block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/jiffies.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16 #include <linux/blktrace_api.h>
  17 #include "blk.h"
  18 #include "blk-cgroup.h"
  19
  20 /*
  21  * tunables
  22  */
  23 /* max queue in one round of service */
  24 static const int cfq_quantum = 8;
  25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  26 /* maximum backwards seek, in KiB */
  27 static const int cfq_back_max = 16 * 1024;
  28 /* penalty of a backwards seek */
  29 static const int cfq_back_penalty = 2;
  30 static const int cfq_slice_sync = HZ / 10;
  31 static int cfq_slice_async = HZ / 25;
  32 static const int cfq_slice_async_rq = 2;
  33 static int cfq_slice_idle = HZ / 125;
  34 static int cfq_group_idle = HZ / 125;
  35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
  36 static const int cfq_hist_divisor = 4;
  37
  38 /*
  39  * offset from end of service tree
  40  */
  41 #define CFQ_IDLE_DELAY          (HZ / 5)
  42
  43 /*
  44  * below this threshold, we consider thinktime immediate
  45  */
  46 #define CFQ_MIN_TT              (2)
  47
  48 #define CFQ_SLICE_SCALE         (5)
  49 #define CFQ_HW_QUEUE_MIN        (5)
  50 #define CFQ_SERVICE_SHIFT       12
  51
  52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  56
  57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  60
  61 static struct kmem_cache *cfq_pool;
  62
  63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  69
  70 /* 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 pd;
 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
 310 struct cfq_io_cq {
 311         struct io_cq            icq;            /* must be the first member */
 312         struct cfq_queue        *cfqq[2];
 313         struct cfq_ttime        ttime;
 314         int                     ioprio;         /* the current ioprio */
 315 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 316         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
 317 #endif
 318 };
 319
 320 /*
 321  * Per block device queue structure
 322  */
 323 struct cfq_data {
 324         struct request_queue *queue;
 325         /* Root service tree for cfq_groups */
 326         struct cfq_rb_root grp_service_tree;
 327         struct cfq_group *root_group;
 328
 329         /*
 330          * The priority currently being served
 331          */
 332         enum wl_class_t serving_wl_class;
 333         enum wl_type_t serving_wl_type;
 334         unsigned long workload_expires;
 335         struct cfq_group *serving_group;
 336
 337         /*
 338          * Each priority tree is sorted by next_request position.  These
 339          * trees are used when determining if two or more queues are
 340          * interleaving requests (see cfq_close_cooperator).
 341          */
 342         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 343
 344         unsigned int busy_queues;
 345         unsigned int busy_sync_queues;
 346
 347         int rq_in_driver;
 348         int rq_in_flight[2];
 349
 350         /*
 351          * queue-depth detection
 352          */
 353         int rq_queued;
 354         int hw_tag;
 355         /*
 356          * hw_tag can be
 357          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 358          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 359          *  0 => no NCQ
 360          */
 361         int hw_tag_est_depth;
 362         unsigned int hw_tag_samples;
 363
 364         /*
 365          * idle window management
 366          */
 367         struct timer_list idle_slice_timer;
 368         struct work_struct unplug_work;
 369
 370         struct cfq_queue *active_queue;
 371         struct cfq_io_cq *active_cic;
 372
 373         /*
 374          * async queue for each priority case
 375          */
 376         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 377         struct cfq_queue *async_idle_cfqq;
 378
 379         sector_t last_position;
 380
 381         /*
 382          * tunables, see top of file
 383          */
 384         unsigned int cfq_quantum;
 385         unsigned int cfq_fifo_expire[2];
 386         unsigned int cfq_back_penalty;
 387         unsigned int cfq_back_max;
 388         unsigned int cfq_slice[2];
 389         unsigned int cfq_slice_async_rq;
 390         unsigned int cfq_slice_idle;
 391         unsigned int cfq_group_idle;
 392         unsigned int cfq_latency;
 393         unsigned int cfq_target_latency;
 394
 395         /*
 396          * Fallback dummy cfqq for extreme OOM conditions
 397          */
 398         struct cfq_queue oom_cfqq;
 399
 400         unsigned long last_delayed_sync;
 401 };
 402
 403 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 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, pd) : 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                                        gfp_t gfp_mask);
 888
 889 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 890 {
 891         /* cic->icq is the first member, %NULL will convert to %NULL */
 892         return container_of(icq, struct cfq_io_cq, icq);
 893 }
 894
 895 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 896                                                struct io_context *ioc)
 897 {
 898         if (ioc)
 899                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 900         return NULL;
 901 }
 902
 903 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 904 {
 905         return cic->cfqq[is_sync];
 906 }
 907
 908 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 909                                 bool is_sync)
 910 {
 911         cic->cfqq[is_sync] = cfqq;
 912 }
 913
 914 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 915 {
 916         return cic->icq.q->elevator->elevator_data;
 917 }
 918
 919 /*
 920  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 921  * set (in which case it could also be direct WRITE).
 922  */
 923 static inline bool cfq_bio_sync(struct bio *bio)
 924 {
 925         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 926 }
 927
 928 /*
 929  * scheduler run of queue, if there are requests pending and no one in the
 930  * driver that will restart queueing
 931  */
 932 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 933 {
 934         if (cfqd->busy_queues) {
 935                 cfq_log(cfqd, "schedule dispatch");
 936                 kblockd_schedule_work(&cfqd->unplug_work);
 937         }
 938 }
 939
 940 /*
 941  * Scale schedule slice based on io priority. Use the sync time slice only
 942  * if a queue is marked sync and has sync io queued. A sync queue with async
 943  * io only, should not get full sync slice length.
 944  */
 945 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 946                                  unsigned short prio)
 947 {
 948         const int base_slice = cfqd->cfq_slice[sync];
 949
 950         WARN_ON(prio >= IOPRIO_BE_NR);
 951
 952         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 953 }
 954
 955 static inline int
 956 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 957 {
 958         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 959 }
 960
 961 /**
 962  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 963  * @charge: disk time being charged
 964  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 965  *
 966  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 967  * scaling is inversely proportional.
 968  *
 969  * scaled = charge / vfraction
 970  *
 971  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 972  */
 973 static inline u64 cfqg_scale_charge(unsigned long charge,
 974                                     unsigned int vfraction)
 975 {
 976         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 977
 978         /* charge / vfraction */
 979         c <<= CFQ_SERVICE_SHIFT;
 980         do_div(c, vfraction);
 981         return c;
 982 }
 983
 984 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 985 {
 986         s64 delta = (s64)(vdisktime - min_vdisktime);
 987         if (delta > 0)
 988                 min_vdisktime = vdisktime;
 989
 990         return min_vdisktime;
 991 }
 992
 993 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 994 {
 995         s64 delta = (s64)(vdisktime - min_vdisktime);
 996         if (delta < 0)
 997                 min_vdisktime = vdisktime;
 998
 999         return min_vdisktime;
1000 }
1001
1002 static void update_min_vdisktime(struct cfq_rb_root *st)
1003 {
1004         struct cfq_group *cfqg;
1005
1006         if (st->left) {
1007                 cfqg = rb_entry_cfqg(st->left);
1008                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
1009                                                   cfqg->vdisktime);
1010         }
1011 }
1012
1013 /*
1014  * get averaged number of queues of RT/BE priority.
1015  * average is updated, with a formula that gives more weight to higher numbers,
1016  * to quickly follows sudden increases and decrease slowly
1017  */
1018
1019 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1020                                         struct cfq_group *cfqg, bool rt)
1021 {
1022         unsigned min_q, max_q;
1023         unsigned mult  = cfq_hist_divisor - 1;
1024         unsigned round = cfq_hist_divisor / 2;
1025         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1026
1027         min_q = min(cfqg->busy_queues_avg[rt], busy);
1028         max_q = max(cfqg->busy_queues_avg[rt], busy);
1029         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1030                 cfq_hist_divisor;
1031         return cfqg->busy_queues_avg[rt];
1032 }
1033
1034 static inline unsigned
1035 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1036 {
1037         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1038 }
1039
1040 static inline unsigned
1041 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1042 {
1043         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1044         if (cfqd->cfq_latency) {
1045                 /*
1046                  * interested queues (we consider only the ones with the same
1047                  * priority class in the cfq group)
1048                  */
1049                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1050                                                 cfq_class_rt(cfqq));
1051                 unsigned sync_slice = cfqd->cfq_slice[1];
1052                 unsigned expect_latency = sync_slice * iq;
1053                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1054
1055                 if (expect_latency > group_slice) {
1056                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1057                         /* scale low_slice according to IO priority
1058                          * and sync vs async */
1059                         unsigned low_slice =
1060                                 min(slice, base_low_slice * slice / sync_slice);
1061                         /* the adapted slice value is scaled to fit all iqs
1062                          * into the target latency */
1063                         slice = max(slice * group_slice / expect_latency,
1064                                     low_slice);
1065                 }
1066         }
1067         return slice;
1068 }
1069
1070 static inline void
1071 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1072 {
1073         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1074
1075         cfqq->slice_start = jiffies;
1076         cfqq->slice_end = jiffies + slice;
1077         cfqq->allocated_slice = slice;
1078         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1079 }
1080
1081 /*
1082  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1083  * isn't valid until the first request from the dispatch is activated
1084  * and the slice time set.
1085  */
1086 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1087 {
1088         if (cfq_cfqq_slice_new(cfqq))
1089                 return false;
1090         if (time_before(jiffies, cfqq->slice_end))
1091                 return false;
1092
1093         return true;
1094 }
1095
1096 /*
1097  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1098  * We choose the request that is closest to the head right now. Distance
1099  * behind the head is penalized and only allowed to a certain extent.
1100  */
1101 static struct request *
1102 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1103 {
1104         sector_t s1, s2, d1 = 0, d2 = 0;
1105         unsigned long back_max;
1106 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1107 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1108         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1109
1110         if (rq1 == NULL || rq1 == rq2)
1111                 return rq2;
1112         if (rq2 == NULL)
1113                 return rq1;
1114
1115         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1116                 return rq_is_sync(rq1) ? rq1 : rq2;
1117
1118         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1119                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1120
1121         s1 = blk_rq_pos(rq1);
1122         s2 = blk_rq_pos(rq2);
1123
1124         /*
1125          * by definition, 1KiB is 2 sectors
1126          */
1127         back_max = cfqd->cfq_back_max * 2;
1128
1129         /*
1130          * Strict one way elevator _except_ in the case where we allow
1131          * short backward seeks which are biased as twice the cost of a
1132          * similar forward seek.
1133          */
1134         if (s1 >= last)
1135                 d1 = s1 - last;
1136         else if (s1 + back_max >= last)
1137                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1138         else
1139                 wrap |= CFQ_RQ1_WRAP;
1140
1141         if (s2 >= last)
1142                 d2 = s2 - last;
1143         else if (s2 + back_max >= last)
1144                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1145         else
1146                 wrap |= CFQ_RQ2_WRAP;
1147
1148         /* Found required data */
1149
1150         /*
1151          * By doing switch() on the bit mask "wrap" we avoid having to
1152          * check two variables for all permutations: --> faster!
