stats->start_idle_time);
 562                 cfqg_stats_clear_idling(stats);
 563         }
 564 }
 565
 566 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 567 {
 568         struct cfqg_stats *stats = &cfqg->stats;
 569
 570         BUG_ON(cfqg_stats_idling(stats));
 571
 572         stats->start_idle_time = sched_clock();
 573         cfqg_stats_mark_idling(stats);
 574 }
 575
 576 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 577 {
 578         struct cfqg_stats *stats = &cfqg->stats;
 579
 580         blkg_stat_add(&stats->avg_queue_size_sum,
 581                       blkg_rwstat_total(&stats->queued));
 582         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 583         cfqg_stats_update_group_wait_time(stats);
 584 }
 585
 586 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 587
 588 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 589 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 590 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 591 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 592 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 593 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 594 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 595
 596 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 597
 598 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 599
 600 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 601 {
 602         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 603 }
 604
 605 static struct cfq_group_data
 606 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 607 {
 608         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 609 }
 610
 611 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 612 {
 613         return pd_to_blkg(&cfqg->pd);
 614 }
 615
 616 static struct blkcg_policy blkcg_policy_cfq;
 617
 618 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 619 {
 620         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 621 }
 622
 623 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 624 {
 625         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 626 }
 627
 628 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 629 {
 630         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 631
 632         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 633 }
 634
 635 static inline void cfqg_get(struct cfq_group *cfqg)
 636 {
 637         return blkg_get(cfqg_to_blkg(cfqg));
 638 }
 639
 640 static inline void cfqg_put(struct cfq_group *cfqg)
 641 {
 642         return blkg_put(cfqg_to_blkg(cfqg));
 643 }
 644
 645 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 646         char __pbuf[128];                                               \
 647                                                                         \
 648         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 649         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 650                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 651                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 652                           __pbuf, ##args);                              \
 653 } while (0)
 654
 655 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 656         char __pbuf[128];                                               \
 657                                                                         \
 658         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 659         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 660 } while (0)
 661
 662 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 663                                             struct cfq_group *curr_cfqg, int rw)
 664 {
 665         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 666         cfqg_stats_end_empty_time(&cfqg->stats);
 667         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 668 }
 669
 670 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 671                         unsigned long time, unsigned long unaccounted_time)
 672 {
 673         blkg_stat_add(&cfqg->stats.time, time);
 674 #ifdef CONFIG_DEBUG_BLK_CGROUP
 675         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 676 #endif
 677 }
 678
 679 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 680 {
 681         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 682 }
 683
 684 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 685 {
 686         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 687 }
 688
 689 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 690                         uint64_t start_time, uint64_t io_start_time, int rw)
 691 {
 692         struct cfqg_stats *stats = &cfqg->stats;
 693         unsigned long long now = sched_clock();
 694
 695         if (time_after64(now, io_start_time))
 696                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 697         if (time_after64(io_start_time, start_time))
 698                 blkg_rwstat_add(&stats->wait_time, rw,
 699                                 io_start_time - start_time);
 700 }
 701
 702 /* @stats = 0 */
 703 static void cfqg_stats_reset(struct cfqg_stats *stats)
 704 {
 705         /* queued stats shouldn't be cleared */
 706         blkg_rwstat_reset(&stats->merged);
 707         blkg_rwstat_reset(&stats->service_time);
 708         blkg_rwstat_reset(&stats->wait_time);
 709         blkg_stat_reset(&stats->time);
 710 #ifdef CONFIG_DEBUG_BLK_CGROUP
 711         blkg_stat_reset(&stats->unaccounted_time);
 712         blkg_stat_reset(&stats->avg_queue_size_sum);
 713         blkg_stat_reset(&stats->avg_queue_size_samples);
 714         blkg_stat_reset(&stats->dequeue);
 715         blkg_stat_reset(&stats->group_wait_time);
 716         blkg_stat_reset(&stats->idle_time);
 717         blkg_stat_reset(&stats->empty_time);
 718 #endif
 719 }
 720
 721 /* @to += @from */
 722 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 723 {
 724         /* queued stats shouldn't be cleared */
 725         blkg_rwstat_add_aux(&to->merged, &from->merged);
 726         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 727         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 728         blkg_stat_add_aux(&from->time, &from->time);
 729 #ifdef CONFIG_DEBUG_BLK_CGROUP
 730         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 731         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 732         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 733         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 734         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 735         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 736         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 737 #endif
 738 }
 739
 740 /*
 741  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 742  * recursive stats can still account for the amount used by this cfqg after
 743  * it's gone.
 744  */
 745 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 746 {
 747         struct cfq_group *parent = cfqg_parent(cfqg);
 748
 749         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 750
 751         if (unlikely(!parent))
 752                 return;
 753
 754         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 755         cfqg_stats_reset(&cfqg->stats);
 756 }
 757
 758 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 759
 760 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 761 static inline void cfqg_get(struct cfq_group *cfqg) { }
 762 static inline void cfqg_put(struct cfq_group *cfqg) { }
 763
 764 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 765         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 766                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 767                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 768                                 ##args)
 769 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 770
 771 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 772                         struct cfq_group *curr_cfqg, int rw) { }
 773 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 774                         unsigned long time, unsigned long unaccounted_time) { }
 775 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 776 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 777 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 778                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 779
 780 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 781
 782 #define cfq_log(cfqd, fmt, args...)     \
 783         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 784
 785 /* Traverses through cfq group service trees */
 786 #define for_each_cfqg_st(cfqg, i, j, st) \
 787         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 788                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 789                         : &cfqg->service_tree_idle; \
 790                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 791                         (i == IDLE_WORKLOAD && j == 0); \
 792                         j++, st = i < IDLE_WORKLOAD ? \
 793                         &cfqg->service_trees[i][j]: NULL) \
 794
 795 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 796         struct cfq_ttime *ttime, bool group_idle)
 797 {
 798         unsigned long slice;
 799         if (!sample_valid(ttime->ttime_samples))
 800                 return false;
 801         if (group_idle)
 802                 slice = cfqd->cfq_group_idle;
 803         else
 804                 slice = cfqd->cfq_slice_idle;
 805         return ttime->ttime_mean > slice;
 806 }
 807
 808 static inline bool iops_mode(struct cfq_data *cfqd)
 809 {
 810         /*
 811          * If we are not idling on queues and it is a NCQ drive, parallel
 812          * execution of requests is on and measuring time is not possible
 813          * in most of the cases until and unless we drive shallower queue
 814          * depths and that becomes a performance bottleneck. In such cases
 815          * switch to start providing fairness in terms of number of IOs.
 816          */
 817         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 818                 return true;
 819         else
 820                 return false;
 821 }
 822
 823 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 824 {
 825         if (cfq_class_idle(cfqq))
 826                 return IDLE_WORKLOAD;
 827         if (cfq_class_rt(cfqq))
 828                 return RT_WORKLOAD;
 829         return BE_WORKLOAD;
 830 }
 831
 832
 833 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 834 {
 835         if (!cfq_cfqq_sync(cfqq))
 836                 return ASYNC_WORKLOAD;
 837         if (!cfq_cfqq_idle_window(cfqq))
 838                 return SYNC_NOIDLE_WORKLOAD;
 839         return SYNC_WORKLOAD;
 840 }
 841
 842 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 843                                         struct cfq_data *cfqd,
 844                                         struct cfq_group *cfqg)
 845 {
 846         if (wl_class == IDLE_WORKLOAD)
 847                 return cfqg->service_tree_idle.count;
 848
 849         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 850                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 851                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 852 }
 853
 854 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 855                                         struct cfq_group *cfqg)
 856 {
 857         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 858                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 859 }
 860
 861 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 862 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 863                                        struct cfq_io_cq *cic, struct bio *bio);
 864
 865 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 866 {
 867         /* cic->icq is the first member, %NULL will convert to %NULL */
 868         return container_of(icq, struct cfq_io_cq, icq);
 869 }
 870
 871 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 872                                                struct io_context *ioc)
 873 {
 874         if (ioc)
 875                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 876         return NULL;
 877 }
 878
 879 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 880 {
 881         return cic->cfqq[is_sync];
 882 }
 883
 884 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 885                                 bool is_sync)
 886 {
 887         cic->cfqq[is_sync] = cfqq;
 888 }
 889
 890 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 891 {
 892         return cic->icq.q->elevator->elevator_data;
 893 }
 894
 895 /*
 896  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 897  * set (in which case it could also be direct WRITE).
 898  */
 899 static inline bool cfq_bio_sync(struct bio *bio)
 900 {
 901         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 902 }
 903
 904 /*
 905  * scheduler run of queue, if there are requests pending and no one in the
 906  * driver that will restart queueing
 907  */
 908 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 909 {
 910         if (cfqd->busy_queues) {
 911                 cfq_log(cfqd, "schedule dispatch");
 912                 kblockd_schedule_work(&cfqd->unplug_work);
 913         }
 914 }
 915
 916 /*
 917  * Scale schedule slice based on io priority. Use the sync time slice only
 918  * if a queue is marked sync and has sync io queued. A sync queue with async
 919  * io only, should not get full sync slice length.
 920  */
 921 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 922                                  unsigned short prio)
 923 {
 924         const int base_slice = cfqd->cfq_slice[sync];
 925
 926         WARN_ON(prio >= IOPRIO_BE_NR);
 927
 928         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 929 }
 930
 931 static inline int
 932 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 933 {
 934         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 935 }
 936
 937 /**
 938  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 939  * @charge: disk time being charged
 940  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 941  *
 942  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 943  * scaling is inversely proportional.
 944  *
 945  * scaled = charge / vfraction
 946  *
 947  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 948  */
 949 static inline u64 cfqg_scale_charge(unsigned long charge,
 950                                     unsigned int vfraction)
 951 {
 952         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 953
 954         /* charge / vfraction */
 955         c <<= CFQ_SERVICE_SHIFT;
 956         do_div(c, vfraction);
 957         return c;
 958 }
 959
 960 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 961 {
 962         s64 delta = (s64)(vdisktime - min_vdisktime);
 963         if (delta > 0)
 964                 min_vdisktime = vdisktime;
 965
 966         return min_vdisktime;
 967 }
 968
 969 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 970 {
 971         s64 delta = (s64)(vdisktime - min_vdisktime);
 972         if (delta < 0)
 973                 min_vdisktime = vdisktime;
 974
 975         return min_vdisktime;
 976 }
 977
 978 static void update_min_vdisktime(struct cfq_rb_root *st)
 979 {
 980         struct cfq_group *cfqg;
 981
 982         if (st->left) {
 983                 cfqg = rb_entry_cfqg(st->left);
 984                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 985                                                   cfqg->vdisktime);
 986         }
 987 }
 988
 989 /*
 990  * get averaged number of queues of RT/BE priority.
 991  * average is updated, with a formula that gives more weight to higher numbers,
 992  * to quickly follows sudden increases and decrease slowly
 993  */
 994
 995 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
 996                                         struct cfq_group *cfqg, bool rt)
 997 {
 998         unsigned min_q, max_q;
 999         unsigned mult  = cfq_hist_divisor - 1;
1000         unsigned round = cfq_hist_divisor / 2;
1001         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1002
1003         min_q = min(cfqg->busy_queues_avg[rt], busy);
1004         max_q = max(cfqg->busy_queues_avg[rt], busy);
1005         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1006                 cfq_hist_divisor;
1007         return cfqg->busy_queues_avg[rt];
1008 }
1009
1010 static inline unsigned
1011 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1012 {
1013         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1014 }
1015
1016 static inline unsigned
1017 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1018 {
1019         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1020         if (cfqd->cfq_latency) {
1021                 /*
1022                  * interested queues (we consider only the ones with the same
1023                  * priority class in the cfq group)
1024                  */
1025                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1026                                                 cfq_class_rt(cfqq));
1027                 unsigned sync_slice = cfqd->cfq_slice[1];
1028                 unsigned expect_latency = sync_slice * iq;
1029                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1030
1031                 if (expect_latency > group_slice) {
1032                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1033                         /* scale low_slice according to IO priority
1034                          * and sync vs async */
1035                         unsigned low_slice =
1036                                 min(slice, base_low_slice * slice / sync_slice);
1037                         /* the adapted slice value is scaled to fit all iqs
1038                          * into the target latency */
1039                         slice = max(slice * group_slice / expect_latency,
1040                                     low_slice);
1041                 }
1042         }
1043         return slice;
1044 }
1045
1046 static inline void
1047 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1048 {
1049         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1050
1051         cfqq->slice_start = jiffies;
1052         cfqq->slice_end = jiffies + slice;
1053         cfqq->allocated_slice = slice;
1054         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1055 }
1056
1057 /*
1058  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1059  * isn't valid until the first request from the dispatch is activated
1060  * and the slice time set.
1061  */
1062 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1063 {
1064         if (cfq_cfqq_slice_new(cfqq))
1065                 return false;
1066         if (time_before(jiffies, cfqq->slice_end))
1067                 return false;
1068
1069         return true;
1070 }
1071
1072 /*
1073  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1074  * We choose the request that is closest to the head right now. Distance
1075  * behind the head is penalized and only allowed to a certain extent.
1076  */
1077 static struct request *
1078 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1079 {
1080         sector_t s1, s2, d1 = 0, d2 = 0;
1081         unsigned long back_max;
1082 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1083 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1084         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1085
1086         if (rq1 == NULL || rq1 == rq2)
1087                 return rq2;
1088         if (rq2 == NULL)
1089                 return rq1;
1090
1091         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1092                 return rq_is_sync(rq1) ? rq1 : rq2;
1093
1094         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1095                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1096
1097         s1 = blk_rq_pos(rq1);
1098         s2 = blk_rq_pos(rq2);
1099
1100         /*
1101          * by definition, 1KiB is 2 sectors
1102          */
1103         back_max = cfqd->cfq_back_max * 2;
1104
1105         /*
1106          * Strict one way elevator _except_ in the case where we allow
1107          * short backward seeks which are biased as twice the cost of a
1108          * similar forward seek.
1109          */
1110         if (s1 >= last)
1111                 d1 = s1 - last;
1112         else if (s1 + back_max >= last)
1113                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1114         else
1115                 wrap |= CFQ_RQ1_WRAP;
1116
1117         if (s2 >= last)
1118                 d2 = s2 - last;
1119         else if (s2 + back_max >= last)
1120                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1121         else
1122                 wrap |= CFQ_RQ2_WRAP;
1123
1124         /* Found required data */
1125
1126         /*
1127          * By doing switch() on the bit mask "wrap" we avoid having to
1128          * check two variables for all permutations: --> faster!
1129          */
1130         switch (wrap) {
1131         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1132                 if (d1 < d2)
1133                         return rq1;
1134                 else if (d2 < d1)
1135                         return rq2;
1136                 else {
1137                         if (s1 >= s2)
1138                                 return rq1;
1139                         else
1140                                 return rq2;
1141                 }
1142
1143         case CFQ_RQ2_WRAP:
1144                 return rq1;
1145         case CFQ_RQ1_WRAP:
1146                 return rq2;
1147         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1148         default:
1149                 /*
1150                  * Since both rqs are wrapped,
1151                  * start with the one that's further behind head
1152                  * (--> only *one* back seek required),
1153                  * since back seek takes more time than forward.
