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