1153          */
1154         switch (wrap) {
1155         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1156                 if (d1 < d2)
1157                         return rq1;
1158                 else if (d2 < d1)
1159                         return rq2;
1160                 else {
1161                         if (s1 >= s2)
1162                                 return rq1;
1163                         else
1164                                 return rq2;
1165                 }
1166
1167         case CFQ_RQ2_WRAP:
1168                 return rq1;
1169         case CFQ_RQ1_WRAP:
1170                 return rq2;
1171         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1172         default:
1173                 /*
1174                  * Since both rqs are wrapped,
1175                  * start with the one that's further behind head
1176                  * (--> only *one* back seek required),
1177                  * since back seek takes more time than forward.
1178                  */
1179                 if (s1 <= s2)
1180                         return rq1;
1181                 else
1182                         return rq2;
1183         }
1184 }
1185
1186 /*
1187  * The below is leftmost cache rbtree addon
1188  */
1189 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1190 {
1191         /* Service tree is empty */
1192         if (!root->count)
1193                 return NULL;
1194
1195         if (!root->left)
1196                 root->left = rb_first(&root->rb);
1197
1198         if (root->left)
1199                 return rb_entry(root->left, struct cfq_queue, rb_node);
1200
1201         return NULL;
1202 }
1203
1204 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1205 {
1206         if (!root->left)
1207                 root->left = rb_first(&root->rb);
1208
1209         if (root->left)
1210                 return rb_entry_cfqg(root->left);
1211
1212         return NULL;
1213 }
1214
1215 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1216 {
1217         rb_erase(n, root);
1218         RB_CLEAR_NODE(n);
1219 }
1220
1221 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1222 {
1223         if (root->left == n)
1224                 root->left = NULL;
1225         rb_erase_init(n, &root->rb);
1226         --root->count;
1227 }
1228
1229 /*
1230  * would be nice to take fifo expire time into account as well
1231  */
1232 static struct request *
1233 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1234                   struct request *last)
1235 {
1236         struct rb_node *rbnext = rb_next(&last->rb_node);
1237         struct rb_node *rbprev = rb_prev(&last->rb_node);
1238         struct request *next = NULL, *prev = NULL;
1239
1240         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1241
1242         if (rbprev)
1243                 prev = rb_entry_rq(rbprev);
1244
1245         if (rbnext)
1246                 next = rb_entry_rq(rbnext);
1247         else {
1248                 rbnext = rb_first(&cfqq->sort_list);
1249                 if (rbnext && rbnext != &last->rb_node)
1250                         next = rb_entry_rq(rbnext);
1251         }
1252
1253         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1254 }
1255
1256 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1257                                       struct cfq_queue *cfqq)
1258 {
1259         /*
1260          * just an approximation, should be ok.
1261          */
1262         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1263                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1264 }
1265
1266 static inline s64
1267 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1268 {
1269         return cfqg->vdisktime - st->min_vdisktime;
1270 }
1271
1272 static void
1273 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1274 {
1275         struct rb_node **node = &st->rb.rb_node;
1276         struct rb_node *parent = NULL;
1277         struct cfq_group *__cfqg;
1278         s64 key = cfqg_key(st, cfqg);
1279         int left = 1;
1280
1281         while (*node != NULL) {
1282                 parent = *node;
1283                 __cfqg = rb_entry_cfqg(parent);
1284
1285                 if (key < cfqg_key(st, __cfqg))
1286                         node = &parent->rb_left;
1287                 else {
1288                         node = &parent->rb_right;
1289                         left = 0;
1290                 }
1291         }
1292
1293         if (left)
1294                 st->left = &cfqg->rb_node;
1295
1296         rb_link_node(&cfqg->rb_node, parent, node);
1297         rb_insert_color(&cfqg->rb_node, &st->rb);
1298 }
1299
1300 /*
1301  * This has to be called only on activation of cfqg
1302  */
1303 static void
1304 cfq_update_group_weight(struct cfq_group *cfqg)
1305 {
1306         if (cfqg->new_weight) {
1307                 cfqg->weight = cfqg->new_weight;
1308                 cfqg->new_weight = 0;
1309         }
1310 }
1311
1312 static void
1313 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1314 {
1315         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1316
1317         if (cfqg->new_leaf_weight) {
1318                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1319                 cfqg->new_leaf_weight = 0;
1320         }
1321 }
1322
1323 static void
1324 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1325 {
1326         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1327         struct cfq_group *pos = cfqg;
1328         struct cfq_group *parent;
1329         bool propagate;
1330
1331         /* add to the service tree */
1332         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1333
1334         /*
1335          * Update leaf_weight.  We cannot update weight at this point
1336          * because cfqg might already have been activated and is
1337          * contributing its current weight to the parent's child_weight.
1338          */
1339         cfq_update_group_leaf_weight(cfqg);
1340         __cfq_group_service_tree_add(st, cfqg);
1341
1342         /*
1343          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1344          * entitled to.  vfraction is calculated by walking the tree
1345          * towards the root calculating the fraction it has at each level.
1346          * The compounded ratio is how much vfraction @cfqg owns.
1347          *
1348          * Start with the proportion tasks in this cfqg has against active
1349          * children cfqgs - its leaf_weight against children_weight.
1350          */
1351         propagate = !pos->nr_active++;
1352         pos->children_weight += pos->leaf_weight;
1353         vfr = vfr * pos->leaf_weight / pos->children_weight;
1354
1355         /*
1356          * Compound ->weight walking up the tree.  Both activation and
1357          * vfraction calculation are done in the same loop.  Propagation
1358          * stops once an already activated node is met.  vfraction
1359          * calculation should always continue to the root.
1360          */
1361         while ((parent = cfqg_parent(pos))) {
1362                 if (propagate) {
1363                         cfq_update_group_weight(pos);
1364                         propagate = !parent->nr_active++;
1365                         parent->children_weight += pos->weight;
1366                 }
1367                 vfr = vfr * pos->weight / parent->children_weight;
1368                 pos = parent;
1369         }
1370
1371         cfqg->vfraction = max_t(unsigned, vfr, 1);
1372 }
1373
1374 static void
1375 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1376 {
1377         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1378         struct cfq_group *__cfqg;
1379         struct rb_node *n;
1380
1381         cfqg->nr_cfqq++;
1382         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1383                 return;
1384
1385         /*
1386          * Currently put the group at the end. Later implement something
1387          * so that groups get lesser vtime based on their weights, so that
1388          * if group does not loose all if it was not continuously backlogged.
1389          */
1390         n = rb_last(&st->rb);
1391         if (n) {
1392                 __cfqg = rb_entry_cfqg(n);
1393                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1394         } else
1395                 cfqg->vdisktime = st->min_vdisktime;
1396         cfq_group_service_tree_add(st, cfqg);
1397 }
1398
1399 static void
1400 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1401 {
1402         struct cfq_group *pos = cfqg;
1403         bool propagate;
1404
1405         /*
1406          * Undo activation from cfq_group_service_tree_add().  Deactivate
1407          * @cfqg and propagate deactivation upwards.
1408          */
1409         propagate = !--pos->nr_active;
1410         pos->children_weight -= pos->leaf_weight;
1411
1412         while (propagate) {
1413                 struct cfq_group *parent = cfqg_parent(pos);
1414
1415                 /* @pos has 0 nr_active at this point */
1416                 WARN_ON_ONCE(pos->children_weight);
1417                 pos->vfraction = 0;
1418
1419                 if (!parent)
1420                         break;
1421
1422                 propagate = !--parent->nr_active;
1423                 parent->children_weight -= pos->weight;
1424                 pos = parent;
1425         }
1426
1427         /* remove from the service tree */
1428         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1429                 cfq_rb_erase(&cfqg->rb_node, st);
1430 }
1431
1432 static void
1433 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1434 {
1435         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1436
1437         BUG_ON(cfqg->nr_cfqq < 1);
1438         cfqg->nr_cfqq--;
1439
1440         /* If there are other cfq queues under this group, don't delete it */
1441         if (cfqg->nr_cfqq)
1442                 return;
1443
1444         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1445         cfq_group_service_tree_del(st, cfqg);
1446         cfqg->saved_wl_slice = 0;
1447         cfqg_stats_update_dequeue(cfqg);
1448 }
1449
1450 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1451                                                 unsigned int *unaccounted_time)
1452 {
1453         unsigned int slice_used;
1454
1455         /*
1456          * Queue got expired before even a single request completed or
1457          * got expired immediately after first request completion.
1458          */
1459         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1460                 /*
1461                  * Also charge the seek time incurred to the group, otherwise
1462                  * if there are mutiple queues in the group, each can dispatch
1463                  * a single request on seeky media and cause lots of seek time
1464                  * and group will never know it.
1465                  */
1466                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1467                                         1);
1468         } else {
1469                 slice_used = jiffies - cfqq->slice_start;
1470                 if (slice_used > cfqq->allocated_slice) {
1471                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1472                         slice_used = cfqq->allocated_slice;
1473                 }
1474                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1475                         *unaccounted_time += cfqq->slice_start -
1476                                         cfqq->dispatch_start;
1477         }
1478
1479         return slice_used;
1480 }
1481
1482 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1483                                 struct cfq_queue *cfqq)
1484 {
1485         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1486         unsigned int used_sl, charge, unaccounted_sl = 0;
1487         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1488                         - cfqg->service_tree_idle.count;
1489         unsigned int vfr;
1490
1491         BUG_ON(nr_sync < 0);
1492         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1493
1494         if (iops_mode(cfqd))
1495                 charge = cfqq->slice_dispatch;
1496         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1497                 charge = cfqq->allocated_slice;
1498
1499         /*
1500          * Can't update vdisktime while on service tree and cfqg->vfraction
1501          * is valid only while on it.  Cache vfr, leave the service tree,
1502          * update vdisktime and go back on.  The re-addition to the tree
1503          * will also update the weights as necessary.
1504          */
1505         vfr = cfqg->vfraction;
1506         cfq_group_service_tree_del(st, cfqg);
1507         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1508         cfq_group_service_tree_add(st, cfqg);
1509
1510         /* This group is being expired. Save the context */
1511         if (time_after(cfqd->workload_expires, jiffies)) {
1512                 cfqg->saved_wl_slice = cfqd->workload_expires
1513                                                 - jiffies;
1514                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1515                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1516         } else
1517                 cfqg->saved_wl_slice = 0;
1518
1519         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1520                                         st->min_vdisktime);
1521         cfq_log_cfqq(cfqq->cfqd, cfqq,
1522                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1523                      used_sl, cfqq->slice_dispatch, charge,
1524                      iops_mode(cfqd), cfqq->nr_sectors);
1525         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1526         cfqg_stats_set_start_empty_time(cfqg);
1527 }
1528
1529 /**
1530  * cfq_init_cfqg_base - initialize base part of a cfq_group
1531  * @cfqg: cfq_group to initialize
1532  *
1533  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1534  * is enabled or not.