1154                  */
1155                 if (s1 <= s2)
1156                         return rq1;
1157                 else
1158                         return rq2;
1159         }
1160 }
1161
1162 /*
1163  * The below is leftmost cache rbtree addon
1164  */
1165 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1166 {
1167         /* Service tree is empty */
1168         if (!root->count)
1169                 return NULL;
1170
1171         if (!root->left)
1172                 root->left = rb_first(&root->rb);
1173
1174         if (root->left)
1175                 return rb_entry(root->left, struct cfq_queue, rb_node);
1176
1177         return NULL;
1178 }
1179
1180 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1181 {
1182         if (!root->left)
1183                 root->left = rb_first(&root->rb);
1184
1185         if (root->left)
1186                 return rb_entry_cfqg(root->left);
1187
1188         return NULL;
1189 }
1190
1191 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1192 {
1193         rb_erase(n, root);
1194         RB_CLEAR_NODE(n);
1195 }
1196
1197 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1198 {
1199         if (root->left == n)
1200                 root->left = NULL;
1201         rb_erase_init(n, &root->rb);
1202         --root->count;
1203 }
1204
1205 /*
1206  * would be nice to take fifo expire time into account as well
1207  */
1208 static struct request *
1209 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1210                   struct request *last)
1211 {
1212         struct rb_node *rbnext = rb_next(&last->rb_node);
1213         struct rb_node *rbprev = rb_prev(&last->rb_node);
1214         struct request *next = NULL, *prev = NULL;
1215
1216         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1217
1218         if (rbprev)
1219                 prev = rb_entry_rq(rbprev);
1220
1221         if (rbnext)
1222                 next = rb_entry_rq(rbnext);
1223         else {
1224                 rbnext = rb_first(&cfqq->sort_list);
1225                 if (rbnext && rbnext != &last->rb_node)
1226                         next = rb_entry_rq(rbnext);
1227         }
1228
1229         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1230 }
1231
1232 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1233                                       struct cfq_queue *cfqq)
1234 {
1235         /*
1236          * just an approximation, should be ok.
1237          */
1238         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1239                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1240 }
1241
1242 static inline s64
1243 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1244 {
1245         return cfqg->vdisktime - st->min_vdisktime;
1246 }
1247
1248 static void
1249 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1250 {
1251         struct rb_node **node = &st->rb.rb_node;
1252         struct rb_node *parent = NULL;
1253         struct cfq_group *__cfqg;
1254         s64 key = cfqg_key(st, cfqg);
1255         int left = 1;
1256
1257         while (*node != NULL) {
1258                 parent = *node;
1259                 __cfqg = rb_entry_cfqg(parent);
1260
1261                 if (key < cfqg_key(st, __cfqg))
1262                         node = &parent->rb_left;
1263                 else {
1264                         node = &parent->rb_right;
1265                         left = 0;
1266                 }
1267         }
1268
1269         if (left)
1270                 st->left = &cfqg->rb_node;
1271
1272         rb_link_node(&cfqg->rb_node, parent, node);
1273         rb_insert_color(&cfqg->rb_node, &st->rb);
1274 }
1275
1276 /*
1277  * This has to be called only on activation of cfqg
1278  */
1279 static void
1280 cfq_update_group_weight(struct cfq_group *cfqg)
1281 {
1282         if (cfqg->new_weight) {
1283                 cfqg->weight = cfqg->new_weight;
1284                 cfqg->new_weight = 0;
1285         }
1286 }
1287
1288 static void
1289 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1290 {
1291         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1292
1293         if (cfqg->new_leaf_weight) {
1294                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1295                 cfqg->new_leaf_weight = 0;
1296         }
1297 }
1298
1299 static void
1300 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1301 {
1302         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1303         struct cfq_group *pos = cfqg;
1304         struct cfq_group *parent;
1305         bool propagate;
1306
1307         /* add to the service tree */
1308         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1309
1310         /*
1311          * Update leaf_weight.  We cannot update weight at this point
1312          * because cfqg might already have been activated and is
1313          * contributing its current weight to the parent's child_weight.
1314          */
1315         cfq_update_group_leaf_weight(cfqg);
1316         __cfq_group_service_tree_add(st, cfqg);
1317
1318         /*
1319          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1320          * entitled to.  vfraction is calculated by walking the tree
1321          * towards the root calculating the fraction it has at each level.
1322          * The compounded ratio is how much vfraction @cfqg owns.
1323          *
1324          * Start with the proportion tasks in this cfqg has against active
1325          * children cfqgs - its leaf_weight against children_weight.
1326          */
1327         propagate = !pos->nr_active++;
1328         pos->children_weight += pos->leaf_weight;
1329         vfr = vfr * pos->leaf_weight / pos->children_weight;
1330
1331         /*
1332          * Compound ->weight walking up the tree.  Both activation and
1333          * vfraction calculation are done in the same loop.  Propagation
1334          * stops once an already activated node is met.  vfraction
1335          * calculation should always continue to the root.
1336          */
1337         while ((parent = cfqg_parent(pos))) {
1338                 if (propagate) {
1339                         cfq_update_group_weight(pos);
1340                         propagate = !parent->nr_active++;
1341                         parent->children_weight += pos->weight;
1342                 }
1343                 vfr = vfr * pos->weight / parent->children_weight;
1344                 pos = parent;
1345         }
1346
1347         cfqg->vfraction = max_t(unsigned, vfr, 1);
1348 }
1349
1350 static void
1351 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1352 {
1353         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1354         struct cfq_group *__cfqg;
1355         struct rb_node *n;
1356
1357         cfqg->nr_cfqq++;
1358         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1359                 return;
1360
1361         /*
1362          * Currently put the group at the end. Later implement something
1363          * so that groups get lesser vtime based on their weights, so that
1364          * if group does not loose all if it was not continuously backlogged.
1365          */
1366         n = rb_last(&st->rb);
1367         if (n) {
1368                 __cfqg = rb_entry_cfqg(n);
1369                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1370         } else
1371                 cfqg->vdisktime = st->min_vdisktime;
1372         cfq_group_service_tree_add(st, cfqg);
1373 }
1374
1375 static void
1376 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1377 {
1378         struct cfq_group *pos = cfqg;
1379         bool propagate;
1380
1381         /*
1382          * Undo activation from cfq_group_service_tree_add().  Deactivate
1383          * @cfqg and propagate deactivation upwards.
1384          */
1385         propagate = !--pos->nr_active;
1386         pos->children_weight -= pos->leaf_weight;
1387
1388         while (propagate) {
1389                 struct cfq_group *parent = cfqg_parent(pos);
1390
1391                 /* @pos has 0 nr_active at this point */
1392                 WARN_ON_ONCE(pos->children_weight);
1393                 pos->vfraction = 0;
1394
1395                 if (!parent)
1396                         break;
1397
1398                 propagate = !--parent->nr_active;
1399                 parent->children_weight -= pos->weight;
1400                 pos = parent;
1401         }
1402
1403         /* remove from the service tree */
1404         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1405                 cfq_rb_erase(&cfqg->rb_node, st);
1406 }
1407
1408 static void
1409 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1410 {
1411         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1412
1413         BUG_ON(cfqg->nr_cfqq < 1);
1414         cfqg->nr_cfqq--;
1415
1416         /* If there are other cfq queues under this group, don't delete it */
1417         if (cfqg->nr_cfqq)
1418                 return;
1419
1420         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1421         cfq_group_service_tree_del(st, cfqg);
1422         cfqg->saved_wl_slice = 0;
1423         cfqg_stats_update_dequeue(cfqg);
1424 }
1425
1426 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1427                                                 unsigned int *unaccounted_time)
1428 {
1429         unsigned int slice_used;
1430
1431         /*
1432          * Queue got expired before even a single request completed or
1433          * got expired immediately after first request completion.
1434          */
1435         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1436                 /*
1437                  * Also charge the seek time incurred to the group, otherwise
1438                  * if there are mutiple queues in the group, each can dispatch
1439                  * a single request on seeky media and cause lots of seek time
1440                  * and group will never know it.
1441                  */
1442                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1443                                         1);
1444         } else {
1445                 slice_used = jiffies - cfqq->slice_start;
1446                 if (slice_used > cfqq->allocated_slice) {
1447                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1448                         slice_used = cfqq->allocated_slice;
1449                 }
1450                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1451                         *unaccounted_time += cfqq->slice_start -
1452                                         cfqq->dispatch_start;
1453         }
1454
1455         return slice_used;
1456 }
1457
1458 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1459                                 struct cfq_queue *cfqq)
1460 {
1461         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1462         unsigned int used_sl, charge, unaccounted_sl = 0;
1463         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1464                         - cfqg->service_tree_idle.count;
1465         unsigned int vfr;
1466
1467         BUG_ON(nr_sync < 0);
1468         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1469
1470         if (iops_mode(cfqd))
1471                 charge = cfqq->slice_dispatch;
1472         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1473                 charge = cfqq->allocated_slice;
1474
1475         /*
1476          * Can't update vdisktime while on service tree and cfqg->vfraction
1477          * is valid only while on it.  Cache vfr, leave the service tree,
1478          * update vdisktime and go back on.  The re-addition to the tree
1479          * will also update the weights as necessary.
1480          */
1481         vfr = cfqg->vfraction;
1482         cfq_group_service_tree_del(st, cfqg);
1483         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1484         cfq_group_service_tree_add(st, cfqg);
1485
1486         /* This group is being expired. Save the context */
1487         if (time_after(cfqd->workload_expires, jiffies)) {
1488                 cfqg->saved_wl_slice = cfqd->workload_expires
1489                                                 - jiffies;
1490                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1491                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1492         } else
1493                 cfqg->saved_wl_slice = 0;
1494
1495         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1496                                         st->min_vdisktime);
1497         cfq_log_cfqq(cfqq->cfqd, cfqq,
1498                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1499                      used_sl, cfqq->slice_dispatch, charge,
1500                      iops_mode(cfqd), cfqq->nr_sectors);
1501         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1502         cfqg_stats_set_start_empty_time(cfqg);
1503 }
1504
1505 /**
1506  * cfq_init_cfqg_base - initialize base part of a cfq_group
1507  * @cfqg: cfq_group to initialize
1508  *
1509  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1510  * is enabled or not.
1511  */
1512 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1513 {
1514         struct cfq_rb_root *st;
1515         int i, j;
1516
1517         for_each_cfqg_st(cfqg, i, j, st)
1518                 *st = CFQ_RB_ROOT;
1519         RB_CLEAR_NODE(&cfqg->rb_node);
1520
1521         cfqg->ttime.last_end_request = jiffies;
1522 }
1523
1524 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1525 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1526                             bool on_dfl, bool reset_dev, bool is_leaf_weight);
1527
1528 static void cfqg_stats_exit(struct cfqg_stats *stats)
1529 {
1530         blkg_rwstat_exit(&stats->merged);
1531         blkg_rwstat_exit(&stats->service_time);
1532         blkg_rwstat_exit(&stats->wait_time);
1533         blkg_rwstat_exit(&stats->queued);
1534         blkg_stat_exit(&stats->time);
1535 #ifdef CONFIG_DEBUG_BLK_CGROUP
1536         blkg_stat_exit(&stats->unaccounted_time);
1537         blkg_stat_exit(&stats->avg_queue_size_sum);
1538         blkg_stat_exit(&stats->avg_queue_size_samples);
1539         blkg_stat_exit(&stats->dequeue);
1540         blkg_stat_exit(&stats->group_wait_time);
1541         blkg_stat_exit(&stats->idle_time);
1542         blkg_stat_exit(&stats->empty_time);
1543 #endif
1544 }
1545
1546 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1547 {
1548         if (blkg_rwstat_init(&stats->merged, gfp) ||
1549             blkg_rwstat_init(&stats->service_time, gfp) ||
1550             blkg_rwstat_init(&stats->wait_time, gfp) ||
1551             blkg_rwstat_init(&stats->queued, gfp) ||
1552             blkg_stat_init(&stats->time, gfp))
1553                 goto err;
1554
1555 #ifdef CONFIG_DEBUG_BLK_CGROUP
1556         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1557             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1558             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1559             blkg_stat_init(&stats->dequeue, gfp) ||
1560             blkg_stat_init(&stats->group_wait_time, gfp) ||
1561             blkg_stat_init(&stats->idle_time, gfp) ||
1562             blkg_stat_init(&stats->empty_time, gfp))
1563                 goto err;
1564 #endif
1565         return 0;
1566 err:
1567         cfqg_stats_exit(stats);
1568         return -ENOMEM;
1569 }
1570
1571 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1572 {
1573         struct cfq_group_data *cgd;
1574
1575         cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1576         if (!cgd)
1577                 return NULL;
1578         return &cgd->cpd;
1579 }
1580
1581 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1582 {
1583         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1584         unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1585                               CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1586
1587         if (cpd_to_blkcg(cpd) == &blkcg_root)
1588                 weight *= 2;
1589
1590         cgd->weight = weight;
1591         cgd->leaf_weight = weight;
1592 }
1593
1594 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1595 {
1596         kfree(cpd_to_cfqgd(cpd));
1597 }
1598
1599 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1600 {
1601         struct blkcg *blkcg = cpd_to_blkcg(cpd);
1602         bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1603         unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1604
1605         if (blkcg == &blkcg_root)
1606                 weight *= 2;
1607
1608         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1609         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1610 }
1611
1612 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1613 {
1614         struct cfq_group *cfqg;
1615
1616         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1617         if (!cfqg)
1618                 return NULL;
1619
1620         cfq_init_cfqg_base(cfqg);
1621         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1622                 kfree(cfqg);
1623                 return NULL;
1624         }
1625
1626         return &cfqg->pd;
1627 }
1628
1629 static void cfq_pd_init(struct blkg_policy_data *pd)
1630 {
1631         struct cfq_group *cfqg = pd_to_cfqg(pd);
1632         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1633
1634         cfqg->weight = cgd->weight;
1635         cfqg->leaf_weight = cgd->leaf_weight;
1636 }
1637
1638 static void cfq_pd_offline(struct blkg_policy_data *pd)
1639 {
1640         struct cfq_group *cfqg = pd_to_cfqg(pd);
1641         int i;
1642
1643         for (i = 0; i < IOPRIO_BE_NR; i++) {
1644                 if (cfqg->async_cfqq[0][i])
1645                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1646                 if (cfqg->async_cfqq[1][i])
1647                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1648         }
1649
1650         if (cfqg->async_idle_cfqq)
1651                 cfq_put_queue(cfqg->async_idle_cfqq);
1652
1653         /*
1654          * @blkg is going offline and will be ignored by
1655          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1656          * that they don't get lost.  If IOs complete after this point, the
1657          * stats for them will be lost.  Oh well...