1535  */
1536 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1537 {
1538         struct cfq_rb_root *st;
1539         int i, j;
1540
1541         for_each_cfqg_st(cfqg, i, j, st)
1542                 *st = CFQ_RB_ROOT;
1543         RB_CLEAR_NODE(&cfqg->rb_node);
1544
1545         cfqg->ttime.last_end_request = jiffies;
1546 }
1547
1548 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1549 static void cfqg_stats_init(struct cfqg_stats *stats)
1550 {
1551         blkg_rwstat_init(&stats->service_bytes);
1552         blkg_rwstat_init(&stats->serviced);
1553         blkg_rwstat_init(&stats->merged);
1554         blkg_rwstat_init(&stats->service_time);
1555         blkg_rwstat_init(&stats->wait_time);
1556         blkg_rwstat_init(&stats->queued);
1557
1558         blkg_stat_init(&stats->sectors);
1559         blkg_stat_init(&stats->time);
1560
1561 #ifdef CONFIG_DEBUG_BLK_CGROUP
1562         blkg_stat_init(&stats->unaccounted_time);
1563         blkg_stat_init(&stats->avg_queue_size_sum);
1564         blkg_stat_init(&stats->avg_queue_size_samples);
1565         blkg_stat_init(&stats->dequeue);
1566         blkg_stat_init(&stats->group_wait_time);
1567         blkg_stat_init(&stats->idle_time);
1568         blkg_stat_init(&stats->empty_time);
1569 #endif
1570 }
1571
1572 static void cfq_cpd_init(const struct blkcg *blkcg)
1573 {
1574         struct cfq_group_data *cgd =
1575                 cpd_to_cfqgd(blkcg->pd[blkcg_policy_cfq.plid]);
1576
1577         if (blkcg == &blkcg_root) {
1578                 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1579                 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1580         } else {
1581                 cgd->weight = CFQ_WEIGHT_DEFAULT;
1582                 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1583         }
1584 }
1585
1586 static void cfq_pd_init(struct blkcg_gq *blkg)
1587 {
1588         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1589         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkg->blkcg);
1590
1591         cfq_init_cfqg_base(cfqg);
1592         cfqg->weight = cgd->weight;
1593         cfqg->leaf_weight = cgd->leaf_weight;
1594         cfqg_stats_init(&cfqg->stats);
1595         cfqg_stats_init(&cfqg->dead_stats);
1596 }
1597
1598 static void cfq_pd_offline(struct blkcg_gq *blkg)
1599 {
1600         /*
1601          * @blkg is going offline and will be ignored by
1602          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1603          * that they don't get lost.  If IOs complete after this point, the
1604          * stats for them will be lost.  Oh well...
1605          */
1606         cfqg_stats_xfer_dead(blkg_to_cfqg(blkg));
1607 }
1608
1609 /* offset delta from cfqg->stats to cfqg->dead_stats */
1610 static const int dead_stats_off_delta = offsetof(struct cfq_group, dead_stats) -
1611                                         offsetof(struct cfq_group, stats);
1612
1613 /* to be used by recursive prfill, sums live and dead stats recursively */
1614 static u64 cfqg_stat_pd_recursive_sum(struct blkg_policy_data *pd, int off)
1615 {
1616         u64 sum = 0;
1617
1618         sum += blkg_stat_recursive_sum(pd, off);
1619         sum += blkg_stat_recursive_sum(pd, off + dead_stats_off_delta);
1620         return sum;
1621 }
1622
1623 /* to be used by recursive prfill, sums live and dead rwstats recursively */
1624 static struct blkg_rwstat cfqg_rwstat_pd_recursive_sum(struct blkg_policy_data *pd,
1625                                                        int off)
1626 {
1627         struct blkg_rwstat a, b;
1628
1629         a = blkg_rwstat_recursive_sum(pd, off);
1630         b = blkg_rwstat_recursive_sum(pd, off + dead_stats_off_delta);
1631         blkg_rwstat_merge(&a, &b);
1632         return a;
1633 }
1634
1635 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
1636 {
1637         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1638
1639         cfqg_stats_reset(&cfqg->stats);
1640         cfqg_stats_reset(&cfqg->dead_stats);
1641 }
1642
1643 /*
1644  * Search for the cfq group current task belongs to. request_queue lock must
1645  * be held.
1646  */
1647 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1648                                                 struct blkcg *blkcg)
1649 {
1650         struct request_queue *q = cfqd->queue;
1651         struct cfq_group *cfqg = NULL;
1652
1653         /* avoid lookup for the common case where there's no blkcg */
1654         if (blkcg == &blkcg_root) {
1655                 cfqg = cfqd->root_group;
1656         } else {
1657                 struct blkcg_gq *blkg;
1658
1659                 blkg = blkg_lookup_create(blkcg, q);
1660                 if (!IS_ERR(blkg))
1661                         cfqg = blkg_to_cfqg(blkg);
1662         }
1663
1664         return cfqg;
1665 }
1666
1667 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1668 {
1669         /* Currently, all async queues are mapped to root group */
1670         if (!cfq_cfqq_sync(cfqq))
1671                 cfqg = cfqq->cfqd->root_group;
1672
1673         cfqq->cfqg = cfqg;
1674         /* cfqq reference on cfqg */
1675         cfqg_get(cfqg);
1676 }
1677
1678 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1679                                      struct blkg_policy_data *pd, int off)
1680 {
1681         struct cfq_group *cfqg = pd_to_cfqg(pd);
1682
1683         if (!cfqg->dev_weight)
1684                 return 0;
1685         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1686 }
1687
1688 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1689 {
1690         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1691                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1692                           0, false);
1693         return 0;
1694 }
1695
1696 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1697                                           struct blkg_policy_data *pd, int off)
1698 {
1699         struct cfq_group *cfqg = pd_to_cfqg(pd);
1700
1701         if (!cfqg->dev_leaf_weight)
1702                 return 0;
1703         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1704 }
1705
1706 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1707 {
1708         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1709                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1710                           0, false);
1711         return 0;
1712 }
1713
1714 static int cfq_print_weight(struct seq_file *sf, void *v)
1715 {
1716         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1717         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1718         unsigned int val = 0;
1719
1720         if (cgd)
1721                 val = cgd->weight;
1722
1723         seq_printf(sf, "%u\n", val);
1724         return 0;
1725 }
1726
1727 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1728 {
1729         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1730         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1731         unsigned int val = 0;
1732
1733         if (cgd)
1734                 val = cgd->leaf_weight;
1735
1736         seq_printf(sf, "%u\n", val);
1737         return 0;
1738 }
1739
1740 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1741                                         char *buf, size_t nbytes, loff_t off,
1742                                         bool is_leaf_weight)
1743 {
1744         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1745         struct blkg_conf_ctx ctx;
1746         struct cfq_group *cfqg;
1747         struct cfq_group_data *cfqgd;
1748         int ret;
1749
1750         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1751         if (ret)
1752                 return ret;
1753
1754         ret = -EINVAL;
1755         cfqg = blkg_to_cfqg(ctx.blkg);
1756         cfqgd = blkcg_to_cfqgd(blkcg);
1757         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1758                 if (!is_leaf_weight) {
1759                         cfqg->dev_weight = ctx.v;
1760                         cfqg->new_weight = ctx.v ?: cfqgd->weight;
1761                 } else {
1762                         cfqg->dev_leaf_weight = ctx.v;
1763                         cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1764                 }
1765                 ret = 0;
1766         }
1767
1768         blkg_conf_finish(&ctx);
1769         return ret ?: nbytes;
1770 }
1771
1772 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1773                                       char *buf, size_t nbytes, loff_t off)
1774 {
1775         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1776 }
1777
1778 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1779                                            char *buf, size_t nbytes, loff_t off)
1780 {
1781         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1782 }
1783
1784 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1785                             u64 val, bool is_leaf_weight)
1786 {
1787         struct blkcg *blkcg = css_to_blkcg(css);
1788         struct blkcg_gq *blkg;
1789         struct cfq_group_data *cfqgd;
1790
1791         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1792                 return -EINVAL;
1793
1794         spin_lock_irq(&blkcg->lock);
1795         cfqgd = blkcg_to_cfqgd(blkcg);
1796
1797         if (!is_leaf_weight)
1798                 cfqgd->weight = val;
1799         else
1800                 cfqgd->leaf_weight = val;
1801
1802         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1803                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1804
1805                 if (!cfqg)
1806                         continue;
1807
1808                 if (!is_leaf_weight) {
1809                         if (!cfqg->dev_weight)
1810                                 cfqg->new_weight = cfqgd->weight;
1811                 } else {
1812                         if (!cfqg->dev_leaf_weight)
1813                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1814                 }
1815         }
1816
1817         spin_unlock_irq(&blkcg->lock);
1818         return 0;
1819 }
1820
1821 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1822                           u64 val)
1823 {
1824         return __cfq_set_weight(css, cft, val, false);
1825 }
1826
1827 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1828                                struct cftype *cft, u64 val)
1829 {
1830         return __cfq_set_weight(css, cft, val, true);
1831 }
1832
1833 static int cfqg_print_stat(struct seq_file *sf, void *v)
1834 {
1835         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1836                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1837         return 0;
1838 }
1839
1840 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1841 {
1842         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1843                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1844         return 0;
1845 }
1846
1847 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1848                                       struct blkg_policy_data *pd, int off)
1849 {
1850         u64 sum = cfqg_stat_pd_recursive_sum(pd, off);
1851
1852         return __blkg_prfill_u64(sf, pd, sum);
1853 }
1854
1855 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1856                                         struct blkg_policy_data *pd, int off)
1857 {
1858         struct blkg_rwstat sum = cfqg_rwstat_pd_recursive_sum(pd, off);
1859
1860         return __blkg_prfill_rwstat(sf, pd, &sum);
1861 }
1862
1863 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1864 {
1865         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1866                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1867                           seq_cft(sf)->private, false);
1868         return 0;
1869 }
1870
1871 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1872 {
1873         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1874                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1875                           seq_cft(sf)->private, true);
1876         return 0;
1877 }
1878
1879 #ifdef CONFIG_DEBUG_BLK_CGROUP
1880 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1881                                       struct blkg_policy_data *pd, int off)
1882 {
1883         struct cfq_group *cfqg = pd_to_cfqg(pd);
1884         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1885         u64 v = 0;
1886
1887         if (samples) {
1888                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1889                 v = div64_u64(v, samples);
1890         }
1891         __blkg_prfill_u64(sf, pd, v);
1892         return 0;
1893 }
1894
1895 /* print avg_queue_size */
1896 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1897 {
1898         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1899                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1900                           0, false);
1901         return 0;
1902 }
1903 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1904
1905 static struct cftype cfq_blkcg_files[] = {
1906         /* on root, weight is mapped to leaf_weight */
1907         {
1908                 .name = "weight_device",
1909                 .flags = CFTYPE_ONLY_ON_ROOT,
1910                 .seq_show = cfqg_print_leaf_weight_device,
1911                 .write = cfqg_set_leaf_weight_device,
1912         },
1913         {
1914                 .name = "weight",
1915                 .flags = CFTYPE_ONLY_ON_ROOT,
1916                 .seq_show = cfq_print_leaf_weight,
1917                 .write_u64 = cfq_set_leaf_weight,
1918         },
1919
1920         /* no such mapping necessary for !roots */
1921         {
1922                 .name = "weight_device",
1923                 .flags = CFTYPE_NOT_ON_ROOT,
1924                 .seq_show = cfqg_print_weight_device,
1925                 .write = cfqg_set_weight_device,
1926         },
1927         {
1928                 .name = "weight",
1929                 .flags = CFTYPE_NOT_ON_ROOT,
1930                 .seq_show = cfq_print_weight,
1931                 .write_u64 = cfq_set_weight,
1932         },
1933
1934         {
1935                 .name = "leaf_weight_device",
1936                 .seq_show = cfqg_print_leaf_weight_device,
1937                 .write = cfqg_set_leaf_weight_device,
1938         },
1939         {
1940                 .name = "leaf_weight",
1941                 .seq_show = cfq_print_leaf_weight,
1942                 .write_u64 = cfq_set_leaf_weight,
1943         },
1944
1945         /* statistics, covers only the tasks in the cfqg */
1946         {
1947                 .name = "time",
1948                 .private = offsetof(struct cfq_group, stats.time),
1949                 .seq_show = cfqg_print_stat,
1950         },
1951         {
1952                 .name = "sectors",
1953                 .private = offsetof(struct cfq_group, stats.sectors),
1954                 .seq_show = cfqg_print_stat,
1955         },
1956         {
1957                 .name = "io_service_bytes",
1958                 .private = offsetof(struct cfq_group, stats.service_bytes),
1959                 .seq_show = cfqg_print_rwstat,
1960         },
1961         {
1962                 .name = "io_serviced",
1963                 .private = offsetof(struct cfq_group, stats.serviced),
1964                 .seq_show = cfqg_print_rwstat,
1965         },
1966         {
1967                 .name = "io_service_time",
1968                 .private = offsetof(struct cfq_group, stats.service_time),
1969                 .seq_show = cfqg_print_rwstat,
1970         },
1971         {
1972                 .name = "io_wait_time",
1973                 .private = offsetof(struct cfq_group, stats.wait_time),
1974                 .seq_show = cfqg_print_rwstat,
1975         },
1976         {
1977                 .name = "io_merged",
1978                 .private = offsetof(struct cfq_group, stats.merged),
1979                 .seq_show = cfqg_print_rwstat,
1980         },
1981         {
1982                 .name = "io_queued",
1983                 .private = offsetof(struct cfq_group, stats.queued),
1984                 .seq_show = cfqg_print_rwstat,
1985         },
1986
1987         /* the same statictics which cover the cfqg and its descendants */
1988         {
1989                 .name = "time_recursive",
1990                 .private = offsetof(struct cfq_group, stats.time),
1991                 .seq_show = cfqg_print_stat_recursive,
1992         },
1993         {
1994                 .name = "sectors_recursive",
1995                 .private = offsetof(struct cfq_group, stats.sectors),
1996                 .seq_show = cfqg_print_stat_recursive,
1997         },
1998         {
1999                 .name = "io_service_bytes_recursive",
2000                 .private = offsetof(struct cfq_group, stats.service_bytes),
2001                 .seq_show = cfqg_print_rwstat_recursive,
2002         },
2003         {
2004                 .name = "io_serviced_recursive",
2005                 .private = offsetof(struct cfq_group, stats.serviced),
2006                 .seq_show = cfqg_print_rwstat_recursive,
2007         },
2008         {
2009                 .name = "io_service_time_recursive",
2010                 .private = offsetof(struct cfq_group, stats.service_time),
2011                 .seq_show = cfqg_print_rwstat_recursive,
2012         },
2013         {
2014                 .name = "io_wait_time_recursive",
2015                 .private = offsetof(struct cfq_group, stats.wait_time),
2016                 .seq_show = cfqg_print_rwstat_recursive,
2017         },
2018         {
2019                 .name = "io_merged_recursive",
2020                 .private = offsetof(struct cfq_group, stats.merged),
2021                 .seq_show = cfqg_print_rwstat_recursive,
2022         },
2023         {
2024                 .name = "io_queued_recursive",
2025                 .private = offsetof(struct cfq_group, stats.queued),
2026                 .seq_show = cfqg_print_rwstat_recursive,
2027         },
2028 #ifdef CONFIG_DEBUG_BLK_CGROUP
2029         {
2030                 .name = "avg_queue_size",
2031                 .seq_show = cfqg_print_avg_queue_size,
2032         },
2033         {
2034                 .name = "group_wait_time",
2035                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2036                 .seq_show = cfqg_print_stat,
2037         },
2038         {
2039                 .name = "idle_time",
2040                 .private = offsetof(struct cfq_group, stats.idle_time),
2041                 .seq_show = cfqg_print_stat,
2042         },
2043         {
2044                 .name = "empty_time",
2045                 .private = offsetof(struct cfq_group, stats.empty_time),
2046                 .seq_show = cfqg_print_stat,
2047         },
2048         {
2049                 .name = "dequeue",
2050                 .private = offsetof(struct cfq_group, stats.dequeue),
2051                 .seq_show = cfqg_print_stat,
2052         },
2053         {
2054                 .name = "unaccounted_time",
2055                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2056                 .seq_show = cfqg_print_stat,
2057         },
2058 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2059         { }     /* terminate */
2060 };
2061 #else /* GROUP_IOSCHED */
2062 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
2063                                                 struct blkcg *blkcg)
2064 {
2065         return cfqd->root_group;
2066 }
2067
2068 static inline void
2069 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2070         cfqq->cfqg = cfqg;
2071 }
2072
2073 #endif /* GROUP_IOSCHED */
2074
2075 /*
2076  * The cfqd->service_trees holds all pending cfq_queue's that have
2077  * requests waiting to be processed. It is sorted in the order that
2078  * we will service the queues.
2079  */
2080 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2081                                  bool add_front)
2082 {
2083         struct rb_node **p, *parent;
2084         struct cfq_queue *__cfqq;
2085         unsigned long rb_key;
2086         struct cfq_rb_root *st;
2087         int left;
2088         int new_cfqq = 1;
2089
2090         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2091         if (cfq_class_idle(cfqq)) {
2092                 rb_key = CFQ_IDLE_DELAY;
2093                 parent = rb_last(&st->rb);
2094                 if (parent && parent != &cfqq->rb_node) {
2095                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2096                         rb_key += __cfqq->rb_key;
2097                 } else
2098                         rb_key += jiffies;
2099         } else if (!add_front) {
2100                 /*
2101                  * Get our rb key offset. Subtract any residual slice
2102                  * value carried from last service. A negative resid
2103                  * count indicates slice overrun, and this should position
2104                  * the next service time further away in the tree.
2105                  */
2106                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2107                 rb_key -= cfqq->slice_resid;
2108                 cfqq->slice_resid = 0;
2109         } else {
2110                 rb_key = -HZ;
2111                 __cfqq = cfq_rb_first(st);
2112                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2113         }
2114
2115         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2116                 new_cfqq = 0;
2117                 /*
2118                  * same position, nothing more to do
2119                  */
2120                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2121                         return;
2122
2123                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2124                 cfqq->service_tree = NULL;
2125         }
2126
2127         left = 1;
2128         parent = NULL;
2129         cfqq->service_tree = st;
2130         p = &st->rb.rb_node;
2131         while (*p) {
2132                 parent = *p;
2133                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2134
2135                 /*
2136                  * sort by key, that represents service time.
2137                  */
2138                 if (time_before(rb_key, __cfqq->rb_key))
2139                         p = &parent->rb_left;
2140                 else {
2141                         p = &parent->rb_right;
2142                         left = 0;
2143                 }
2144         }
2145
2146         if (left)
2147                 st->left = &cfqq->rb_node;
2148
2149         cfqq->rb_key = rb_key;
2150         rb_link_node(&cfqq->rb_node, parent, p);
2151         rb_insert_color(&cfqq->rb_node, &st->rb);
2152         st->count++;
2153         if (add_front || !new_cfqq)
2154                 return;
2155         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2156 }
2157
2158 static struct cfq_queue *
2159 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2160                      sector_t sector, struct rb_node **ret_parent,
2161                      struct rb_node ***rb_link)
2162 {
2163         struct rb_node **p, *parent;
2164         struct cfq_queue *cfqq = NULL;
2165
2166         parent = NULL;
2167         p = &root->rb_node;
2168         while (*p) {
2169                 struct rb_node **n;
2170
2171                 parent = *p;
2172                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2173
2174                 /*
2175                  * Sort strictly based on sector.  Smallest to the left,
2176                  * largest to the right.
2177                  */
2178                 if (sector > blk_rq_pos(cfqq->next_rq))
2179                         n = &(*p)->rb_right;
2180                 else if (sector < blk_rq_pos(cfqq->next_rq))
2181                         n = &(*p)->rb_left;
2182                 else
2183                         break;
2184                 p = n;
2185                 cfqq = NULL;
2186         }
2187
2188         *ret_parent = parent;
2189         if (rb_link)
2190                 *rb_link = p;
2191         return cfqq;
2192 }
2193
2194 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2195 {
2196         struct rb_node **p, *parent;
2197         struct cfq_queue *__cfqq;
2198
2199         if (cfqq->p_root) {
2200                 rb_erase(&cfqq->p_node, cfqq->p_root);
2201                 cfqq->p_root = NULL;
2202         }
2203
2204         if (cfq_class_idle(cfqq))
2205                 return;
2206         if (!cfqq->next_rq)
2207                 return;
2208
2209         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2210         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2211                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2212         if (!__cfqq) {
2213                 rb_link_node(&cfqq->p_node, parent, p);
2214                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2215         } else
2216                 cfqq->p_root = NULL;
2217 }
2218
2219 /*
2220  * Update cfqq's position in the service tree.
2221  */
2222 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2223 {
2224         /*
2225          * Resorting requires the cfqq to be on the RR list already.
2226          */
2227         if (cfq_cfqq_on_rr(cfqq)) {
2228                 cfq_service_tree_add(cfqd, cfqq, 0);
2229                 cfq_prio_tree_add(cfqd, cfqq);
2230         }
2231 }
2232
2233 /*
2234  * add to busy list of queues for service, trying to be fair in ordering
2235  * the pending list according to last request service
2236  */
2237 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2238 {
2239         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2240         BUG_ON(cfq_cfqq_on_rr(cfqq));
2241         cfq_mark_cfqq_on_rr(cfqq);
2242         cfqd->busy_queues++;
2243         if (cfq_cfqq_sync(cfqq))
2244                 cfqd->busy_sync_queues++;
2245
2246         cfq_resort_rr_list(cfqd, cfqq);
2247 }
2248
2249 /*
2250  * Called when the cfqq no longer has requests pending, remove it from
2251  * the service tree.