1658          */
1659         cfqg_stats_xfer_dead(cfqg);
1660 }
1661
1662 static void cfq_pd_free(struct blkg_policy_data *pd)
1663 {
1664         struct cfq_group *cfqg = pd_to_cfqg(pd);
1665
1666         cfqg_stats_exit(&cfqg->stats);
1667         return kfree(cfqg);
1668 }
1669
1670 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1671 {
1672         struct cfq_group *cfqg = pd_to_cfqg(pd);
1673
1674         cfqg_stats_reset(&cfqg->stats);
1675 }
1676
1677 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1678                                          struct blkcg *blkcg)
1679 {
1680         struct blkcg_gq *blkg;
1681
1682         blkg = blkg_lookup(blkcg, cfqd->queue);
1683         if (likely(blkg))
1684                 return blkg_to_cfqg(blkg);
1685         return NULL;
1686 }
1687
1688 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1689 {
1690         cfqq->cfqg = cfqg;
1691         /* cfqq reference on cfqg */
1692         cfqg_get(cfqg);
1693 }
1694
1695 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1696                                      struct blkg_policy_data *pd, int off)
1697 {
1698         struct cfq_group *cfqg = pd_to_cfqg(pd);
1699
1700         if (!cfqg->dev_weight)
1701                 return 0;
1702         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1703 }
1704
1705 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1706 {
1707         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1708                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1709                           0, false);
1710         return 0;
1711 }
1712
1713 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1714                                           struct blkg_policy_data *pd, int off)
1715 {
1716         struct cfq_group *cfqg = pd_to_cfqg(pd);
1717
1718         if (!cfqg->dev_leaf_weight)
1719                 return 0;
1720         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1721 }
1722
1723 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1724 {
1725         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1726                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1727                           0, false);
1728         return 0;
1729 }
1730
1731 static int cfq_print_weight(struct seq_file *sf, void *v)
1732 {
1733         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1734         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1735         unsigned int val = 0;
1736
1737         if (cgd)
1738                 val = cgd->weight;
1739
1740         seq_printf(sf, "%u\n", val);
1741         return 0;
1742 }
1743
1744 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1745 {
1746         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1747         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1748         unsigned int val = 0;
1749
1750         if (cgd)
1751                 val = cgd->leaf_weight;
1752
1753         seq_printf(sf, "%u\n", val);
1754         return 0;
1755 }
1756
1757 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1758                                         char *buf, size_t nbytes, loff_t off,
1759                                         bool on_dfl, bool is_leaf_weight)
1760 {
1761         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1762         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1763         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1764         struct blkg_conf_ctx ctx;
1765         struct cfq_group *cfqg;
1766         struct cfq_group_data *cfqgd;
1767         int ret;
1768         u64 v;
1769
1770         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1771         if (ret)
1772                 return ret;
1773
1774         if (sscanf(ctx.body, "%llu", &v) == 1) {
1775                 /* require "default" on dfl */
1776                 ret = -ERANGE;
1777                 if (!v && on_dfl)
1778                         goto out_finish;
1779         } else if (!strcmp(strim(ctx.body), "default")) {
1780                 v = 0;
1781         } else {
1782                 ret = -EINVAL;
1783                 goto out_finish;
1784         }
1785
1786         cfqg = blkg_to_cfqg(ctx.blkg);
1787         cfqgd = blkcg_to_cfqgd(blkcg);
1788
1789         ret = -ERANGE;
1790         if (!v || (v >= min && v <= max)) {
1791                 if (!is_leaf_weight) {
1792                         cfqg->dev_weight = v;
1793                         cfqg->new_weight = v ?: cfqgd->weight;
1794                 } else {
1795                         cfqg->dev_leaf_weight = v;
1796                         cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1797                 }
1798                 ret = 0;
1799         }
1800 out_finish:
1801         blkg_conf_finish(&ctx);
1802         return ret ?: nbytes;
1803 }
1804
1805 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1806                                       char *buf, size_t nbytes, loff_t off)
1807 {
1808         return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1809 }
1810
1811 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1812                                            char *buf, size_t nbytes, loff_t off)
1813 {
1814         return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1815 }
1816
1817 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1818                             bool on_dfl, bool reset_dev, bool is_leaf_weight)
1819 {
1820         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1821         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1822         struct blkcg *blkcg = css_to_blkcg(css);
1823         struct blkcg_gq *blkg;
1824         struct cfq_group_data *cfqgd;
1825         int ret = 0;
1826
1827         if (val < min || val > max)
1828                 return -ERANGE;
1829
1830         spin_lock_irq(&blkcg->lock);
1831         cfqgd = blkcg_to_cfqgd(blkcg);
1832         if (!cfqgd) {
1833                 ret = -EINVAL;
1834                 goto out;
1835         }
1836
1837         if (!is_leaf_weight)
1838                 cfqgd->weight = val;
1839         else
1840                 cfqgd->leaf_weight = val;
1841
1842         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1843                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1844
1845                 if (!cfqg)
1846                         continue;
1847
1848                 if (!is_leaf_weight) {
1849                         if (reset_dev)
1850                                 cfqg->dev_weight = 0;
1851                         if (!cfqg->dev_weight)
1852                                 cfqg->new_weight = cfqgd->weight;
1853                 } else {
1854                         if (reset_dev)
1855                                 cfqg->dev_leaf_weight = 0;
1856                         if (!cfqg->dev_leaf_weight)
1857                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1858                 }
1859         }
1860
1861 out:
1862         spin_unlock_irq(&blkcg->lock);
1863         return ret;
1864 }
1865
1866 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1867                           u64 val)
1868 {
1869         return __cfq_set_weight(css, val, false, false, false);
1870 }
1871
1872 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1873                                struct cftype *cft, u64 val)
1874 {
1875         return __cfq_set_weight(css, val, false, false, true);
1876 }
1877
1878 static int cfqg_print_stat(struct seq_file *sf, void *v)
1879 {
1880         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1881                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1882         return 0;
1883 }
1884
1885 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1886 {
1887         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1888                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1889         return 0;
1890 }
1891
1892 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1893                                       struct blkg_policy_data *pd, int off)
1894 {
1895         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1896                                           &blkcg_policy_cfq, off);
1897         return __blkg_prfill_u64(sf, pd, sum);
1898 }
1899
1900 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1901                                         struct blkg_policy_data *pd, int off)
1902 {
1903         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1904                                                         &blkcg_policy_cfq, off);
1905         return __blkg_prfill_rwstat(sf, pd, &sum);
1906 }
1907
1908 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1909 {
1910         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1911                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1912                           seq_cft(sf)->private, false);
1913         return 0;
1914 }
1915
1916 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1917 {
1918         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1919                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1920                           seq_cft(sf)->private, true);
1921         return 0;
1922 }
1923
1924 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1925                                int off)
1926 {
1927         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1928
1929         return __blkg_prfill_u64(sf, pd, sum >> 9);
1930 }
1931
1932 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1933 {
1934         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1935                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1936         return 0;
1937 }
1938
1939 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1940                                          struct blkg_policy_data *pd, int off)
1941 {
1942         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1943                                         offsetof(struct blkcg_gq, stat_bytes));
1944         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1945                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1946
1947         return __blkg_prfill_u64(sf, pd, sum >> 9);
1948 }
1949
1950 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1951 {
1952         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1953                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1954                           false);
1955         return 0;
1956 }
1957
1958 #ifdef CONFIG_DEBUG_BLK_CGROUP
1959 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1960                                       struct blkg_policy_data *pd, int off)
1961 {
1962         struct cfq_group *cfqg = pd_to_cfqg(pd);
1963         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1964         u64 v = 0;
1965
1966         if (samples) {
1967                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1968                 v = div64_u64(v, samples);
1969         }
1970         __blkg_prfill_u64(sf, pd, v);
1971         return 0;
1972 }
1973
1974 /* print avg_queue_size */
1975 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1976 {
1977         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1978                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1979                           0, false);
1980         return 0;
1981 }
1982 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1983
1984 static struct cftype cfq_blkcg_legacy_files[] = {
1985         /* on root, weight is mapped to leaf_weight */
1986         {
1987                 .name = "weight_device",
1988                 .flags = CFTYPE_ONLY_ON_ROOT,
1989                 .seq_show = cfqg_print_leaf_weight_device,
1990                 .write = cfqg_set_leaf_weight_device,
1991         },
1992         {
1993                 .name = "weight",
1994                 .flags = CFTYPE_ONLY_ON_ROOT,
1995                 .seq_show = cfq_print_leaf_weight,
1996                 .write_u64 = cfq_set_leaf_weight,
1997         },
1998
1999         /* no such mapping necessary for !roots */
2000         {
2001                 .name = "weight_device",
2002                 .flags = CFTYPE_NOT_ON_ROOT,
2003                 .seq_show = cfqg_print_weight_device,
2004                 .write = cfqg_set_weight_device,
2005         },
2006         {
2007                 .name = "weight",
2008                 .flags = CFTYPE_NOT_ON_ROOT,
2009                 .seq_show = cfq_print_weight,
2010                 .write_u64 = cfq_set_weight,
2011         },
2012
2013         {
2014                 .name = "leaf_weight_device",
2015                 .seq_show = cfqg_print_leaf_weight_device,
2016                 .write = cfqg_set_leaf_weight_device,
2017         },
2018         {
2019                 .name = "leaf_weight",
2020                 .seq_show = cfq_print_leaf_weight,
2021                 .write_u64 = cfq_set_leaf_weight,
2022         },
2023
2024         /* statistics, covers only the tasks in the cfqg */
2025         {
2026                 .name = "time",
2027                 .private = offsetof(struct cfq_group, stats.time),
2028                 .seq_show = cfqg_print_stat,
2029         },
2030         {
2031                 .name = "sectors",
2032                 .seq_show = cfqg_print_stat_sectors,
2033         },
2034         {
2035                 .name = "io_service_bytes",
2036                 .private = (unsigned long)&blkcg_policy_cfq,
2037                 .seq_show = blkg_print_stat_bytes,
2038         },
2039         {
2040                 .name = "io_serviced",
2041                 .private = (unsigned long)&blkcg_policy_cfq,
2042                 .seq_show = blkg_print_stat_ios,
2043         },
2044         {
2045                 .name = "io_service_time",
2046                 .private = offsetof(struct cfq_group, stats.service_time),
2047                 .seq_show = cfqg_print_rwstat,
2048         },
2049         {
2050                 .name = "io_wait_time",
2051                 .private = offsetof(struct cfq_group, stats.wait_time),
2052                 .seq_show = cfqg_print_rwstat,
2053         },
2054         {
2055                 .name = "io_merged",
2056                 .private = offsetof(struct cfq_group, stats.merged),
2057                 .seq_show = cfqg_print_rwstat,
2058         },
2059         {
2060                 .name = "io_queued",
2061                 .private = offsetof(struct cfq_group, stats.queued),
2062                 .seq_show = cfqg_print_rwstat,
2063         },
2064
2065         /* the same statictics which cover the cfqg and its descendants */
2066         {
2067                 .name = "time_recursive",
2068                 .private = offsetof(struct cfq_group, stats.time),
2069                 .seq_show = cfqg_print_stat_recursive,
2070         },
2071         {
2072                 .name = "sectors_recursive",
2073                 .seq_show = cfqg_print_stat_sectors_recursive,
2074         },
2075         {
2076                 .name = "io_service_bytes_recursive",
2077                 .private = (unsigned long)&blkcg_policy_cfq,
2078                 .seq_show = blkg_print_stat_bytes_recursive,
2079         },
2080         {
2081                 .name = "io_serviced_recursive",
2082                 .private = (unsigned long)&blkcg_policy_cfq,
2083                 .seq_show = blkg_print_stat_ios_recursive,
2084         },
2085         {
2086                 .name = "io_service_time_recursive",
2087                 .private = offsetof(struct cfq_group, stats.service_time),
2088                 .seq_show = cfqg_print_rwstat_recursive,
2089         },
2090         {
2091                 .name = "io_wait_time_recursive",
2092                 .private = offsetof(struct cfq_group, stats.wait_time),
2093                 .seq_show = cfqg_print_rwstat_recursive,
2094         },
2095         {
2096                 .name = "io_merged_recursive",
2097                 .private = offsetof(struct cfq_group, stats.merged),
2098                 .seq_show = cfqg_print_rwstat_recursive,
2099         },
2100         {
2101                 .name = "io_queued_recursive",
2102                 .private = offsetof(struct cfq_group, stats.queued),
2103                 .seq_show = cfqg_print_rwstat_recursive,
2104         },
2105 #ifdef CONFIG_DEBUG_BLK_CGROUP
2106         {
2107                 .name = "avg_queue_size",
2108                 .seq_show = cfqg_print_avg_queue_size,
2109         },
2110         {
2111                 .name = "group_wait_time",
2112                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2113                 .seq_show = cfqg_print_stat,
2114         },
2115         {
2116                 .name = "idle_time",
2117                 .private = offsetof(struct cfq_group, stats.idle_time),
2118                 .seq_show = cfqg_print_stat,
2119         },
2120         {
2121                 .name = "empty_time",
2122                 .private = offsetof(struct cfq_group, stats.empty_time),
2123                 .seq_show = cfqg_print_stat,
2124         },
2125         {
2126                 .name = "dequeue",
2127                 .private = offsetof(struct cfq_group, stats.dequeue),
2128                 .seq_show = cfqg_print_stat,
2129         },
2130         {
2131                 .name = "unaccounted_time",
2132                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2133                 .seq_show = cfqg_print_stat,
2134         },
2135 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2136         { }     /* terminate */
2137 };
2138
2139 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2140 {
2141         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2142         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2143
2144         seq_printf(sf, "default %u\n", cgd->weight);
2145         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2146                           &blkcg_policy_cfq, 0, false);
2147         return 0;
2148 }
2149
2150 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2151                                      char *buf, size_t nbytes, loff_t off)
2152 {
2153         char *endp;
2154         int ret;
2155         u64 v;
2156
2157         buf = strim(buf);
2158
2159         /* "WEIGHT" or "default WEIGHT" sets the default weight */
2160         v = simple_strtoull(buf, &endp, 0);
2161         if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2162                 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2163                 return ret ?: nbytes;
2164         }
2165
2166         /* "MAJ:MIN WEIGHT" */
2167         return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2168 }
2169
2170 static struct cftype cfq_blkcg_files[] = {
2171         {
2172                 .name = "weight",
2173                 .flags = CFTYPE_NOT_ON_ROOT,
2174                 .seq_show = cfq_print_weight_on_dfl,
2175                 .write = cfq_set_weight_on_dfl,
2176         },
2177         { }     /* terminate */
2178 };
2179
2180 #else /* GROUP_IOSCHED */
2181 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2182                                          struct blkcg *blkcg)
2183 {
2184         return cfqd->root_group;
2185 }
2186
2187 static inline void
2188 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2189         cfqq->cfqg = cfqg;
2190 }
2191
2192 #endif /* GROUP_IOSCHED */
2193
2194 /*
2195  * The cfqd->service_trees holds all pending cfq_queue's that have
2196  * requests waiting to be processed. It is sorted in the order that
2197  * we will service the queues.