2252  */
2253 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2254 {
2255         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2256         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2257         cfq_clear_cfqq_on_rr(cfqq);
2258
2259         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2260                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2261                 cfqq->service_tree = NULL;
2262         }
2263         if (cfqq->p_root) {
2264                 rb_erase(&cfqq->p_node, cfqq->p_root);
2265                 cfqq->p_root = NULL;
2266         }
2267
2268         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2269         BUG_ON(!cfqd->busy_queues);
2270         cfqd->busy_queues--;
2271         if (cfq_cfqq_sync(cfqq))
2272                 cfqd->busy_sync_queues--;
2273 }
2274
2275 /*
2276  * rb tree support functions
2277  */
2278 static void cfq_del_rq_rb(struct request *rq)
2279 {
2280         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2281         const int sync = rq_is_sync(rq);
2282
2283         BUG_ON(!cfqq->queued[sync]);
2284         cfqq->queued[sync]--;
2285
2286         elv_rb_del(&cfqq->sort_list, rq);
2287
2288         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2289                 /*
2290                  * Queue will be deleted from service tree when we actually
2291                  * expire it later. Right now just remove it from prio tree
2292                  * as it is empty.
2293                  */
2294                 if (cfqq->p_root) {
2295                         rb_erase(&cfqq->p_node, cfqq->p_root);
2296                         cfqq->p_root = NULL;
2297                 }
2298         }
2299 }
2300
2301 static void cfq_add_rq_rb(struct request *rq)
2302 {
2303         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2304         struct cfq_data *cfqd = cfqq->cfqd;
2305         struct request *prev;
2306
2307         cfqq->queued[rq_is_sync(rq)]++;
2308
2309         elv_rb_add(&cfqq->sort_list, rq);
2310
2311         if (!cfq_cfqq_on_rr(cfqq))
2312                 cfq_add_cfqq_rr(cfqd, cfqq);
2313
2314         /*
2315          * check if this request is a better next-serve candidate
2316          */
2317         prev = cfqq->next_rq;
2318         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2319
2320         /*
2321          * adjust priority tree position, if ->next_rq changes
2322          */
2323         if (prev != cfqq->next_rq)
2324                 cfq_prio_tree_add(cfqd, cfqq);
2325
2326         BUG_ON(!cfqq->next_rq);
2327 }
2328
2329 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2330 {
2331         elv_rb_del(&cfqq->sort_list, rq);
2332         cfqq->queued[rq_is_sync(rq)]--;
2333         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2334         cfq_add_rq_rb(rq);
2335         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2336                                  rq->cmd_flags);
2337 }
2338
2339 static struct request *
2340 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2341 {
2342         struct task_struct *tsk = current;
2343         struct cfq_io_cq *cic;
2344         struct cfq_queue *cfqq;
2345
2346         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2347         if (!cic)
2348                 return NULL;
2349
2350         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2351         if (cfqq)
2352                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2353
2354         return NULL;
2355 }
2356
2357 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2358 {
2359         struct cfq_data *cfqd = q->elevator->elevator_data;
2360
2361         cfqd->rq_in_driver++;
2362         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2363                                                 cfqd->rq_in_driver);
2364
2365         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2366 }
2367
2368 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2369 {
2370         struct cfq_data *cfqd = q->elevator->elevator_data;
2371
2372         WARN_ON(!cfqd->rq_in_driver);
2373         cfqd->rq_in_driver--;
2374         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2375                                                 cfqd->rq_in_driver);
2376 }
2377
2378 static void cfq_remove_request(struct request *rq)
2379 {
2380         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2381
2382         if (cfqq->next_rq == rq)
2383                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2384
2385         list_del_init(&rq->queuelist);
2386         cfq_del_rq_rb(rq);
2387
2388         cfqq->cfqd->rq_queued--;
2389         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2390         if (rq->cmd_flags & REQ_PRIO) {
2391                 WARN_ON(!cfqq->prio_pending);
2392                 cfqq->prio_pending--;
2393         }
2394 }
2395
2396 static int cfq_merge(struct request_queue *q, struct request **req,
2397                      struct bio *bio)
2398 {
2399         struct cfq_data *cfqd = q->elevator->elevator_data;
2400         struct request *__rq;
2401
2402         __rq = cfq_find_rq_fmerge(cfqd, bio);
2403         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2404                 *req = __rq;
2405                 return ELEVATOR_FRONT_MERGE;
2406         }
2407
2408         return ELEVATOR_NO_MERGE;
2409 }
2410
2411 static void cfq_merged_request(struct request_queue *q, struct request *req,
2412                                int type)
2413 {
2414         if (type == ELEVATOR_FRONT_MERGE) {
2415                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2416
2417                 cfq_reposition_rq_rb(cfqq, req);
2418         }
2419 }
2420
2421 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2422                                 struct bio *bio)
2423 {
2424         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2425 }
2426
2427 static void
2428 cfq_merged_requests(struct request_queue *q, struct request *rq,
2429                     struct request *next)
2430 {
2431         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2432         struct cfq_data *cfqd = q->elevator->elevator_data;
2433
2434         /*
2435          * reposition in fifo if next is older than rq
2436          */
2437         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2438             time_before(next->fifo_time, rq->fifo_time) &&
2439             cfqq == RQ_CFQQ(next)) {
2440                 list_move(&rq->queuelist, &next->queuelist);
2441                 rq->fifo_time = next->fifo_time;
2442         }
2443
2444         if (cfqq->next_rq == next)
2445                 cfqq->next_rq = rq;
2446         cfq_remove_request(next);
2447         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2448
2449         cfqq = RQ_CFQQ(next);
2450         /*
2451          * all requests of this queue are merged to other queues, delete it
2452          * from the service tree. If it's the active_queue,
2453          * cfq_dispatch_requests() will choose to expire it or do idle
2454          */
2455         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2456             cfqq != cfqd->active_queue)
2457                 cfq_del_cfqq_rr(cfqd, cfqq);
2458 }
2459
2460 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2461                            struct bio *bio)
2462 {
2463         struct cfq_data *cfqd = q->elevator->elevator_data;
2464         struct cfq_io_cq *cic;
2465         struct cfq_queue *cfqq;
2466
2467         /*
2468          * Disallow merge of a sync bio into an async request.
2469          */
2470         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2471                 return false;
2472
2473         /*
2474          * Lookup the cfqq that this bio will be queued with and allow
2475          * merge only if rq is queued there.
2476          */
2477         cic = cfq_cic_lookup(cfqd, current->io_context);
2478         if (!cic)
2479                 return false;
2480
2481         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2482         return cfqq == RQ_CFQQ(rq);
2483 }
2484
2485 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2486 {
2487         del_timer(&cfqd->idle_slice_timer);
2488         cfqg_stats_update_idle_time(cfqq->cfqg);
2489 }
2490
2491 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2492                                    struct cfq_queue *cfqq)
2493 {
2494         if (cfqq) {
2495                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2496                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2497                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2498                 cfqq->slice_start = 0;
2499                 cfqq->dispatch_start = jiffies;
2500                 cfqq->allocated_slice = 0;
2501                 cfqq->slice_end = 0;
2502                 cfqq->slice_dispatch = 0;
2503                 cfqq->nr_sectors = 0;
2504
2505                 cfq_clear_cfqq_wait_request(cfqq);
2506                 cfq_clear_cfqq_must_dispatch(cfqq);
2507                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2508                 cfq_clear_cfqq_fifo_expire(cfqq);
2509                 cfq_mark_cfqq_slice_new(cfqq);
2510
2511                 cfq_del_timer(cfqd, cfqq);
2512         }
2513
2514         cfqd->active_queue = cfqq;
2515 }
2516
2517 /*
2518  * current cfqq expired its slice (or was too idle), select new one
2519  */
2520 static void
2521 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2522                     bool timed_out)
2523 {
2524         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2525
2526         if (cfq_cfqq_wait_request(cfqq))
2527                 cfq_del_timer(cfqd, cfqq);
2528
2529         cfq_clear_cfqq_wait_request(cfqq);
2530         cfq_clear_cfqq_wait_busy(cfqq);
2531
2532         /*
2533          * If this cfqq is shared between multiple processes, check to
2534          * make sure that those processes are still issuing I/Os within
2535          * the mean seek distance.  If not, it may be time to break the
2536          * queues apart again.
2537          */
2538         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2539                 cfq_mark_cfqq_split_coop(cfqq);
2540
2541         /*
2542          * store what was left of this slice, if the queue idled/timed out
2543          */
2544         if (timed_out) {
2545                 if (cfq_cfqq_slice_new(cfqq))
2546                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2547                 else
2548                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2549                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2550         }
2551
2552         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2553
2554         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2555                 cfq_del_cfqq_rr(cfqd, cfqq);
2556
2557         cfq_resort_rr_list(cfqd, cfqq);
2558
2559         if (cfqq == cfqd->active_queue)
2560                 cfqd->active_queue = NULL;
2561
2562         if (cfqd->active_cic) {
2563                 put_io_context(cfqd->active_cic->icq.ioc);
2564                 cfqd->active_cic = NULL;
2565         }
2566 }
2567
2568 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2569 {
2570         struct cfq_queue *cfqq = cfqd->active_queue;
2571
2572         if (cfqq)
2573                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2574 }
2575
2576 /*
2577  * Get next queue for service. Unless we have a queue preemption,
2578  * we'll simply select the first cfqq in the service tree.
2579  */
2580 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2581 {
2582         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2583                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2584
2585         if (!cfqd->rq_queued)
2586                 return NULL;
2587
2588         /* There is nothing to dispatch */
2589         if (!st)
2590                 return NULL;
2591         if (RB_EMPTY_ROOT(&st->rb))
2592                 return NULL;
2593         return cfq_rb_first(st);
2594 }
2595
2596 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2597 {
2598         struct cfq_group *cfqg;
2599         struct cfq_queue *cfqq;
2600         int i, j;
2601         struct cfq_rb_root *st;
2602
2603         if (!cfqd->rq_queued)
2604                 return NULL;
2605
2606         cfqg = cfq_get_next_cfqg(cfqd);
2607         if (!cfqg)
2608                 return NULL;
2609
2610         for_each_cfqg_st(cfqg, i, j, st)
2611                 if ((cfqq = cfq_rb_first(st)) != NULL)
2612                         return cfqq;
2613         return NULL;
2614 }
2615
2616 /*
2617  * Get and set a new active queue for service.
2618  */
2619 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2620                                               struct cfq_queue *cfqq)
2621 {
2622         if (!cfqq)
2623                 cfqq = cfq_get_next_queue(cfqd);
2624
2625         __cfq_set_active_queue(cfqd, cfqq);
2626         return cfqq;
2627 }
2628
2629 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2630                                           struct request *rq)
2631 {
2632         if (blk_rq_pos(rq) >= cfqd->last_position)
2633                 return blk_rq_pos(rq) - cfqd->last_position;
2634         else
2635                 return cfqd->last_position - blk_rq_pos(rq);
2636 }
2637
2638 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2639                                struct request *rq)
2640 {
2641         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2642 }
2643
2644 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2645                                     struct cfq_queue *cur_cfqq)
2646 {
2647         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2648         struct rb_node *parent, *node;
2649         struct cfq_queue *__cfqq;
2650         sector_t sector = cfqd->last_position;
2651
2652         if (RB_EMPTY_ROOT(root))
2653                 return NULL;
2654
2655         /*
2656          * First, if we find a request starting at the end of the last
2657          * request, choose it.
2658          */
2659         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2660         if (__cfqq)
2661                 return __cfqq;
2662
2663         /*
2664          * If the exact sector wasn't found, the parent of the NULL leaf
2665          * will contain the closest sector.
2666          */
2667         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2668         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2669                 return __cfqq;
2670
2671         if (blk_rq_pos(__cfqq->next_rq) < sector)
2672                 node = rb_next(&__cfqq->p_node);
2673         else
2674                 node = rb_prev(&__cfqq->p_node);
2675         if (!node)
2676                 return NULL;
2677
2678         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2679         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2680                 return __cfqq;
2681
2682         return NULL;
2683 }
2684
2685 /*
2686  * cfqd - obvious
2687  * cur_cfqq - passed in so that we don't decide that the current queue is
2688  *            closely cooperating with itself.
2689  *
2690  * So, basically we're assuming that that cur_cfqq has dispatched at least
2691  * one request, and that cfqd->last_position reflects a position on the disk
2692  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2693  * assumption.
2694  */
2695 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2696                                               struct cfq_queue *cur_cfqq)
2697 {
2698         struct cfq_queue *cfqq;
2699
2700         if (cfq_class_idle(cur_cfqq))
2701                 return NULL;
2702         if (!cfq_cfqq_sync(cur_cfqq))
2703                 return NULL;
2704         if (CFQQ_SEEKY(cur_cfqq))
2705                 return NULL;
2706
2707         /*
2708          * Don't search priority tree if it's the only queue in the group.
2709          */
2710         if (cur_cfqq->cfqg->nr_cfqq == 1)
2711                 return NULL;
2712
2713         /*
2714          * We should notice if some of the queues are cooperating, eg
2715          * working closely on the same area of the disk. In that case,
2716          * we can group them together and don't waste time idling.
2717          */
2718         cfqq = cfqq_close(cfqd, cur_cfqq);
2719         if (!cfqq)
2720                 return NULL;
2721
2722         /* If new queue belongs to different cfq_group, don't choose it */
2723         if (cur_cfqq->cfqg != cfqq->cfqg)
2724                 return NULL;
2725
2726         /*
2727          * It only makes sense to merge sync queues.
2728          */
2729         if (!cfq_cfqq_sync(cfqq))
2730                 return NULL;
2731         if (CFQQ_SEEKY(cfqq))
2732                 return NULL;
2733
2734         /*
2735          * Do not merge queues of different priority classes
2736          */
2737         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2738                 return NULL;
2739
2740         return cfqq;
2741 }
2742
2743 /*
2744  * Determine whether we should enforce idle window for this queue.
2745  */
2746
2747 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2748 {
2749         enum wl_class_t wl_class = cfqq_class(cfqq);
2750         struct cfq_rb_root *st = cfqq->service_tree;
2751
2752         BUG_ON(!st);
2753         BUG_ON(!st->count);
2754
2755         if (!cfqd->cfq_slice_idle)
2756                 return false;
2757
2758         /* We never do for idle class queues. */
2759         if (wl_class == IDLE_WORKLOAD)
2760                 return false;
2761
2762         /* We do for queues that were marked with idle window flag. */
2763         if (cfq_cfqq_idle_window(cfqq) &&
2764            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2765                 return true;
2766
2767         /*
2768          * Otherwise, we do only if they are the last ones
2769          * in their service tree.
2770          */
2771         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2772            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2773                 return true;
2774         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2775         return false;
2776 }
2777
2778 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2779 {
2780         struct cfq_queue *cfqq = cfqd->active_queue;
2781         struct cfq_io_cq *cic;
2782         unsigned long sl, group_idle = 0;
2783
2784         /*
2785          * SSD device without seek penalty, disable idling. But only do so
2786          * for devices that support queuing, otherwise we still have a problem
2787          * with sync vs async workloads.
2788          */
2789         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2790                 return;
2791
2792         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2793         WARN_ON(cfq_cfqq_slice_new(cfqq));
2794
2795         /*
2796          * idle is disabled, either manually or by past process history
2797          */
2798         if (!cfq_should_idle(cfqd, cfqq)) {
2799                 /* no queue idling. Check for group idling */
2800                 if (cfqd->cfq_group_idle)
2801                         group_idle = cfqd->cfq_group_idle;
2802                 else
2803                         return;
2804         }
2805
2806         /*
2807          * still active requests from this queue, don't idle
2808          */
2809         if (cfqq->dispatched)
2810                 return;
2811
2812         /*
2813          * task has exited, don't wait
2814          */
2815         cic = cfqd->active_cic;
2816         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2817                 return;
2818
2819         /*
2820          * If our average think time is larger than the remaining time
2821          * slice, then don't idle. This avoids overrunning the allotted
2822          * time slice.
2823          */
2824         if (sample_valid(cic->ttime.ttime_samples) &&
2825             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2826                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2827                              cic->ttime.ttime_mean);
2828                 return;
2829         }
2830
2831         /* There are other queues in the group, don't do group idle */
2832         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2833                 return;
2834
2835         cfq_mark_cfqq_wait_request(cfqq);
2836
2837         if (group_idle)
2838                 sl = cfqd->cfq_group_idle;
2839         else
2840                 sl = cfqd->cfq_slice_idle;
2841
2842         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2843         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2844         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2845                         group_idle ? 1 : 0);
2846 }
2847
2848 /*
2849  * Move request from internal lists to the request queue dispatch list.
2850  */
2851 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2852 {
2853         struct cfq_data *cfqd = q->elevator->elevator_data;
2854         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2855
2856         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2857
2858         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2859         cfq_remove_request(rq);
2860         cfqq->dispatched++;
2861         (RQ_CFQG(rq))->dispatched++;
2862         elv_dispatch_sort(q, rq);
2863
2864         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2865         cfqq->nr_sectors += blk_rq_sectors(rq);
2866         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2867 }
2868
2869 /*
2870  * return expired entry, or NULL to just start from scratch in rbtree
2871  */
2872 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2873 {
2874         struct request *rq = NULL;
2875
2876         if (cfq_cfqq_fifo_expire(cfqq))
2877                 return NULL;
2878
2879         cfq_mark_cfqq_fifo_expire(cfqq);
2880
2881         if (list_empty(&cfqq->fifo))
2882                 return NULL;
2883
2884         rq = rq_entry_fifo(cfqq->fifo.next);
2885         if (time_before(jiffies, rq->fifo_time))
2886                 rq = NULL;
2887
2888         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2889         return rq;
2890 }
2891
2892 static inline int
2893 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2894 {
2895         const int base_rq = cfqd->cfq_slice_async_rq;
2896
2897         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2898
2899         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2900 }
2901
2902 /*
2903  * Must be called with the queue_lock held.
2904  */
2905 static int cfqq_process_refs(struct cfq_queue *cfqq)
2906 {
2907         int process_refs, io_refs;
2908
2909         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2910         process_refs = cfqq->ref - io_refs;
2911         BUG_ON(process_refs < 0);
2912         return process_refs;
2913 }
2914
2915 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2916 {
2917         int process_refs, new_process_refs;
2918         struct cfq_queue *__cfqq;
2919
2920         /*
2921          * If there are no process references on the new_cfqq, then it is
2922          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2923          * chain may have dropped their last reference (not just their
2924          * last process reference).
2925          */
2926         if (!cfqq_process_refs(new_cfqq))
2927                 return;
2928
2929         /* Avoid a circular list and skip interim queue merges */
2930         while ((__cfqq = new_cfqq->new_cfqq)) {
2931                 if (__cfqq == cfqq)
2932                         return;
2933                 new_cfqq = __cfqq;
2934         }
2935
2936         process_refs = cfqq_process_refs(cfqq);
2937         new_process_refs = cfqq_process_refs(new_cfqq);
2938         /*
2939          * If the process for the cfqq has gone away, there is no
2940          * sense in merging the queues.
2941          */
2942         if (process_refs == 0 || new_process_refs == 0)
2943                 return;
2944
2945         /*
2946          * Merge in the direction of the lesser amount of work.
2947          */
2948         if (new_process_refs >= process_refs) {
2949                 cfqq->new_cfqq = new_cfqq;
2950                 new_cfqq->ref += process_refs;
2951         } else {
2952                 new_cfqq->new_cfqq = cfqq;
2953                 cfqq->ref += new_process_refs;
2954         }
2955 }
2956
2957 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2958                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2959 {
2960         struct cfq_queue *queue;
2961         int i;
2962         bool key_valid = false;
2963         unsigned long lowest_key = 0;
2964         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2965
2966         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2967                 /* select the one with lowest rb_key */
2968                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2969                 if (queue &&
2970                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2971                         lowest_key = queue->rb_key;
2972                         cur_best = i;
2973                         key_valid = true;
2974                 }
2975         }
2976
2977         return cur_best;
2978 }
2979
2980 static void
2981 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2982 {
2983         unsigned slice;
2984         unsigned count;
2985         struct cfq_rb_root *st;
2986         unsigned group_slice;
2987         enum wl_class_t original_class = cfqd->serving_wl_class;
2988
2989         /* Choose next priority. RT > BE > IDLE */
2990         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2991                 cfqd->serving_wl_class = RT_WORKLOAD;
2992         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2993                 cfqd->serving_wl_class = BE_WORKLOAD;
2994         else {
2995                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2996                 cfqd->workload_expires = jiffies + 1;
2997                 return;
2998         }
2999
3000         if (original_class != cfqd->serving_wl_class)
3001                 goto new_workload;
3002
3003         /*
3004          * For RT and BE, we have to choose also the type
3005          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3006          * expiration time
3007          */
3008         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3009         count = st->count;
3010
3011         /*
3012          * check workload expiration, and that we still have other queues ready
3013          */
3014         if (count && !time_after(jiffies, cfqd->workload_expires))
3015                 return;
3016
3017 new_workload:
3018         /* otherwise select new workload type */
3019         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3020                                         cfqd->serving_wl_class);
3021         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3022         count = st->count;
3023
3024         /*
3025          * the workload slice is computed as a fraction of target latency
3026          * proportional to the number of queues in that workload, over
3027          * all the queues in the same priority class
3028          */
3029         group_slice = cfq_group_slice(cfqd, cfqg);
3030
3031         slice = group_slice * count /
3032                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3033                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3034                                         cfqg));
3035
3036         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3037                 unsigned int tmp;
3038
3039                 /*
3040                  * Async queues are currently system wide. Just taking
3041                  * proportion of queues with-in same group will lead to higher
3042                  * async ratio system wide as generally root group is going
3043                  * to have higher weight. A more accurate thing would be to
3044                  * calculate system wide asnc/sync ratio.