2198  */
2199 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2200                                  bool add_front)
2201 {
2202         struct rb_node **p, *parent;
2203         struct cfq_queue *__cfqq;
2204         unsigned long rb_key;
2205         struct cfq_rb_root *st;
2206         int left;
2207         int new_cfqq = 1;
2208
2209         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2210         if (cfq_class_idle(cfqq)) {
2211                 rb_key = CFQ_IDLE_DELAY;
2212                 parent = rb_last(&st->rb);
2213                 if (parent && parent != &cfqq->rb_node) {
2214                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2215                         rb_key += __cfqq->rb_key;
2216                 } else
2217                         rb_key += jiffies;
2218         } else if (!add_front) {
2219                 /*
2220                  * Get our rb key offset. Subtract any residual slice
2221                  * value carried from last service. A negative resid
2222                  * count indicates slice overrun, and this should position
2223                  * the next service time further away in the tree.
2224                  */
2225                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2226                 rb_key -= cfqq->slice_resid;
2227                 cfqq->slice_resid = 0;
2228         } else {
2229                 rb_key = -HZ;
2230                 __cfqq = cfq_rb_first(st);
2231                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2232         }
2233
2234         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2235                 new_cfqq = 0;
2236                 /*
2237                  * same position, nothing more to do
2238                  */
2239                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2240                         return;
2241
2242                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2243                 cfqq->service_tree = NULL;
2244         }
2245
2246         left = 1;
2247         parent = NULL;
2248         cfqq->service_tree = st;
2249         p = &st->rb.rb_node;
2250         while (*p) {
2251                 parent = *p;
2252                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2253
2254                 /*
2255                  * sort by key, that represents service time.
2256                  */
2257                 if (time_before(rb_key, __cfqq->rb_key))
2258                         p = &parent->rb_left;
2259                 else {
2260                         p = &parent->rb_right;
2261                         left = 0;
2262                 }
2263         }
2264
2265         if (left)
2266                 st->left = &cfqq->rb_node;
2267
2268         cfqq->rb_key = rb_key;
2269         rb_link_node(&cfqq->rb_node, parent, p);
2270         rb_insert_color(&cfqq->rb_node, &st->rb);
2271         st->count++;
2272         if (add_front || !new_cfqq)
2273                 return;
2274         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2275 }
2276
2277 static struct cfq_queue *
2278 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2279                      sector_t sector, struct rb_node **ret_parent,
2280                      struct rb_node ***rb_link)
2281 {
2282         struct rb_node **p, *parent;
2283         struct cfq_queue *cfqq = NULL;
2284
2285         parent = NULL;
2286         p = &root->rb_node;
2287         while (*p) {
2288                 struct rb_node **n;
2289
2290                 parent = *p;
2291                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2292
2293                 /*
2294                  * Sort strictly based on sector.  Smallest to the left,
2295                  * largest to the right.
2296                  */
2297                 if (sector > blk_rq_pos(cfqq->next_rq))
2298                         n = &(*p)->rb_right;
2299                 else if (sector < blk_rq_pos(cfqq->next_rq))
2300                         n = &(*p)->rb_left;
2301                 else
2302                         break;
2303                 p = n;
2304                 cfqq = NULL;
2305         }
2306
2307         *ret_parent = parent;
2308         if (rb_link)
2309                 *rb_link = p;
2310         return cfqq;
2311 }
2312
2313 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2314 {
2315         struct rb_node **p, *parent;
2316         struct cfq_queue *__cfqq;
2317
2318         if (cfqq->p_root) {
2319                 rb_erase(&cfqq->p_node, cfqq->p_root);
2320                 cfqq->p_root = NULL;
2321         }
2322
2323         if (cfq_class_idle(cfqq))
2324                 return;
2325         if (!cfqq->next_rq)
2326                 return;
2327
2328         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2329         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2330                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2331         if (!__cfqq) {
2332                 rb_link_node(&cfqq->p_node, parent, p);
2333                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2334         } else
2335                 cfqq->p_root = NULL;
2336 }
2337
2338 /*
2339  * Update cfqq's position in the service tree.
2340  */
2341 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2342 {
2343         /*
2344          * Resorting requires the cfqq to be on the RR list already.
2345          */
2346         if (cfq_cfqq_on_rr(cfqq)) {
2347                 cfq_service_tree_add(cfqd, cfqq, 0);
2348                 cfq_prio_tree_add(cfqd, cfqq);
2349         }
2350 }
2351
2352 /*
2353  * add to busy list of queues for service, trying to be fair in ordering
2354  * the pending list according to last request service
2355  */
2356 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2357 {
2358         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2359         BUG_ON(cfq_cfqq_on_rr(cfqq));
2360         cfq_mark_cfqq_on_rr(cfqq);
2361         cfqd->busy_queues++;
2362         if (cfq_cfqq_sync(cfqq))
2363                 cfqd->busy_sync_queues++;
2364
2365         cfq_resort_rr_list(cfqd, cfqq);
2366 }
2367
2368 /*
2369  * Called when the cfqq no longer has requests pending, remove it from
2370  * the service tree.
2371  */
2372 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2373 {
2374         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2375         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2376         cfq_clear_cfqq_on_rr(cfqq);
2377
2378         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2379                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2380                 cfqq->service_tree = NULL;
2381         }
2382         if (cfqq->p_root) {
2383                 rb_erase(&cfqq->p_node, cfqq->p_root);
2384                 cfqq->p_root = NULL;
2385         }
2386
2387         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2388         BUG_ON(!cfqd->busy_queues);
2389         cfqd->busy_queues--;
2390         if (cfq_cfqq_sync(cfqq))
2391                 cfqd->busy_sync_queues--;
2392 }
2393
2394 /*
2395  * rb tree support functions
2396  */
2397 static void cfq_del_rq_rb(struct request *rq)
2398 {
2399         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2400         const int sync = rq_is_sync(rq);
2401
2402         BUG_ON(!cfqq->queued[sync]);
2403         cfqq->queued[sync]--;
2404
2405         elv_rb_del(&cfqq->sort_list, rq);
2406
2407         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2408                 /*
2409                  * Queue will be deleted from service tree when we actually
2410                  * expire it later. Right now just remove it from prio tree
2411                  * as it is empty.
2412                  */
2413                 if (cfqq->p_root) {
2414                         rb_erase(&cfqq->p_node, cfqq->p_root);
2415                         cfqq->p_root = NULL;
2416                 }
2417         }
2418 }
2419
2420 static void cfq_add_rq_rb(struct request *rq)
2421 {
2422         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2423         struct cfq_data *cfqd = cfqq->cfqd;
2424         struct request *prev;
2425
2426         cfqq->queued[rq_is_sync(rq)]++;
2427
2428         elv_rb_add(&cfqq->sort_list, rq);
2429
2430         if (!cfq_cfqq_on_rr(cfqq))
2431                 cfq_add_cfqq_rr(cfqd, cfqq);
2432
2433         /*
2434          * check if this request is a better next-serve candidate
2435          */
2436         prev = cfqq->next_rq;
2437         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2438
2439         /*
2440          * adjust priority tree position, if ->next_rq changes
2441          */
2442         if (prev != cfqq->next_rq)
2443                 cfq_prio_tree_add(cfqd, cfqq);
2444
2445         BUG_ON(!cfqq->next_rq);
2446 }
2447
2448 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2449 {
2450         elv_rb_del(&cfqq->sort_list, rq);
2451         cfqq->queued[rq_is_sync(rq)]--;
2452         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2453         cfq_add_rq_rb(rq);
2454         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2455                                  rq->cmd_flags);
2456 }
2457
2458 static struct request *
2459 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2460 {
2461         struct task_struct *tsk = current;
2462         struct cfq_io_cq *cic;
2463         struct cfq_queue *cfqq;
2464
2465         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2466         if (!cic)
2467                 return NULL;
2468
2469         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2470         if (cfqq)
2471                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2472
2473         return NULL;
2474 }
2475
2476 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2477 {
2478         struct cfq_data *cfqd = q->elevator->elevator_data;
2479
2480         cfqd->rq_in_driver++;
2481         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2482                                                 cfqd->rq_in_driver);
2483
2484         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2485 }
2486
2487 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2488 {
2489         struct cfq_data *cfqd = q->elevator->elevator_data;
2490
2491         WARN_ON(!cfqd->rq_in_driver);
2492         cfqd->rq_in_driver--;
2493         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2494                                                 cfqd->rq_in_driver);
2495 }
2496
2497 static void cfq_remove_request(struct request *rq)
2498 {
2499         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2500
2501         if (cfqq->next_rq == rq)
2502                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2503
2504         list_del_init(&rq->queuelist);
2505         cfq_del_rq_rb(rq);
2506
2507         cfqq->cfqd->rq_queued--;
2508         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2509         if (rq->cmd_flags & REQ_PRIO) {
2510                 WARN_ON(!cfqq->prio_pending);
2511                 cfqq->prio_pending--;
2512         }
2513 }
2514
2515 static int cfq_merge(struct request_queue *q, struct request **req,
2516                      struct bio *bio)
2517 {
2518         struct cfq_data *cfqd = q->elevator->elevator_data;
2519         struct request *__rq;
2520
2521         __rq = cfq_find_rq_fmerge(cfqd, bio);
2522         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2523                 *req = __rq;
2524                 return ELEVATOR_FRONT_MERGE;
2525         }
2526
2527         return ELEVATOR_NO_MERGE;
2528 }
2529
2530 static void cfq_merged_request(struct request_queue *q, struct request *req,
2531                                int type)
2532 {
2533         if (type == ELEVATOR_FRONT_MERGE) {
2534                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2535
2536                 cfq_reposition_rq_rb(cfqq, req);
2537         }
2538 }
2539
2540 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2541                                 struct bio *bio)
2542 {
2543         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2544 }
2545
2546 static void
2547 cfq_merged_requests(struct request_queue *q, struct request *rq,
2548                     struct request *next)
2549 {
2550         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2551         struct cfq_data *cfqd = q->elevator->elevator_data;
2552
2553         /*
2554          * reposition in fifo if next is older than rq
2555          */
2556         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2557             time_before(next->fifo_time, rq->fifo_time) &&
2558             cfqq == RQ_CFQQ(next)) {
2559                 list_move(&rq->queuelist, &next->queuelist);
2560                 rq->fifo_time = next->fifo_time;
2561         }
2562
2563         if (cfqq->next_rq == next)
2564                 cfqq->next_rq = rq;
2565         cfq_remove_request(next);
2566         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2567
2568         cfqq = RQ_CFQQ(next);
2569         /*
2570          * all requests of this queue are merged to other queues, delete it
2571          * from the service tree. If it's the active_queue,
2572          * cfq_dispatch_requests() will choose to expire it or do idle
2573          */
2574         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2575             cfqq != cfqd->active_queue)
2576                 cfq_del_cfqq_rr(cfqd, cfqq);
2577 }
2578
2579 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2580                            struct bio *bio)
2581 {
2582         struct cfq_data *cfqd = q->elevator->elevator_data;
2583         struct cfq_io_cq *cic;
2584         struct cfq_queue *cfqq;
2585
2586         /*
2587          * Disallow merge of a sync bio into an async request.
2588          */
2589         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2590                 return false;
2591
2592         /*
2593          * Lookup the cfqq that this bio will be queued with and allow
2594          * merge only if rq is queued there.
2595          */
2596         cic = cfq_cic_lookup(cfqd, current->io_context);
2597         if (!cic)
2598                 return false;
2599
2600         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2601         return cfqq == RQ_CFQQ(rq);
2602 }
2603
2604 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2605 {
2606         del_timer(&cfqd->idle_slice_timer);
2607         cfqg_stats_update_idle_time(cfqq->cfqg);
2608 }
2609
2610 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2611                                    struct cfq_queue *cfqq)
2612 {
2613         if (cfqq) {
2614                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2615                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2616                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2617                 cfqq->slice_start = 0;
2618                 cfqq->dispatch_start = jiffies;
2619                 cfqq->allocated_slice = 0;
2620                 cfqq->slice_end = 0;
2621                 cfqq->slice_dispatch = 0;
2622                 cfqq->nr_sectors = 0;
2623
2624                 cfq_clear_cfqq_wait_request(cfqq);
2625                 cfq_clear_cfqq_must_dispatch(cfqq);
2626                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2627                 cfq_clear_cfqq_fifo_expire(cfqq);
2628                 cfq_mark_cfqq_slice_new(cfqq);
2629
2630                 cfq_del_timer(cfqd, cfqq);
2631         }
2632
2633         cfqd->active_queue = cfqq;
2634 }
2635
2636 /*
2637  * current cfqq expired its slice (or was too idle), select new one
2638  */
2639 static void
2640 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2641                     bool timed_out)
2642 {
2643         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2644
2645         if (cfq_cfqq_wait_request(cfqq))
2646                 cfq_del_timer(cfqd, cfqq);
2647
2648         cfq_clear_cfqq_wait_request(cfqq);
2649         cfq_clear_cfqq_wait_busy(cfqq);
2650
2651         /*
2652          * If this cfqq is shared between multiple processes, check to
2653          * make sure that those processes are still issuing I/Os within
2654          * the mean seek distance.  If not, it may be time to break the
2655          * queues apart again.
2656          */
2657         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2658                 cfq_mark_cfqq_split_coop(cfqq);
2659
2660         /*
2661          * store what was left of this slice, if the queue idled/timed out
2662          */
2663         if (timed_out) {
2664                 if (cfq_cfqq_slice_new(cfqq))
2665                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2666                 else
2667                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2668                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2669         }
2670
2671         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2672
2673         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2674                 cfq_del_cfqq_rr(cfqd, cfqq);
2675
2676         cfq_resort_rr_list(cfqd, cfqq);
2677
2678         if (cfqq == cfqd->active_queue)
2679                 cfqd->active_queue = NULL;
2680
2681         if (cfqd->active_cic) {
2682                 put_io_context(cfqd->active_cic->icq.ioc);
2683                 cfqd->active_cic = NULL;
2684         }
2685 }
2686
2687 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2688 {
2689         struct cfq_queue *cfqq = cfqd->active_queue;
2690
2691         if (cfqq)
2692                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2693 }
2694
2695 /*
2696  * Get next queue for service. Unless we have a queue preemption,
2697  * we'll simply select the first cfqq in the service tree.
2698  */
2699 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2700 {
2701         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2702                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2703
2704         if (!cfqd->rq_queued)
2705                 return NULL;
2706
2707         /* There is nothing to dispatch */
2708         if (!st)
2709                 return NULL;
2710         if (RB_EMPTY_ROOT(&st->rb))
2711                 return NULL;
2712         return cfq_rb_first(st);
2713 }
2714
2715 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2716 {
2717         struct cfq_group *cfqg;
2718         struct cfq_queue *cfqq;
2719         int i, j;
2720         struct cfq_rb_root *st;
2721
2722         if (!cfqd->rq_queued)
2723                 return NULL;
2724
2725         cfqg = cfq_get_next_cfqg(cfqd);
2726         if (!cfqg)
2727                 return NULL;
2728
2729         for_each_cfqg_st(cfqg, i, j, st)
2730                 if ((cfqq = cfq_rb_first(st)) != NULL)
2731                         return cfqq;
2732         return NULL;
2733 }
2734
2735 /*
2736  * Get and set a new active queue for service.