3045                  */
3046                 tmp = cfqd->cfq_target_latency *
3047                         cfqg_busy_async_queues(cfqd, cfqg);
3048                 tmp = tmp/cfqd->busy_queues;
3049                 slice = min_t(unsigned, slice, tmp);
3050
3051                 /* async workload slice is scaled down according to
3052                  * the sync/async slice ratio. */
3053                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3054         } else
3055                 /* sync workload slice is at least 2 * cfq_slice_idle */
3056                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3057
3058         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3059         cfq_log(cfqd, "workload slice:%d", slice);
3060         cfqd->workload_expires = jiffies + slice;
3061 }
3062
3063 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3064 {
3065         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3066         struct cfq_group *cfqg;
3067
3068         if (RB_EMPTY_ROOT(&st->rb))
3069                 return NULL;
3070         cfqg = cfq_rb_first_group(st);
3071         update_min_vdisktime(st);
3072         return cfqg;
3073 }
3074
3075 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3076 {
3077         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3078
3079         cfqd->serving_group = cfqg;
3080
3081         /* Restore the workload type data */
3082         if (cfqg->saved_wl_slice) {
3083                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3084                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3085                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3086         } else
3087                 cfqd->workload_expires = jiffies - 1;
3088
3089         choose_wl_class_and_type(cfqd, cfqg);
3090 }
3091
3092 /*
3093  * Select a queue for service. If we have a current active queue,
3094  * check whether to continue servicing it, or retrieve and set a new one.
3095  */
3096 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3097 {
3098         struct cfq_queue *cfqq, *new_cfqq = NULL;
3099
3100         cfqq = cfqd->active_queue;
3101         if (!cfqq)
3102                 goto new_queue;
3103
3104         if (!cfqd->rq_queued)
3105                 return NULL;
3106
3107         /*
3108          * We were waiting for group to get backlogged. Expire the queue
3109          */
3110         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3111                 goto expire;
3112
3113         /*
3114          * The active queue has run out of time, expire it and select new.
3115          */
3116         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3117                 /*
3118                  * If slice had not expired at the completion of last request
3119                  * we might not have turned on wait_busy flag. Don't expire
3120                  * the queue yet. Allow the group to get backlogged.
3121                  *
3122                  * The very fact that we have used the slice, that means we
3123                  * have been idling all along on this queue and it should be
3124                  * ok to wait for this request to complete.
3125                  */
3126                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3127                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3128                         cfqq = NULL;
3129                         goto keep_queue;
3130                 } else
3131                         goto check_group_idle;
3132         }
3133
3134         /*
3135          * The active queue has requests and isn't expired, allow it to
3136          * dispatch.
3137          */
3138         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3139                 goto keep_queue;
3140
3141         /*
3142          * If another queue has a request waiting within our mean seek
3143          * distance, let it run.  The expire code will check for close
3144          * cooperators and put the close queue at the front of the service
3145          * tree.  If possible, merge the expiring queue with the new cfqq.
3146          */
3147         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3148         if (new_cfqq) {
3149                 if (!cfqq->new_cfqq)
3150                         cfq_setup_merge(cfqq, new_cfqq);
3151                 goto expire;
3152         }
3153
3154         /*
3155          * No requests pending. If the active queue still has requests in
3156          * flight or is idling for a new request, allow either of these
3157          * conditions to happen (or time out) before selecting a new queue.
3158          */
3159         if (timer_pending(&cfqd->idle_slice_timer)) {
3160                 cfqq = NULL;
3161                 goto keep_queue;
3162         }
3163
3164         /*
3165          * This is a deep seek queue, but the device is much faster than
3166          * the queue can deliver, don't idle
3167          **/
3168         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3169             (cfq_cfqq_slice_new(cfqq) ||
3170             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3171                 cfq_clear_cfqq_deep(cfqq);
3172                 cfq_clear_cfqq_idle_window(cfqq);
3173         }
3174
3175         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3176                 cfqq = NULL;
3177                 goto keep_queue;
3178         }
3179
3180         /*
3181          * If group idle is enabled and there are requests dispatched from
3182          * this group, wait for requests to complete.
3183          */
3184 check_group_idle:
3185         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3186             cfqq->cfqg->dispatched &&
3187             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3188                 cfqq = NULL;
3189                 goto keep_queue;
3190         }
3191
3192 expire:
3193         cfq_slice_expired(cfqd, 0);
3194 new_queue:
3195         /*
3196          * Current queue expired. Check if we have to switch to a new
3197          * service tree
3198          */
3199         if (!new_cfqq)
3200                 cfq_choose_cfqg(cfqd);
3201
3202         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3203 keep_queue:
3204         return cfqq;
3205 }
3206
3207 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3208 {
3209         int dispatched = 0;
3210
3211         while (cfqq->next_rq) {
3212                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3213                 dispatched++;
3214         }
3215
3216         BUG_ON(!list_empty(&cfqq->fifo));
3217
3218         /* By default cfqq is not expired if it is empty. Do it explicitly */
3219         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3220         return dispatched;
3221 }
3222
3223 /*
3224  * Drain our current requests. Used for barriers and when switching
3225  * io schedulers on-the-fly.
3226  */
3227 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3228 {
3229         struct cfq_queue *cfqq;
3230         int dispatched = 0;
3231
3232         /* Expire the timeslice of the current active queue first */
3233         cfq_slice_expired(cfqd, 0);
3234         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3235                 __cfq_set_active_queue(cfqd, cfqq);
3236                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3237         }
3238
3239         BUG_ON(cfqd->busy_queues);
3240
3241         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3242         return dispatched;
3243 }
3244
3245 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3246         struct cfq_queue *cfqq)
3247 {
3248         /* the queue hasn't finished any request, can't estimate */
3249         if (cfq_cfqq_slice_new(cfqq))
3250                 return true;
3251         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3252                 cfqq->slice_end))
3253                 return true;
3254
3255         return false;
3256 }
3257
3258 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3259 {
3260         unsigned int max_dispatch;
3261
3262         /*
3263          * Drain async requests before we start sync IO
3264          */
3265         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3266                 return false;
3267
3268         /*
3269          * If this is an async queue and we have sync IO in flight, let it wait
3270          */
3271         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3272                 return false;
3273
3274         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3275         if (cfq_class_idle(cfqq))
3276                 max_dispatch = 1;
3277
3278         /*
3279          * Does this cfqq already have too much IO in flight?
3280          */
3281         if (cfqq->dispatched >= max_dispatch) {
3282                 bool promote_sync = false;
3283                 /*
3284                  * idle queue must always only have a single IO in flight
3285                  */
3286                 if (cfq_class_idle(cfqq))
3287                         return false;
3288
3289                 /*
3290                  * If there is only one sync queue
3291                  * we can ignore async queue here and give the sync
3292                  * queue no dispatch limit. The reason is a sync queue can
3293                  * preempt async queue, limiting the sync queue doesn't make
3294                  * sense. This is useful for aiostress test.
3295                  */
3296                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3297                         promote_sync = true;
3298
3299                 /*
3300                  * We have other queues, don't allow more IO from this one
3301                  */
3302                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3303                                 !promote_sync)
3304                         return false;
3305
3306                 /*
3307                  * Sole queue user, no limit
3308                  */
3309                 if (cfqd->busy_queues == 1 || promote_sync)
3310                         max_dispatch = -1;
3311                 else
3312                         /*
3313                          * Normally we start throttling cfqq when cfq_quantum/2
3314                          * requests have been dispatched. But we can drive
3315                          * deeper queue depths at the beginning of slice
3316                          * subjected to upper limit of cfq_quantum.
3317                          * */
3318                         max_dispatch = cfqd->cfq_quantum;
3319         }
3320
3321         /*
3322          * Async queues must wait a bit before being allowed dispatch.
3323          * We also ramp up the dispatch depth gradually for async IO,
3324          * based on the last sync IO we serviced
3325          */
3326         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3327                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3328                 unsigned int depth;
3329
3330                 depth = last_sync / cfqd->cfq_slice[1];
3331                 if (!depth && !cfqq->dispatched)
3332                         depth = 1;
3333                 if (depth < max_dispatch)
3334                         max_dispatch = depth;
3335         }
3336
3337         /*
3338          * If we're below the current max, allow a dispatch
3339          */
3340         return cfqq->dispatched < max_dispatch;
3341 }
3342
3343 /*
3344  * Dispatch a request from cfqq, moving them to the request queue
3345  * dispatch list.
3346  */
3347 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3348 {
3349         struct request *rq;
3350
3351         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3352
3353         if (!cfq_may_dispatch(cfqd, cfqq))
3354                 return false;
3355
3356         /*
3357          * follow expired path, else get first next available
3358          */
3359         rq = cfq_check_fifo(cfqq);
3360         if (!rq)
3361                 rq = cfqq->next_rq;
3362
3363         /*
3364          * insert request into driver dispatch list
3365          */
3366         cfq_dispatch_insert(cfqd->queue, rq);
3367
3368         if (!cfqd->active_cic) {
3369                 struct cfq_io_cq *cic = RQ_CIC(rq);
3370
3371                 atomic_long_inc(&cic->icq.ioc->refcount);
3372                 cfqd->active_cic = cic;
3373         }
3374
3375         return true;
3376 }
3377
3378 /*
3379  * Find the cfqq that we need to service and move a request from that to the
3380  * dispatch list
3381  */
3382 static int cfq_dispatch_requests(struct request_queue *q, int force)
3383 {
3384         struct cfq_data *cfqd = q->elevator->elevator_data;
3385         struct cfq_queue *cfqq;
3386
3387         if (!cfqd->busy_queues)
3388                 return 0;
3389
3390         if (unlikely(force))
3391                 return cfq_forced_dispatch(cfqd);
3392
3393         cfqq = cfq_select_queue(cfqd);
3394         if (!cfqq)
3395                 return 0;
3396
3397         /*
3398          * Dispatch a request from this cfqq, if it is allowed
3399          */
3400         if (!cfq_dispatch_request(cfqd, cfqq))
3401                 return 0;
3402
3403         cfqq->slice_dispatch++;
3404         cfq_clear_cfqq_must_dispatch(cfqq);
3405
3406         /*
3407          * expire an async queue immediately if it has used up its slice. idle
3408          * queue always expire after 1 dispatch round.
3409          */
3410         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3411             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3412             cfq_class_idle(cfqq))) {
3413                 cfqq->slice_end = jiffies + 1;
3414                 cfq_slice_expired(cfqd, 0);
3415         }
3416
3417         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3418         return 1;
3419 }
3420
3421 /*
3422  * task holds one reference to the queue, dropped when task exits. each rq
3423  * in-flight on this queue also holds a reference, dropped when rq is freed.