2737  */
2738 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2739                                               struct cfq_queue *cfqq)
2740 {
2741         if (!cfqq)
2742                 cfqq = cfq_get_next_queue(cfqd);
2743
2744         __cfq_set_active_queue(cfqd, cfqq);
2745         return cfqq;
2746 }
2747
2748 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2749                                           struct request *rq)
2750 {
2751         if (blk_rq_pos(rq) >= cfqd->last_position)
2752                 return blk_rq_pos(rq) - cfqd->last_position;
2753         else
2754                 return cfqd->last_position - blk_rq_pos(rq);
2755 }
2756
2757 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2758                                struct request *rq)
2759 {
2760         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2761 }
2762
2763 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2764                                     struct cfq_queue *cur_cfqq)
2765 {
2766         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2767         struct rb_node *parent, *node;
2768         struct cfq_queue *__cfqq;
2769         sector_t sector = cfqd->last_position;
2770
2771         if (RB_EMPTY_ROOT(root))
2772                 return NULL;
2773
2774         /*
2775          * First, if we find a request starting at the end of the last
2776          * request, choose it.
2777          */
2778         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2779         if (__cfqq)
2780                 return __cfqq;
2781
2782         /*
2783          * If the exact sector wasn't found, the parent of the NULL leaf
2784          * will contain the closest sector.
2785          */
2786         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2787         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2788                 return __cfqq;
2789
2790         if (blk_rq_pos(__cfqq->next_rq) < sector)
2791                 node = rb_next(&__cfqq->p_node);
2792         else
2793                 node = rb_prev(&__cfqq->p_node);
2794         if (!node)
2795                 return NULL;
2796
2797         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2798         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2799                 return __cfqq;
2800
2801         return NULL;
2802 }
2803
2804 /*
2805  * cfqd - obvious
2806  * cur_cfqq - passed in so that we don't decide that the current queue is
2807  *            closely cooperating with itself.
2808  *
2809  * So, basically we're assuming that that cur_cfqq has dispatched at least
2810  * one request, and that cfqd->last_position reflects a position on the disk
2811  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2812  * assumption.
2813  */
2814 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2815                                               struct cfq_queue *cur_cfqq)
2816 {
2817         struct cfq_queue *cfqq;
2818
2819         if (cfq_class_idle(cur_cfqq))
2820                 return NULL;
2821         if (!cfq_cfqq_sync(cur_cfqq))
2822                 return NULL;
2823         if (CFQQ_SEEKY(cur_cfqq))
2824                 return NULL;
2825
2826         /*
2827          * Don't search priority tree if it's the only queue in the group.
2828          */
2829         if (cur_cfqq->cfqg->nr_cfqq == 1)
2830                 return NULL;
2831
2832         /*
2833          * We should notice if some of the queues are cooperating, eg
2834          * working closely on the same area of the disk. In that case,
2835          * we can group them together and don't waste time idling.
2836          */
2837         cfqq = cfqq_close(cfqd, cur_cfqq);
2838         if (!cfqq)
2839                 return NULL;
2840
2841         /* If new queue belongs to different cfq_group, don't choose it */
2842         if (cur_cfqq->cfqg != cfqq->cfqg)
2843                 return NULL;
2844
2845         /*
2846          * It only makes sense to merge sync queues.
2847          */
2848         if (!cfq_cfqq_sync(cfqq))
2849                 return NULL;
2850         if (CFQQ_SEEKY(cfqq))
2851                 return NULL;
2852
2853         /*
2854          * Do not merge queues of different priority classes
2855          */
2856         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2857                 return NULL;
2858
2859         return cfqq;
2860 }
2861
2862 /*
2863  * Determine whether we should enforce idle window for this queue.
2864  */
2865
2866 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2867 {
2868         enum wl_class_t wl_class = cfqq_class(cfqq);
2869         struct cfq_rb_root *st = cfqq->service_tree;
2870
2871         BUG_ON(!st);
2872         BUG_ON(!st->count);
2873
2874         if (!cfqd->cfq_slice_idle)
2875                 return false;
2876
2877         /* We never do for idle class queues. */
2878         if (wl_class == IDLE_WORKLOAD)
2879                 return false;
2880
2881         /* We do for queues that were marked with idle window flag. */
2882         if (cfq_cfqq_idle_window(cfqq) &&
2883            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2884                 return true;
2885
2886         /*
2887          * Otherwise, we do only if they are the last ones
2888          * in their service tree.
2889          */
2890         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2891            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2892                 return true;
2893         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2894         return false;
2895 }
2896
2897 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2898 {
2899         struct cfq_queue *cfqq = cfqd->active_queue;
2900         struct cfq_io_cq *cic;
2901         unsigned long sl, group_idle = 0;
2902
2903         /*
2904          * SSD device without seek penalty, disable idling. But only do so
2905          * for devices that support queuing, otherwise we still have a problem
2906          * with sync vs async workloads.
2907          */
2908         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2909                 return;
2910
2911         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2912         WARN_ON(cfq_cfqq_slice_new(cfqq));
2913
2914         /*
2915          * idle is disabled, either manually or by past process history
2916          */
2917         if (!cfq_should_idle(cfqd, cfqq)) {
2918                 /* no queue idling. Check for group idling */
2919                 if (cfqd->cfq_group_idle)
2920                         group_idle = cfqd->cfq_group_idle;
2921                 else
2922                         return;
2923         }
2924
2925         /*
2926          * still active requests from this queue, don't idle
2927          */
2928         if (cfqq->dispatched)
2929                 return;
2930
2931         /*
2932          * task has exited, don't wait
2933          */
2934         cic = cfqd->active_cic;
2935         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2936                 return;
2937
2938         /*
2939          * If our average think time is larger than the remaining time
2940          * slice, then don't idle. This avoids overrunning the allotted
2941          * time slice.
2942          */
2943         if (sample_valid(cic->ttime.ttime_samples) &&
2944             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2945                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2946                              cic->ttime.ttime_mean);
2947                 return;
2948         }
2949
2950         /* There are other queues in the group, don't do group idle */
2951         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2952                 return;
2953
2954         cfq_mark_cfqq_wait_request(cfqq);
2955
2956         if (group_idle)
2957                 sl = cfqd->cfq_group_idle;
2958         else
2959                 sl = cfqd->cfq_slice_idle;
2960
2961         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2962         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2963         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2964                         group_idle ? 1 : 0);
2965 }
2966
2967 /*
2968  * Move request from internal lists to the request queue dispatch list.
2969  */
2970 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2971 {
2972         struct cfq_data *cfqd = q->elevator->elevator_data;
2973         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2974
2975         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2976
2977         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2978         cfq_remove_request(rq);
2979         cfqq->dispatched++;
2980         (RQ_CFQG(rq))->dispatched++;
2981         elv_dispatch_sort(q, rq);
2982
2983         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2984         cfqq->nr_sectors += blk_rq_sectors(rq);
2985 }
2986
2987 /*
2988  * return expired entry, or NULL to just start from scratch in rbtree
2989  */
2990 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2991 {
2992         struct request *rq = NULL;
2993
2994         if (cfq_cfqq_fifo_expire(cfqq))
2995                 return NULL;
2996
2997         cfq_mark_cfqq_fifo_expire(cfqq);
2998
2999         if (list_empty(&cfqq->fifo))
3000                 return NULL;
3001
3002         rq = rq_entry_fifo(cfqq->fifo.next);
3003         if (time_before(jiffies, rq->fifo_time))
3004                 rq = NULL;
3005
3006         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
3007         return rq;
3008 }
3009
3010 static inline int
3011 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3012 {
3013         const int base_rq = cfqd->cfq_slice_async_rq;
3014
3015         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3016
3017         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3018 }
3019
3020 /*
3021  * Must be called with the queue_lock held.
3022  */
3023 static int cfqq_process_refs(struct cfq_queue *cfqq)
3024 {
3025         int process_refs, io_refs;
3026
3027         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3028         process_refs = cfqq->ref - io_refs;
3029         BUG_ON(process_refs < 0);
3030         return process_refs;
3031 }
3032
3033 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3034 {
3035         int process_refs, new_process_refs;
3036         struct cfq_queue *__cfqq;
3037
3038         /*
3039          * If there are no process references on the new_cfqq, then it is
3040          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3041          * chain may have dropped their last reference (not just their
3042          * last process reference).
3043          */
3044         if (!cfqq_process_refs(new_cfqq))
3045                 return;
3046
3047         /* Avoid a circular list and skip interim queue merges */
3048         while ((__cfqq = new_cfqq->new_cfqq)) {
3049                 if (__cfqq == cfqq)
3050                         return;
3051                 new_cfqq = __cfqq;
3052         }
3053
3054         process_refs = cfqq_process_refs(cfqq);
3055         new_process_refs = cfqq_process_refs(new_cfqq);
3056         /*
3057          * If the process for the cfqq has gone away, there is no
3058          * sense in merging the queues.
3059          */
3060         if (process_refs == 0 || new_process_refs == 0)
3061                 return;
3062
3063         /*
3064          * Merge in the direction of the lesser amount of work.
3065          */
3066         if (new_process_refs >= process_refs) {
3067                 cfqq->new_cfqq = new_cfqq;
3068                 new_cfqq->ref += process_refs;
3069         } else {
3070                 new_cfqq->new_cfqq = cfqq;
3071                 cfqq->ref += new_process_refs;
3072         }
3073 }
3074
3075 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3076                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3077 {
3078         struct cfq_queue *queue;
3079         int i;
3080         bool key_valid = false;
3081         unsigned long lowest_key = 0;
3082         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3083
3084         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3085                 /* select the one with lowest rb_key */
3086                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3087                 if (queue &&
3088                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
3089                         lowest_key = queue->rb_key;
3090                         cur_best = i;
3091                         key_valid = true;
3092                 }
3093         }
3094
3095         return cur_best;
3096 }
3097
3098 static void
3099 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3100 {
3101         unsigned slice;
3102         unsigned count;
3103         struct cfq_rb_root *st;
3104         unsigned group_slice;
3105         enum wl_class_t original_class = cfqd->serving_wl_class;
3106
3107         /* Choose next priority. RT > BE > IDLE */
3108         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3109                 cfqd->serving_wl_class = RT_WORKLOAD;
3110         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3111                 cfqd->serving_wl_class = BE_WORKLOAD;
3112         else {
3113                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3114                 cfqd->workload_expires = jiffies + 1;
3115                 return;
3116         }
3117
3118         if (original_class != cfqd->serving_wl_class)
3119                 goto new_workload;
3120
3121         /*
3122          * For RT and BE, we have to choose also the type
3123          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3124          * expiration time
3125          */
3126         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3127         count = st->count;
3128
3129         /*
3130          * check workload expiration, and that we still have other queues ready
3131          */
3132         if (count && !time_after(jiffies, cfqd->workload_expires))
3133                 return;
3134
3135 new_workload:
3136         /* otherwise select new workload type */
3137         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3138                                         cfqd->serving_wl_class);
3139         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3140         count = st->count;
3141
3142         /*
3143          * the workload slice is computed as a fraction of target latency
3144          * proportional to the number of queues in that workload, over
3145          * all the queues in the same priority class
3146          */
3147         group_slice = cfq_group_slice(cfqd, cfqg);
3148
3149         slice = group_slice * count /
3150                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3151                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3152                                         cfqg));
3153
3154         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3155                 unsigned int tmp;
3156
3157                 /*
3158                  * Async queues are currently system wide. Just taking
3159                  * proportion of queues with-in same group will lead to higher
3160                  * async ratio system wide as generally root group is going
3161                  * to have higher weight. A more accurate thing would be to
3162                  * calculate system wide asnc/sync ratio.
3163                  */
3164                 tmp = cfqd->cfq_target_latency *
3165                         cfqg_busy_async_queues(cfqd, cfqg);
3166                 tmp = tmp/cfqd->busy_queues;
3167                 slice = min_t(unsigned, slice, tmp);
3168
3169                 /* async workload slice is scaled down according to
3170                  * the sync/async slice ratio. */
3171                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3172         } else
3173                 /* sync workload slice is at least 2 * cfq_slice_idle */
3174                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3175
3176         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3177         cfq_log(cfqd, "workload slice:%d", slice);
3178         cfqd->workload_expires = jiffies + slice;
3179 }
3180
3181 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3182 {
3183         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3184         struct cfq_group *cfqg;
3185
3186         if (RB_EMPTY_ROOT(&st->rb))
3187                 return NULL;
3188         cfqg = cfq_rb_first_group(st);
3189         update_min_vdisktime(st);
3190         return cfqg;
3191 }
3192
3193 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3194 {
3195         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3196
3197         cfqd->serving_group = cfqg;
3198
3199         /* Restore the workload type data */
3200         if (cfqg->saved_wl_slice) {
3201                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3202                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3203                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3204         } else
3205                 cfqd->workload_expires = jiffies - 1;
3206
3207         choose_wl_class_and_type(cfqd, cfqg);
3208 }
3209
3210 /*
3211  * Select a queue for service. If we have a current active queue,
3212  * check whether to continue servicing it, or retrieve and set a new one.
3213  */
3214 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3215 {
3216         struct cfq_queue *cfqq, *new_cfqq = NULL;
3217
3218         cfqq = cfqd->active_queue;
3219         if (!cfqq)
3220                 goto new_queue;
3221
3222         if (!cfqd->rq_queued)
3223                 return NULL;
3224
3225         /*
3226          * We were waiting for group to get backlogged. Expire the queue
3227          */
3228         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3229                 goto expire;
3230
3231         /*
3232          * The active queue has run out of time, expire it and select new.
3233          */
3234         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3235                 /*
3236                  * If slice had not expired at the completion of last request
3237                  * we might not have turned on wait_busy flag. Don't expire
3238                  * the queue yet. Allow the group to get backlogged.
3239                  *
3240                  * The very fact that we have used the slice, that means we
3241                  * have been idling all along on this queue and it should be
3242                  * ok to wait for this request to complete.
3243                  */
3244                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3245                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3246                         cfqq = NULL;
3247                         goto keep_queue;
3248                 } else
3249                         goto check_group_idle;
3250         }
3251
3252         /*
3253          * The active queue has requests and isn't expired, allow it to
3254          * dispatch.
3255          */
3256         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3257                 goto keep_queue;
3258
3259         /*
3260          * If another queue has a request waiting within our mean seek
3261          * distance, let it run.  The expire code will check for close
3262          * cooperators and put the close queue at the front of the service
3263          * tree.  If possible, merge the expiring queue with the new cfqq.
3264          */
3265         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3266         if (new_cfqq) {
3267                 if (!cfqq->new_cfqq)
3268                         cfq_setup_merge(cfqq, new_cfqq);
3269                 goto expire;
3270         }
3271
3272         /*
3273          * No requests pending. If the active queue still has requests in
3274          * flight or is idling for a new request, allow either of these
3275          * conditions to happen (or time out) before selecting a new queue.
3276          */
3277         if (timer_pending(&cfqd->idle_slice_timer)) {
3278                 cfqq = NULL;
3279                 goto keep_queue;
3280         }
3281
3282         /*
3283          * This is a deep seek queue, but the device is much faster than
3284          * the queue can deliver, don't idle
3285          **/
3286         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3287             (cfq_cfqq_slice_new(cfqq) ||
3288             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3289                 cfq_clear_cfqq_deep(cfqq);
3290                 cfq_clear_cfqq_idle_window(cfqq);
3291         }
3292
3293         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3294                 cfqq = NULL;
3295                 goto keep_queue;
3296         }
3297
3298         /*
3299          * If group idle is enabled and there are requests dispatched from
3300          * this group, wait for requests to complete.
3301          */
3302 check_group_idle:
3303         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3304             cfqq->cfqg->dispatched &&
3305             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3306                 cfqq = NULL;
3307                 goto keep_queue;
3308         }
3309
3310 expire:
3311         cfq_slice_expired(cfqd, 0);
3312 new_queue:
3313         /*
3314          * Current queue expired. Check if we have to switch to a new
3315          * service tree
3316          */
3317         if (!new_cfqq)
3318                 cfq_choose_cfqg(cfqd);
3319
3320         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3321 keep_queue:
3322         return cfqq;
3323 }
3324
3325 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3326 {
3327         int dispatched = 0;
3328
3329         while (cfqq->next_rq) {
3330                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3331                 dispatched++;
3332         }
3333
3334         BUG_ON(!list_empty(&cfqq->fifo));
3335
3336         /* By default cfqq is not expired if it is empty. Do it explicitly */
3337         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3338         return dispatched;
3339 }
3340
3341 /*
3342  * Drain our current requests. Used for barriers and when switching
3343  * io schedulers on-the-fly.
3344  */
3345 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3346 {
3347         struct cfq_queue *cfqq;
3348         int dispatched = 0;
3349
3350         /* Expire the timeslice of the current active queue first */
3351         cfq_slice_expired(cfqd, 0);
3352         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3353                 __cfq_set_active_queue(cfqd, cfqq);
3354                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3355         }
3356
3357         BUG_ON(cfqd->busy_queues);
3358
3359         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3360         return dispatched;
3361 }
3362
3363 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3364         struct cfq_queue *cfqq)
3365 {
3366         /* the queue hasn't finished any request, can't estimate */
3367         if (cfq_cfqq_slice_new(cfqq))
3368                 return true;
3369         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3370                 cfqq->slice_end))
3371                 return true;
3372
3373         return false;
3374 }
3375
3376 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3377 {
3378         unsigned int max_dispatch;
3379
3380         /*
3381          * Drain async requests before we start sync IO
3382          */
3383         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3384                 return false;
3385
3386         /*
3387          * If this is an async queue and we have sync IO in flight, let it wait
3388          */
3389         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3390                 return false;
3391
3392         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3393         if (cfq_class_idle(cfqq))
3394                 max_dispatch = 1;
3395
3396         /*
3397          * Does this cfqq already have too much IO in flight?
3398          */
3399         if (cfqq->dispatched >= max_dispatch) {
3400                 bool promote_sync = false;
3401                 /*
3402                  * idle queue must always only have a single IO in flight
3403                  */
3404                 if (cfq_class_idle(cfqq))
3405                         return false;
3406
3407                 /*
3408                  * If there is only one sync queue
3409                  * we can ignore async queue here and give the sync
3410                  * queue no dispatch limit. The reason is a sync queue can
3411                  * preempt async queue, limiting the sync queue doesn't make
3412                  * sense. This is useful for aiostress test.
3413                  */
3414                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3415                         promote_sync = true;
3416
3417                 /*
3418                  * We have other queues, don't allow more IO from this one
3419                  */
3420                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3421                                 !promote_sync)
3422                         return false;
3423
3424                 /*
3425                  * Sole queue user, no limit
3426                  */
3427                 if (cfqd->busy_queues == 1 || promote_sync)
3428                         max_dispatch = -1;
3429                 else
3430                         /*
3431                          * Normally we start throttling cfqq when cfq_quantum/2
3432                          * requests have been dispatched. But we can drive
3433                          * deeper queue depths at the beginning of slice
3434                          * subjected to upper limit of cfq_quantum.
3435                          * */
3436                         max_dispatch = cfqd->cfq_quantum;
3437         }
3438
3439         /*
3440          * Async queues must wait a bit before being allowed dispatch.
3441          * We also ramp up the dispatch depth gradually for async IO,
3442          * based on the last sync IO we serviced
3443          */
3444         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3445                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3446                 unsigned int depth;
3447
3448                 depth = last_sync / cfqd->cfq_slice[1];
3449                 if (!depth && !cfqq->dispatched)
3450                         depth = 1;
3451                 if (depth < max_dispatch)
3452                         max_dispatch = depth;
3453         }
3454
3455         /*
3456          * If we're below the current max, allow a dispatch
3457          */
3458         return cfqq->dispatched < max_dispatch;
3459 }
3460
3461 /*
3462  * Dispatch a request from cfqq, moving them to the request queue
3463  * dispatch list.
3464  */
3465 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3466 {
3467         struct request *rq;
3468
3469         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3470
3471         if (!cfq_may_dispatch(cfqd, cfqq))
3472                 return false;
3473
3474         /*
3475          * follow expired path, else get first next available
3476          */
3477         rq = cfq_check_fifo(cfqq);
3478         if (!rq)
3479                 rq = cfqq->next_rq;
3480
3481         /*
3482          * insert request into driver dispatch list
3483          */
3484         cfq_dispatch_insert(cfqd->queue, rq);
3485
3486         if (!cfqd->active_cic) {
3487                 struct cfq_io_cq *cic = RQ_CIC(rq);
3488
3489                 atomic_long_inc(&cic->icq.ioc->refcount);
3490                 cfqd->active_cic = cic;
3491         }
3492
3493         return true;
3494 }
3495
3496 /*
3497  * Find the cfqq that we need to service and move a request from that to the
3498  * dispatch list
3499  */
3500 static int cfq_dispatch_requests(struct request_queue *q, int force)
3501 {
3502         struct cfq_data *cfqd = q->elevator->elevator_data;
3503         struct cfq_queue *cfqq;
3504
3505         if (!cfqd->busy_queues)
3506                 return 0;
3507
3508         if (unlikely(force))
3509                 return cfq_forced_dispatch(cfqd);
3510
3511         cfqq = cfq_select_queue(cfqd);
3512         if (!cfqq)
3513                 return 0;
3514
3515         /*
3516          * Dispatch a request from this cfqq, if it is allowed
3517          */
3518         if (!cfq_dispatch_request(cfqd, cfqq))
3519                 return 0;
3520
3521         cfqq->slice_dispatch++;
3522         cfq_clear_cfqq_must_dispatch(cfqq);
3523
3524         /*
3525          * expire an async queue immediately if it has used up its slice. idle
3526          * queue always expire after 1 dispatch round.
3527          */
3528         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3529             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3530             cfq_class_idle(cfqq))) {
3531                 cfqq->slice_end = jiffies + 1;
3532                 cfq_slice_expired(cfqd, 0);
3533         }
3534
3535         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3536         return 1;
3537 }
3538
3539 /*
3540  * task holds one reference to the queue, dropped when task exits. each rq
3541  * in-flight on this queue also holds a reference, dropped when rq is freed.
3542  *
3543  * Each cfq queue took a reference on the parent group. Drop it now.
3544  * queue lock must be held here.
3545  */
3546 static void cfq_put_queue(struct cfq_queue *cfqq)
3547 {
3548         struct cfq_data *cfqd = cfqq->cfqd;
3549         struct cfq_group *cfqg;
3550
3551         BUG_ON(cfqq->ref <= 0);
3552
3553         cfqq->ref--;
3554         if (cfqq->ref)
3555                 return;
3556
3557         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3558         BUG_ON(rb_first(&cfqq->sort_list));
3559         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3560         cfqg = cfqq->cfqg;
3561
3562         if (unlikely(cfqd->active_queue == cfqq)) {
3563                 __cfq_slice_expired(cfqd, cfqq, 0);
3564                 cfq_schedule_dispatch(cfqd);
3565         }
3566
3567         BUG_ON(cfq_cfqq_on_rr(cfqq));
3568         kmem_cache_free(cfq_pool, cfqq);
3569         cfqg_put(cfqg);
3570 }
3571
3572 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3573 {
3574         struct cfq_queue *__cfqq, *next;
3575
3576         /*
3577          * If this queue was scheduled to merge with another queue, be
3578          * sure to drop the reference taken on that queue (and others in
3579          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3580          */
3581         __cfqq = cfqq->new_cfqq;
3582         while (__cfqq) {
3583                 if (__cfqq == cfqq) {
3584                         WARN(1, "cfqq->new_cfqq loop detected\n");
3585                         break;
3586                 }
3587                 next = __cfqq->new_cfqq;
3588                 cfq_put_queue(__cfqq);
3589                 __cfqq = next;
3590         }
3591 }
3592
3593 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3594 {
3595         if (unlikely(cfqq == cfqd->active_queue)) {
3596                 __cfq_slice_expired(cfqd, cfqq, 0);
3597                 cfq_schedule_dispatch(cfqd);
3598         }
3599
3600         cfq_put_cooperator(cfqq);
3601
3602         cfq_put_queue(cfqq);
3603 }
3604
3605 static void cfq_init_icq(struct io_cq *icq)
3606 {
3607         struct cfq_io_cq *cic = icq_to_cic(icq);
3608
3609         cic->ttime.last_end_request = jiffies;
3610 }
3611
3612 static void cfq_exit_icq(struct io_cq *icq)
3613 {
3614         struct cfq_io_cq *cic = icq_to_cic(icq);
3615         struct cfq_data *cfqd = cic_to_cfqd(cic);
3616
3617         if (cic_to_cfqq(cic, false)) {
3618                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3619                 cic_set_cfqq(cic, NULL, false);
3620         }
3621
3622         if (cic_to_cfqq(cic, true)) {
3623                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3624                 cic_set_cfqq(cic, NULL, true);
3625         }
3626 }
3627
3628 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3629 {
3630         struct task_struct *tsk = current;
3631         int ioprio_class;
3632
3633         if (!cfq_cfqq_prio_changed(cfqq))
3634                 return;
3635
3636         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3637         switch (ioprio_class) {
3638         default:
3639                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3640         case IOPRIO_CLASS_NONE:
3641                 /*
3642                  * no prio set, inherit CPU scheduling settings
3643                  */
3644                 cfqq->ioprio = task_nice_ioprio(tsk);
3645                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3646                 break;
3647         case IOPRIO_CLASS_RT:
3648                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3649                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3650                 break;
3651         case IOPRIO_CLASS_BE:
3652                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3653                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3654                 break;
3655         case IOPRIO_CLASS_IDLE:
3656                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3657                 cfqq->ioprio = 7;
3658                 cfq_clear_cfqq_idle_window(cfqq);
3659                 break;
3660         }
3661
3662         /*
3663          * keep track of original prio settings in case we have to temporarily
3664          * elevate the priority of this queue
3665          */
3666         cfqq->org_ioprio = cfqq->ioprio;
3667         cfq_clear_cfqq_prio_changed(cfqq);
3668 }
3669
3670 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3671 {
3672         int ioprio = cic->icq.ioc->ioprio;
3673         struct cfq_data *cfqd = cic_to_cfqd(cic);
3674         struct cfq_queue *cfqq;
3675
3676         /*
3677          * Check whether ioprio has changed.  The condition may trigger
3678          * spuriously on a newly created cic but there's no harm.
3679          */
3680         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3681                 return;
3682
3683         cfqq = cic_to_cfqq(cic, false);
3684         if (cfqq) {
3685                 cfq_put_queue(cfqq);
3686                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3687                 cic_set_cfqq(cic, cfqq, false);
3688         }
3689
3690         cfqq = cic_to_cfqq(cic, true);
3691         if (cfqq)
3692                 cfq_mark_cfqq_prio_changed(cfqq);
3693
3694         cic->ioprio = ioprio;
3695 }
3696
3697 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3698                           pid_t pid, bool is_sync)
3699 {
3700         RB_CLEAR_NODE(&cfqq->rb_node);
3701         RB_CLEAR_NODE(&cfqq->p_node);
3702         INIT_LIST_HEAD(&cfqq->fifo);
3703
3704         cfqq->ref = 0;
3705         cfqq->cfqd = cfqd;
3706
3707         cfq_mark_cfqq_prio_changed(cfqq);
3708
3709         if (is_sync) {
3710                 if (!cfq_class_idle(cfqq))
3711                         cfq_mark_cfqq_idle_window(cfqq);
3712                 cfq_mark_cfqq_sync(cfqq);
3713         }
3714         cfqq->pid = pid;
3715 }
3716
3717 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3718 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3719 {
3720         struct cfq_data *cfqd = cic_to_cfqd(cic);
3721         struct cfq_queue *cfqq;
3722         uint64_t serial_nr;
3723
3724         rcu_read_lock();
3725         serial_nr = bio_blkcg(bio)->css.serial_nr;
3726         rcu_read_unlock();
3727
3728         /*
3729          * Check whether blkcg has changed.  The condition may trigger
3730          * spuriously on a newly created cic but there's no harm.
3731          */
3732         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3733                 return;
3734
3735         /*
3736          * Drop reference to queues.  New queues will be assigned in new
3737          * group upon arrival of fresh requests.
3738          */
3739         cfqq = cic_to_cfqq(cic, false);
3740         if (cfqq) {
3741                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3742                 cic_set_cfqq(cic, NULL, false);
3743                 cfq_put_queue(cfqq);
3744         }
3745
3746         cfqq = cic_to_cfqq(cic, true);
3747         if (cfqq) {
3748                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3749                 cic_set_cfqq(cic, NULL, true);
3750                 cfq_put_queue(cfqq);
3751         }
3752
3753         cic->blkcg_serial_nr = serial_nr;
3754 }
3755 #else
3756 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3757 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3758
3759 static struct cfq_queue **
3760 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3761 {
3762         switch (ioprio_class) {
3763         case IOPRIO_CLASS_RT:
3764                 return &cfqg->async_cfqq[0][ioprio];
3765         case IOPRIO_CLASS_NONE:
3766                 ioprio = IOPRIO_NORM;
3767                 /* fall through */
3768         case IOPRIO_CLASS_BE:
3769                 return &cfqg->async_cfqq[1][ioprio];
3770         case IOPRIO_CLASS_IDLE:
3771                 return &cfqg->async_idle_cfqq;
3772         default:
3773                 BUG();
3774         }
3775 }
3776
3777 static struct cfq_queue *
3778 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3779               struct bio *bio)
3780 {
3781         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3782         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3783         struct cfq_queue **async_cfqq = NULL;
3784         struct cfq_queue *cfqq;
3785         struct cfq_group *cfqg;
3786
3787         rcu_read_lock();
3788         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3789         if (!cfqg) {
3790                 cfqq = &cfqd->oom_cfqq;
3791                 goto out;
3792         }
3793
3794         if (!is_sync) {
3795                 if (!ioprio_valid(cic->ioprio)) {
3796                         struct task_struct *tsk = current;
3797                         ioprio = task_nice_ioprio(tsk);
3798                         ioprio_class = task_nice_ioclass(tsk);
3799                 }
3800                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3801                 cfqq = *async_cfqq;
3802                 if (cfqq)
3803                         goto out;
3804         }
3805
3806         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3807                                      cfqd->queue->node);
3808         if (!cfqq) {
3809                 cfqq = &cfqd->oom_cfqq;
3810                 goto out;
3811         }
3812
3813         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3814         cfq_init_prio_data(cfqq, cic);
3815         cfq_link_cfqq_cfqg(cfqq, cfqg);
3816         cfq_log_cfqq(cfqd, cfqq, "alloced");
3817
3818         if (async_cfqq) {
3819                 /* a new async queue is created, pin and remember */
3820                 cfqq->ref++;
3821                 *async_cfqq = cfqq;
3822         }
3823 out:
3824         cfqq->ref++;
3825         rcu_read_unlock();
3826         return cfqq;
3827 }
3828
3829 static void
3830 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3831 {
3832         unsigned long elapsed = jiffies - ttime->last_end_request;
3833         elapsed = min(elapsed, 2UL * slice_idle);
3834
3835         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3836         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3837         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3838 }
3839
3840 static void
3841 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3842                         struct cfq_io_cq *cic)
3843 {
3844         if (cfq_cfqq_sync(cfqq)) {
3845                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3846                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3847                         cfqd->cfq_slice_idle);
3848         }
3849 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3850         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3851 #endif
3852 }
3853
3854 static void
3855 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3856                        struct request *rq)
3857 {
3858         sector_t sdist = 0;
3859         sector_t n_sec = blk_rq_sectors(rq);
3860         if (cfqq->last_request_pos) {
3861                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3862                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3863                 else
3864                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3865         }
3866
3867         cfqq->seek_history <<= 1;
3868         if (blk_queue_nonrot(cfqd->queue))
3869                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3870         else
3871                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3872 }
3873
3874 /*
3875  * Disable idle window if the process thinks too long or seeks so much that
3876  * it doesn't matter
3877  */
3878 static void
3879 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3880                        struct cfq_io_cq *cic)
3881 {
3882         int old_idle, enable_idle;
3883
3884         /*
3885          * Don't idle for async or idle io prio class
3886          */
3887         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3888                 return;
3889
3890         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3891
3892         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3893                 cfq_mark_cfqq_deep(cfqq);
3894
3895         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3896                 enable_idle = 0;
3897         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3898                  !cfqd->cfq_slice_idle ||
3899                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3900                 enable_idle = 0;
3901         else if (sample_valid(cic->ttime.ttime_samples)) {
3902                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3903                         enable_idle = 0;
3904                 else
3905                         enable_idle = 1;
3906         }
3907
3908         if (old_idle != enable_idle) {
3909                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3910                 if (enable_idle)
3911                         cfq_mark_cfqq_idle_window(cfqq);
3912                 else
3913                         cfq_clear_cfqq_idle_window(cfqq);
3914         }
3915 }
3916
3917 /*
3918  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3919  * no or if we aren't sure, a 1 will cause a preempt.
3920  */
3921 static bool
3922 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3923                    struct request *rq)
3924 {
3925         struct cfq_queue *cfqq;
3926
3927         cfqq = cfqd->active_queue;
3928         if (!cfqq)
3929                 return false;
3930
3931         if (cfq_class_idle(new_cfqq))
3932                 return false;
3933
3934         if (cfq_class_idle(cfqq))
3935                 return true;
3936
3937         /*
3938          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3939          */
3940         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3941                 return false;
3942
3943         /*
3944          * if the new request is sync, but the currently running queue is
3945          * not, let the sync request have priority.
3946          */
3947         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3948                 return true;
3949
3950         if (new_cfqq->cfqg != cfqq->cfqg)
3951                 return false;
3952
3953         if (cfq_slice_used(cfqq))
3954                 return true;
3955
3956         /* Allow preemption only if we are idling on sync-noidle tree */
3957         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3958             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3959             new_cfqq->service_tree->count == 2 &&
3960             RB_EMPTY_ROOT(&cfqq->sort_list))
3961                 return true;
3962
3963         /*
3964          * So both queues are sync. Let the new request get disk time if
3965          * it's a metadata request and the current queue is doing regular IO.
3966          */
3967         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3968                 return true;
3969
3970         /*
3971          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3972          */
3973         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3974                 return true;
3975
3976         /* An idle queue should not be idle now for some reason */
3977         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3978                 return true;
3979
3980         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3981                 return false;
3982
3983         /*
3984          * if this request is as-good as one we would expect from the
3985          * current cfqq, let it preempt
3986          */
3987         if (cfq_rq_close(cfqd, cfqq, rq))
3988                 return true;
3989
3990         return false;
3991 }
3992
3993 /*
3994  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3995  * let it have half of its nominal slice.
3996  */
3997 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3998 {
3999         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
4000
4001         cfq_log_cfqq(cfqd, cfqq, "preempt");
4002         cfq_slice_expired(cfqd, 1);
4003
4004         /*
4005          * workload type is changed, don't save slice, otherwise preempt
4006          * doesn't happen
4007          */
4008         if (old_type != cfqq_type(cfqq))
4009                 cfqq->cfqg->saved_wl_slice = 0;
4010
4011         /*
4012          * Put the new queue at the front of the of the current list,
4013          * so we know that it will be selected next.
4014          */
4015         BUG_ON(!cfq_cfqq_on_rr(cfqq));
4016
4017         cfq_service_tree_add(cfqd, cfqq, 1);
4018
4019         cfqq->slice_end = 0;
4020         cfq_mark_cfqq_slice_new(cfqq);
4021 }
4022
4023 /*
4024  * Called when a new fs request (rq) is added (to cfqq). Check if there's
4025  * something we should do about it
4026  */
4027 static void
4028 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
4029                 struct request *rq)
4030 {
4031         struct cfq_io_cq *cic = RQ_CIC(rq);
4032
4033         cfqd->rq_queued++;
4034         if (rq->cmd_flags & REQ_PRIO)
4035                 cfqq->prio_pending++;
4036
4037         cfq_update_io_thinktime(cfqd, cfqq, cic);
4038         cfq_update_io_seektime(cfqd, cfqq, rq);
4039         cfq_update_idle_window(cfqd, cfqq, cic);
4040
4041         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
4042
4043         if (cfqq == cfqd->active_queue) {
4044                 /*
4045                  * Remember that we saw a request from this process, but
4046                  * don't start queuing just yet. Otherwise we risk seeing lots
4047                  * of tiny requests, because we disrupt the normal plugging
4048                  * and merging. If the request is already larger than a single
4049                  * page, let it rip immediately. For that case we assume that
4050                  * merging is already done. Ditto for a busy system that
4051                  * has other work pending, don't risk delaying until the
4052                  * idle timer unplug to continue working.
4053                  */
4054                 if (cfq_cfqq_wait_request(cfqq)) {
4055                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
4056                             cfqd->busy_queues > 1) {
4057                                 cfq_del_timer(cfqd, cfqq);
4058                                 cfq_clear_cfqq_wait_request(cfqq);
4059                                 __blk_run_queue(cfqd->queue);
4060                         } else {
4061                                 cfqg_stats_update_idle_time(cfqq->cfqg);
4062                                 cfq_mark_cfqq_must_dispatch(cfqq);
4063                         }
4064                 }
4065         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
4066                 /*
4067                  * not the active queue - expire current slice if it is
4068                  * idle and has expired it's mean thinktime or this new queue
4069                  * has some old slice time left and is of higher priority or
4070                  * this new queue is RT and the current one is BE
4071                  */
4072                 cfq_preempt_queue(cfqd, cfqq);
4073                 __blk_run_queue(cfqd->queue);
4074         }
4075 }
4076
4077 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4078 {
4079         struct cfq_data *cfqd = q->elevator->elevator_data;
4080         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4081
4082         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4083         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4084
4085         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4086         list_add_tail(&rq->queuelist, &cfqq->fifo);
4087         cfq_add_rq_rb(rq);
4088         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4089                                  rq->cmd_flags);
4090         cfq_rq_enqueued(cfqd, cfqq, rq);
4091 }
4092
4093 /*
4094  * Update hw_tag based on peak queue depth over 50 samples under
4095  * sufficient load.
4096  */
4097 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4098 {
4099         struct cfq_queue *cfqq = cfqd->active_queue;
4100
4101         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4102                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4103
4104         if (cfqd->hw_tag == 1)
4105                 return;
4106
4107         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4108             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4109                 return;
4110
4111         /*
4112          * If active queue hasn't enough requests and can idle, cfq might not
4113          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4114          * case
4115          */
4116         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4117             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4118             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4119                 return;
4120
4121         if (cfqd->hw_tag_samples++ < 50)
4122                 return;
4123
4124         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4125                 cfqd->hw_tag = 1;
4126         else
4127                 cfqd->hw_tag = 0;
4128 }
4129
4130 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4131 {
4132         struct cfq_io_cq *cic = cfqd->active_cic;
4133
4134         /* If the queue already has requests, don't wait */
4135         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4136                 return false;
4137
4138         /* If there are other queues in the group, don't wait */
4139         if (cfqq->cfqg->nr_cfqq > 1)
4140                 return false;
4141
4142         /* the only queue in the group, but think time is big */
4143         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4144                 return false;
4145
4146         if (cfq_slice_used(cfqq))
4147                 return true;
4148
4149         /* if slice left is less than think time, wait busy */
4150         if (cic && sample_valid(cic->ttime.ttime_samples)
4151             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4152                 return true;
4153
4154         /*
4155          * If think times is less than a jiffy than ttime_mean=0 and above
4156          * will not be true. It might happen that slice has not expired yet
4157          * but will expire soon (4-5 ns) during select_queue(). To cover the
4158          * case where think time is less than a jiffy, mark the queue wait
4159          * busy if only 1 jiffy is left in the slice.
4160          */
4161         if (cfqq->slice_end - jiffies == 1)
4162                 return true;
4163
4164         return false;
4165 }
4166
4167 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4168 {
4169         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4170         struct cfq_data *cfqd = cfqq->cfqd;
4171         const int sync = rq_is_sync(rq);
4172         unsigned long now;
4173
4174         now = jiffies;
4175         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4176                      !!(rq->cmd_flags & REQ_NOIDLE));
4177
4178         cfq_update_hw_tag(cfqd);
4179
4180         WARN_ON(!cfqd->rq_in_driver);
4181         WARN_ON(!cfqq->dispatched);
4182         cfqd->rq_in_driver--;
4183         cfqq->dispatched--;
4184         (RQ_CFQG(rq))->dispatched--;
4185         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4186                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4187
4188         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4189
4190         if (sync) {
4191                 struct cfq_rb_root *st;
4192
4193                 RQ_CIC(rq)->ttime.last_end_request = now;
4194
4195                 if (cfq_cfqq_on_rr(cfqq))
4196                         st = cfqq->service_tree;
4197                 else
4198                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4199                                         cfqq_type(cfqq));
4200
4201                 st->ttime.last_end_request = now;
4202                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4203                         cfqd->last_delayed_sync = now;
4204         }
4205
4206 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4207         cfqq->cfqg->ttime.last_end_request = now;
4208 #endif
4209
4210         /*
4211          * If this is the active queue, check if it needs to be expired,
4212          * or if we want to idle in case it has no pending requests.
4213          */
4214         if (cfqd->active_queue == cfqq) {
4215                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4216
4217                 if (cfq_cfqq_slice_new(cfqq)) {
4218                         cfq_set_prio_slice(cfqd, cfqq);
4219                         cfq_clear_cfqq_slice_new(cfqq);
4220                 }
4221
4222                 /*
4223                  * Should we wait for next request to come in before we expire
4224                  * the queue.
4225                  */
4226                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4227                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4228                         if (!cfqd->cfq_slice_idle)
4229                                 extend_sl = cfqd->cfq_group_idle;
4230                         cfqq->slice_end = jiffies + extend_sl;
4231                         cfq_mark_cfqq_wait_busy(cfqq);
4232                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4233                 }
4234
4235                 /*
4236                  * Idling is not enabled on:
4237                  * - expired queues
4238                  * - idle-priority queues
4239                  * - async queues
4240                  * - queues with still some requests queued
4241                  * - when there is a close cooperator
4242                  */
4243                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4244                         cfq_slice_expired(cfqd, 1);
4245                 else if (sync && cfqq_empty &&
4246                          !cfq_close_cooperator(cfqd, cfqq)) {
4247                         cfq_arm_slice_timer(cfqd);
4248                 }
4249         }
4250
4251         if (!cfqd->rq_in_driver)
4252                 cfq_schedule_dispatch(cfqd);
4253 }
4254
4255 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4256 {
4257         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4258                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4259                 return ELV_MQUEUE_MUST;
4260         }
4261
4262         return ELV_MQUEUE_MAY;
4263 }
4264
4265 static int cfq_may_queue(struct request_queue *q, int rw)
4266 {
4267         struct cfq_data *cfqd = q->elevator->elevator_data;
4268         struct task_struct *tsk = current;
4269         struct cfq_io_cq *cic;
4270         struct cfq_queue *cfqq;
4271
4272         /*
4273          * don't force setup of a queue from here, as a call to may_queue
4274          * does not necessarily imply that a request actually will be queued.
4275          * so just lookup a possibly existing queue, or return 'may queue'
4276          * if that fails
4277          */
4278         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4279         if (!cic)
4280                 return ELV_MQUEUE_MAY;
4281
4282         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4283         if (cfqq) {
4284                 cfq_init_prio_data(cfqq, cic);
4285
4286                 return __cfq_may_queue(cfqq);
4287         }
4288
4289         return ELV_MQUEUE_MAY;
4290 }
4291
4292 /*
4293  * queue lock held here
4294  */
4295 static void cfq_put_request(struct request *rq)
4296 {
4297         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4298
4299         if (cfqq) {
4300                 const int rw = rq_data_dir(rq);
4301
4302                 BUG_ON(!cfqq->allocated[rw]);
4303                 cfqq->allocated[rw]--;
4304
4305                 /* Put down rq reference on cfqg */
4306                 cfqg_put(RQ_CFQG(rq));
4307                 rq->elv.priv[0] = NULL;
4308                 rq->elv.priv[1] = NULL;
4309
4310                 cfq_put_queue(cfqq);
4311         }
4312 }
4313
4314 static struct cfq_queue *
4315 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4316                 struct cfq_queue *cfqq)
4317 {
4318         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4319         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4320         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4321         cfq_put_queue(cfqq);
4322         return cic_to_cfqq(cic, 1);
4323 }
4324
4325 /*
4326  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4327  * was the last process referring to said cfqq.
4328  */
4329 static struct cfq_queue *
4330 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4331 {
4332         if (cfqq_process_refs(cfqq) == 1) {
4333                 cfqq->pid = current->pid;
4334                 cfq_clear_cfqq_coop(cfqq);
4335                 cfq_clear_cfqq_split_coop(cfqq);
4336                 return cfqq;
4337         }
4338
4339         cic_set_cfqq(cic, NULL, 1);
4340
4341         cfq_put_cooperator(cfqq);
4342
4343         cfq_put_queue(cfqq);
4344         return NULL;
4345 }
4346 /*
4347  * Allocate cfq data structures associated with this request.
4348  */
4349 static int
4350 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4351                 gfp_t gfp_mask)
4352 {
4353         struct cfq_data *cfqd = q->elevator->elevator_data;
4354         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4355         const int rw = rq_data_dir(rq);
4356         const bool is_sync = rq_is_sync(rq);
4357         struct cfq_queue *cfqq;
4358
4359         spin_lock_irq(q->queue_lock);
4360
4361         check_ioprio_changed(cic, bio);
4362         check_blkcg_changed(cic, bio);
4363 new_queue:
4364         cfqq = cic_to_cfqq(cic, is_sync);
4365         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4366                 if (cfqq)
4367                         cfq_put_queue(cfqq);
4368                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4369                 cic_set_cfqq(cic, cfqq, is_sync);
4370         } else {
4371                 /*
4372                  * If the queue was seeky for too long, break it apart.
4373                  */
4374                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4375                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4376                         cfqq = split_cfqq(cic, cfqq);
4377                         if (!cfqq)
4378                                 goto new_queue;
4379                 }
4380
4381                 /*
4382                  * Check to see if this queue is scheduled to merge with
4383                  * another, closely cooperating queue.  The merging of
4384                  * queues happens here as it must be done in process context.
4385                  * The reference on new_cfqq was taken in merge_cfqqs.
4386                  */
4387                 if (cfqq->new_cfqq)
4388                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4389         }
4390
4391         cfqq->allocated[rw]++;
4392
4393         cfqq->ref++;
4394         cfqg_get(cfqq->cfqg);
4395         rq->elv.priv[0] = cfqq;
4396         rq->elv.priv[1] = cfqq->cfqg;
4397         spin_unlock_irq(q->queue_lock);
4398         return 0;
4399 }
4400
4401 static void cfq_kick_queue(struct work_struct *work)
4402 {
4403         struct cfq_data *cfqd =
4404                 container_of(work, struct cfq_data, unplug_work);
4405         struct request_queue *q = cfqd->queue;
4406
4407         spin_lock_irq(q->queue_lock);
4408         __blk_run_queue(cfqd->queue);
4409         spin_unlock_irq(q->queue_lock);
4410 }
4411
4412 /*
4413  * Timer running if the active_queue is currently idling inside its time slice
4414  */
4415 static void cfq_idle_slice_timer(unsigned long data)
4416 {
4417         struct cfq_data *cfqd = (struct cfq_data *) data;
4418         struct cfq_queue *cfqq;
4419         unsigned long flags;
4420         int timed_out = 1;
4421
4422         cfq_log(cfqd, "idle timer fired");
4423
4424         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4425
4426         cfqq = cfqd->active_queue;
4427         if (cfqq) {
4428                 timed_out = 0;
4429
4430                 /*
4431                  * We saw a request before the queue expired, let it through
4432                  */
4433                 if (cfq_cfqq_must_dispatch(cfqq))
4434                         goto out_kick;
4435
4436                 /*
4437                  * expired
4438                  */
4439                 if (cfq_slice_used(cfqq))
4440                         goto expire;
4441
4442                 /*
4443                  * only expire and reinvoke request handler, if there are
4444                  * other queues with pending requests
4445                  */
4446                 if (!cfqd->busy_queues)
4447                         goto out_cont;
4448
4449                 /*
4450                  * not expired and it has a request pending, let it dispatch
4451                  */
4452                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4453                         goto out_kick;
4454
4455                 /*
4456                  * Queue depth flag is reset only when the idle didn't succeed
4457                  */
4458                 cfq_clear_cfqq_deep(cfqq);
4459         }
4460 expire:
4461         cfq_slice_expired(cfqd, timed_out);
4462 out_kick:
4463         cfq_schedule_dispatch(cfqd);
4464 out_cont:
4465         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4466 }
4467
4468 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4469 {
4470         del_timer_sync(&cfqd->idle_slice_timer);
4471         cancel_work_sync(&cfqd->unplug_work);
4472 }
4473
4474 static void cfq_exit_queue(struct elevator_queue *e)
4475 {
4476         struct cfq_data *cfqd = e->elevator_data;
4477         struct request_queue *q = cfqd->queue;
4478
4479         cfq_shutdown_timer_wq(cfqd);
4480
4481         spin_lock_irq(q->queue_lock);
4482
4483         if (cfqd->active_queue)
4484                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4485
4486         spin_unlock_irq(q->queue_lock);
4487
4488         cfq_shutdown_timer_wq(cfqd);
4489
4490 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4491         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4492 #else
4493         kfree(cfqd->root_group);
4494 #endif
4495         kfree(cfqd);
4496 }
4497
4498 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4499 {
4500         struct cfq_data *cfqd;
4501         struct blkcg_gq *blkg __maybe_unused;
4502         int i, ret;
4503         struct elevator_queue *eq;
4504
4505         eq = elevator_alloc(q, e);
4506         if (!eq)
4507                 return -ENOMEM;
4508
4509         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4510         if (!cfqd) {
4511                 kobject_put(&eq->kobj);
4512                 return -ENOMEM;
4513         }
4514         eq->elevator_data = cfqd;
4515
4516         cfqd->queue = q;
4517         spin_lock_irq(q->queue_lock);
4518         q->elevator = eq;
4519         spin_unlock_irq(q->queue_lock);
4520
4521         /* Init root service tree */
4522         cfqd->grp_service_tree = CFQ_RB_ROOT;
4523
4524         /* Init root group and prefer root group over other groups by default */
4525 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4526         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4527         if (ret)
4528                 goto out_free;
4529
4530         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4531 #else
4532         ret = -ENOMEM;
4533         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4534                                         GFP_KERNEL, cfqd->queue->node);
4535         if (!cfqd->root_group)
4536                 goto out_free;
4537
4538         cfq_init_cfqg_base(cfqd->root_group);
4539         cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4540         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL;
4541 #endif
4542
4543         /*
4544          * Not strictly needed (since RB_ROOT just clears the node and we
4545          * zeroed cfqd on alloc), but better be safe in case someone decides
4546          * to add magic to the rb code
4547          */
4548         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4549                 cfqd->prio_trees[i] = RB_ROOT;
4550
4551         /*
4552          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4553          * Grab a permanent reference to it, so that the normal code flow
4554          * will not attempt to free it.  oom_cfqq is linked to root_group
4555          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4556          * the reference from linking right away.
4557          */
4558         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4559         cfqd->oom_cfqq.ref++;
4560
4561         spin_lock_irq(q->queue_lock);
4562         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4563         cfqg_put(cfqd->root_group);
4564         spin_unlock_irq(q->queue_lock);
4565
4566         init_timer(&cfqd->idle_slice_timer);
4567         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4568         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4569
4570         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4571
4572         cfqd->cfq_quantum = cfq_quantum;
4573         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4574         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4575         cfqd->cfq_back_max = cfq_back_max;
4576         cfqd->cfq_back_penalty = cfq_back_penalty;
4577         cfqd->cfq_slice[0] = cfq_slice_async;
4578         cfqd->cfq_slice[1] = cfq_slice_sync;
4579         cfqd->cfq_target_latency = cfq_target_latency;
4580         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4581         cfqd->cfq_slice_idle = cfq_slice_idle;
4582         cfqd->cfq_group_idle = cfq_group_idle;
4583         cfqd->cfq_latency = 1;
4584         cfqd->hw_tag = -1;
4585         /*
4586          * we optimistically start assuming sync ops weren't delayed in last
4587          * second, in order to have larger depth for async operations.
4588          */
4589         cfqd->last_delayed_sync = jiffies - HZ;
4590         return 0;
4591
4592 out_free:
4593         kfree(cfqd);
4594         kobject_put(&eq->kobj);
4595         return ret;
4596 }
4597
4598 static void cfq_registered_queue(struct request_queue *q)
4599 {
4600         struct elevator_queue *e = q->elevator;
4601         struct cfq_data *cfqd = e->elevator_data;
4602
4603         /*
4604          * Default to IOPS mode with no idling for SSDs
4605          */
4606         if (blk_queue_nonrot(q))
4607                 cfqd->cfq_slice_idle = 0;
4608 }
4609
4610 /*
4611  * sysfs parts below -->
4612  */
4613 static ssize_t
4614 cfq_var_show(unsigned int var, char *page)
4615 {
4616         return sprintf(page, "%u\n", var);
4617 }
4618
4619 static ssize_t
4620 cfq_var_store(unsigned int *var, const char *page, size_t count)
4621 {
4622         char *p = (char *) page;
4623
4624         *var = simple_strtoul(p, &p, 10);
4625         return count;
4626 }
4627
4628 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4629 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4630 {                                                                       \
4631         struct cfq_data *cfqd = e->elevator_data;                       \
4632         unsigned int __data = __VAR;                                    \
4633         if (__CONV)                                                     \
4634                 __data = jiffies_to_msecs(__data);                      \
4635         return cfq_var_show(__data, (page));                            \
4636 }
4637 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4638 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4639 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4640 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4641 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4642 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4643 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4644 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4645 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4646 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4647 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4648 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4649 #undef SHOW_FUNCTION
4650
4651 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4652 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4653 {                                                                       \
4654         struct cfq_data *cfqd = e->elevator_data;                       \
4655         unsigned int __data;                                            \
4656         int ret = cfq_var_store(&__data, (page), count);                \
4657         if (__data < (MIN))                                             \
4658                 __data = (MIN);                                         \
4659         else if (__data > (MAX))                                        \
4660                 __data = (MAX);                                         \
4661         if (__CONV)                                                     \
4662                 *(__PTR) = msecs_to_jiffies(__data);                    \
4663         else                                                            \
4664                 *(__PTR) = __data;                                      \
4665         return ret;                                                     \
4666 }
4667 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4668 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4669                 UINT_MAX, 1);
4670 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4671                 UINT_MAX, 1);
4672 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4673 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4674                 UINT_MAX, 0);
4675 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4676 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4677 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4678 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4679 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4680                 UINT_MAX, 0);
4681 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4682 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4683 #undef STORE_FUNCTION
4684
4685 #define CFQ_ATTR(name) \
4686         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4687
4688 static struct elv_fs_entry cfq_attrs[] = {
4689         CFQ_ATTR(quantum),
4690         CFQ_ATTR(fifo_expire_sync),
4691         CFQ_ATTR(fifo_expire_async),
4692         CFQ_ATTR(back_seek_max),
4693         CFQ_ATTR(back_seek_penalty),
4694         CFQ_ATTR(slice_sync),
4695         CFQ_ATTR(slice_async),
4696         CFQ_ATTR(slice_async_rq),
4697         CFQ_ATTR(slice_idle),
4698         CFQ_ATTR(group_idle),
4699         CFQ_ATTR(low_latency),
4700         CFQ_ATTR(target_latency),
4701         __ATTR_NULL
4702 };
4703
4704 static struct elevator_type iosched_cfq = {
4705         .ops = {
4706                 .elevator_merge_fn =            cfq_merge,
4707                 .elevator_merged_fn =           cfq_merged_request,
4708                 .elevator_merge_req_fn =        cfq_merged_requests,
4709                 .elevator_allow_merge_fn =      cfq_allow_merge,
4710                 .elevator_bio_merged_fn =       cfq_bio_merged,
4711                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4712                 .elevator_add_req_fn =          cfq_insert_request,
4713                 .elevator_activate_req_fn =     cfq_activate_request,
4714                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4715                 .elevator_completed_req_fn =    cfq_completed_request,
4716                 .elevator_former_req_fn =       elv_rb_former_request,
4717                 .elevator_latter_req_fn =       elv_rb_latter_request,
4718                 .elevator_init_icq_fn =         cfq_init_icq,
4719                 .elevator_exit_icq_fn =         cfq_exit_icq,
4720                 .elevator_set_req_fn =          cfq_set_request,
4721                 .elevator_put_req_fn =          cfq_put_request,
4722                 .elevator_may_queue_fn =        cfq_may_queue,
4723                 .elevator_init_fn =             cfq_init_queue,
4724                 .elevator_exit_fn =             cfq_exit_queue,
4725                 .elevator_registered_fn =       cfq_registered_queue,
4726         },
4727         .icq_size       =       sizeof(struct cfq_io_cq),
4728         .icq_align      =       __alignof__(struct cfq_io_cq),
4729         .elevator_attrs =       cfq_attrs,
4730         .elevator_name  =       "cfq",
4731         .elevator_owner =       THIS_MODULE,
4732 };
4733
4734 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4735 static struct blkcg_policy blkcg_policy_cfq = {
4736         .dfl_cftypes            = cfq_blkcg_files,
4737         .legacy_cftypes         = cfq_blkcg_legacy_files,
4738
4739         .cpd_alloc_fn           = cfq_cpd_alloc,
4740         .cpd_init_fn            = cfq_cpd_init,
4741         .cpd_free_fn            = cfq_cpd_free,
4742         .cpd_bind_fn            = cfq_cpd_bind,
4743
4744         .pd_alloc_fn            = cfq_pd_alloc,
4745         .pd_init_fn             = cfq_pd_init,
4746         .pd_offline_fn          = cfq_pd_offline,
4747         .pd_free_fn             = cfq_pd_free,
4748         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4749 };
4750 #endif
4751
4752 static int __init cfq_init(void)
4753 {
4754         int ret;
4755
4756         /*
4757          * could be 0 on HZ < 1000 setups
4758          */
4759         if (!cfq_slice_async)
4760                 cfq_slice_async = 1;
4761         if (!cfq_slice_idle)
4762                 cfq_slice_idle = 1;
4763
4764 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4765         if (!cfq_group_idle)
4766                 cfq_group_idle = 1;
4767
4768         ret = blkcg_policy_register(&blkcg_policy_cfq);
4769         if (ret)
4770                 return ret;
4771 #else
4772         cfq_group_idle = 0;
4773 #endif
4774
4775         ret = -ENOMEM;
4776         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4777         if (!cfq_pool)
4778                 goto err_pol_unreg;
4779
4780         ret = elv_register(&iosched_cfq);
4781         if (ret)
4782                 goto err_free_pool;
4783
4784         return 0;
4785
4786 err_free_pool:
4787         kmem_cache_destroy(cfq_pool);
4788 err_pol_unreg:
4789 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4790         blkcg_policy_unregister(&blkcg_policy_cfq);
4791 #endif
4792         return ret;
4793 }
4794
4795 static void __exit cfq_exit(void)
4796 {
4797 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4798         blkcg_policy_unregister(&blkcg_policy_cfq);
4799 #endif
4800         elv_unregister(&iosched_cfq);
4801         kmem_cache_destroy(cfq_pool);
4802 }
4803
4804 module_init(cfq_init);
4805 module_exit(cfq_exit);
4806
4807 MODULE_AUTHOR("Jens Axboe");
4808 MODULE_LICENSE("GPL");
4809 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");