3424  *
3425  * Each cfq queue took a reference on the parent group. Drop it now.
3426  * queue lock must be held here.
3427  */
3428 static void cfq_put_queue(struct cfq_queue *cfqq)
3429 {
3430         struct cfq_data *cfqd = cfqq->cfqd;
3431         struct cfq_group *cfqg;
3432
3433         BUG_ON(cfqq->ref <= 0);
3434
3435         cfqq->ref--;
3436         if (cfqq->ref)
3437                 return;
3438
3439         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3440         BUG_ON(rb_first(&cfqq->sort_list));
3441         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3442         cfqg = cfqq->cfqg;
3443
3444         if (unlikely(cfqd->active_queue == cfqq)) {
3445                 __cfq_slice_expired(cfqd, cfqq, 0);
3446                 cfq_schedule_dispatch(cfqd);
3447         }
3448
3449         BUG_ON(cfq_cfqq_on_rr(cfqq));
3450         kmem_cache_free(cfq_pool, cfqq);
3451         cfqg_put(cfqg);
3452 }
3453
3454 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3455 {
3456         struct cfq_queue *__cfqq, *next;
3457
3458         /*
3459          * If this queue was scheduled to merge with another queue, be
3460          * sure to drop the reference taken on that queue (and others in
3461          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3462          */
3463         __cfqq = cfqq->new_cfqq;
3464         while (__cfqq) {
3465                 if (__cfqq == cfqq) {
3466                         WARN(1, "cfqq->new_cfqq loop detected\n");
3467                         break;
3468                 }
3469                 next = __cfqq->new_cfqq;
3470                 cfq_put_queue(__cfqq);
3471                 __cfqq = next;
3472         }
3473 }
3474
3475 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3476 {
3477         if (unlikely(cfqq == cfqd->active_queue)) {
3478                 __cfq_slice_expired(cfqd, cfqq, 0);
3479                 cfq_schedule_dispatch(cfqd);
3480         }
3481
3482         cfq_put_cooperator(cfqq);
3483
3484         cfq_put_queue(cfqq);
3485 }
3486
3487 static void cfq_init_icq(struct io_cq *icq)
3488 {
3489         struct cfq_io_cq *cic = icq_to_cic(icq);
3490
3491         cic->ttime.last_end_request = jiffies;
3492 }
3493
3494 static void cfq_exit_icq(struct io_cq *icq)
3495 {
3496         struct cfq_io_cq *cic = icq_to_cic(icq);
3497         struct cfq_data *cfqd = cic_to_cfqd(cic);
3498
3499         if (cic->cfqq[BLK_RW_ASYNC]) {
3500                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3501                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3502         }
3503
3504         if (cic->cfqq[BLK_RW_SYNC]) {
3505                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3506                 cic->cfqq[BLK_RW_SYNC] = NULL;
3507         }
3508 }
3509
3510 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3511 {
3512         struct task_struct *tsk = current;
3513         int ioprio_class;
3514
3515         if (!cfq_cfqq_prio_changed(cfqq))
3516                 return;
3517
3518         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3519         switch (ioprio_class) {
3520         default:
3521                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3522         case IOPRIO_CLASS_NONE:
3523                 /*
3524                  * no prio set, inherit CPU scheduling settings
3525                  */
3526                 cfqq->ioprio = task_nice_ioprio(tsk);
3527                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3528                 break;
3529         case IOPRIO_CLASS_RT:
3530                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3531                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3532                 break;
3533         case IOPRIO_CLASS_BE:
3534                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3535                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3536                 break;
3537         case IOPRIO_CLASS_IDLE:
3538                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3539                 cfqq->ioprio = 7;
3540                 cfq_clear_cfqq_idle_window(cfqq);
3541                 break;
3542         }
3543
3544         /*
3545          * keep track of original prio settings in case we have to temporarily
3546          * elevate the priority of this queue
3547          */
3548         cfqq->org_ioprio = cfqq->ioprio;
3549         cfq_clear_cfqq_prio_changed(cfqq);
3550 }
3551
3552 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3553 {
3554         int ioprio = cic->icq.ioc->ioprio;
3555         struct cfq_data *cfqd = cic_to_cfqd(cic);
3556         struct cfq_queue *cfqq;
3557
3558         /*
3559          * Check whether ioprio has changed.  The condition may trigger
3560          * spuriously on a newly created cic but there's no harm.
3561          */
3562         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3563                 return;
3564
3565         cfqq = cic->cfqq[BLK_RW_ASYNC];
3566         if (cfqq) {
3567                 struct cfq_queue *new_cfqq;
3568                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3569                                          GFP_ATOMIC);
3570                 if (new_cfqq) {
3571                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3572                         cfq_put_queue(cfqq);
3573                 }
3574         }
3575
3576         cfqq = cic->cfqq[BLK_RW_SYNC];
3577         if (cfqq)
3578                 cfq_mark_cfqq_prio_changed(cfqq);
3579
3580         cic->ioprio = ioprio;
3581 }
3582
3583 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3584                           pid_t pid, bool is_sync)
3585 {
3586         RB_CLEAR_NODE(&cfqq->rb_node);
3587         RB_CLEAR_NODE(&cfqq->p_node);
3588         INIT_LIST_HEAD(&cfqq->fifo);
3589
3590         cfqq->ref = 0;
3591         cfqq->cfqd = cfqd;
3592
3593         cfq_mark_cfqq_prio_changed(cfqq);
3594
3595         if (is_sync) {
3596                 if (!cfq_class_idle(cfqq))
3597                         cfq_mark_cfqq_idle_window(cfqq);
3598                 cfq_mark_cfqq_sync(cfqq);
3599         }
3600         cfqq->pid = pid;
3601 }
3602
3603 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3604 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3605 {
3606         struct cfq_data *cfqd = cic_to_cfqd(cic);
3607         struct cfq_queue *sync_cfqq;
3608         uint64_t serial_nr;
3609
3610         rcu_read_lock();
3611         serial_nr = bio_blkcg(bio)->css.serial_nr;
3612         rcu_read_unlock();
3613
3614         /*
3615          * Check whether blkcg has changed.  The condition may trigger
3616          * spuriously on a newly created cic but there's no harm.
3617          */
3618         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3619                 return;
3620
3621         sync_cfqq = cic_to_cfqq(cic, 1);
3622         if (sync_cfqq) {
3623                 /*
3624                  * Drop reference to sync queue. A new sync queue will be
3625                  * assigned in new group upon arrival of a fresh request.
3626                  */
3627                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3628                 cic_set_cfqq(cic, NULL, 1);
3629                 cfq_put_queue(sync_cfqq);
3630         }
3631
3632         cic->blkcg_serial_nr = serial_nr;
3633 }
3634 #else
3635 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3636 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3637
3638 static struct cfq_queue *
3639 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3640                      struct bio *bio, gfp_t gfp_mask)
3641 {
3642         struct blkcg *blkcg;
3643         struct cfq_queue *cfqq, *new_cfqq = NULL;
3644         struct cfq_group *cfqg;
3645
3646 retry:
3647         rcu_read_lock();
3648
3649         blkcg = bio_blkcg(bio);
3650         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3651         if (!cfqg) {
3652                 cfqq = &cfqd->oom_cfqq;
3653                 goto out;
3654         }
3655
3656         cfqq = cic_to_cfqq(cic, is_sync);
3657
3658         /*
3659          * Always try a new alloc if we fell back to the OOM cfqq
3660          * originally, since it should just be a temporary situation.
3661          */
3662         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3663                 cfqq = NULL;
3664                 if (new_cfqq) {
3665                         cfqq = new_cfqq;
3666                         new_cfqq = NULL;
3667                 } else if (gfp_mask & __GFP_WAIT) {
3668                         rcu_read_unlock();
3669                         spin_unlock_irq(cfqd->queue->queue_lock);
3670                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3671                                         gfp_mask | __GFP_ZERO,
3672                                         cfqd->queue->node);
3673                         spin_lock_irq(cfqd->queue->queue_lock);
3674                         if (new_cfqq)
3675                                 goto retry;
3676                         else
3677                                 return &cfqd->oom_cfqq;
3678                 } else {
3679                         cfqq = kmem_cache_alloc_node(cfq_pool,
3680                                         gfp_mask | __GFP_ZERO,
3681                                         cfqd->queue->node);
3682                 }
3683
3684                 if (cfqq) {
3685                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3686                         cfq_init_prio_data(cfqq, cic);
3687                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3688                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3689                 } else
3690                         cfqq = &cfqd->oom_cfqq;
3691         }
3692 out:
3693         if (new_cfqq)
3694                 kmem_cache_free(cfq_pool, new_cfqq);
3695
3696         rcu_read_unlock();
3697         return cfqq;
3698 }
3699
3700 static struct cfq_queue **
3701 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3702 {
3703         switch (ioprio_class) {
3704         case IOPRIO_CLASS_RT:
3705                 return &cfqd->async_cfqq[0][ioprio];
3706         case IOPRIO_CLASS_NONE:
3707                 ioprio = IOPRIO_NORM;
3708                 /* fall through */
3709         case IOPRIO_CLASS_BE:
3710                 return &cfqd->async_cfqq[1][ioprio];
3711         case IOPRIO_CLASS_IDLE:
3712                 return &cfqd->async_idle_cfqq;
3713         default:
3714                 BUG();
3715         }
3716 }
3717
3718 static struct cfq_queue *
3719 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3720               struct bio *bio, gfp_t gfp_mask)
3721 {
3722         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3723         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3724         struct cfq_queue **async_cfqq = NULL;
3725         struct cfq_queue *cfqq = NULL;
3726
3727         if (!is_sync) {
3728                 if (!ioprio_valid(cic->ioprio)) {
3729                         struct task_struct *tsk = current;
3730                         ioprio = task_nice_ioprio(tsk);
3731                         ioprio_class = task_nice_ioclass(tsk);
3732                 }
3733                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3734                 cfqq = *async_cfqq;
3735         }
3736
3737         if (!cfqq)
3738                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3739
3740         /*
3741          * pin the queue now that it's allocated, scheduler exit will prune it
3742          */
3743         if (!is_sync && !(*async_cfqq)) {
3744                 cfqq->ref++;
3745                 *async_cfqq = cfqq;
3746         }
3747
3748         cfqq->ref++;
3749         return cfqq;
3750 }
3751
3752 static void
3753 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3754 {
3755         unsigned long elapsed = jiffies - ttime->last_end_request;
3756         elapsed = min(elapsed, 2UL * slice_idle);
3757
3758         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3759         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3760         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3761 }
3762
3763 static void
3764 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3765                         struct cfq_io_cq *cic)
3766 {
3767         if (cfq_cfqq_sync(cfqq)) {
3768                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3769                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3770                         cfqd->cfq_slice_idle);
3771         }
3772 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3773         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3774 #endif
3775 }
3776
3777 static void
3778 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3779                        struct request *rq)
3780 {
3781         sector_t sdist = 0;
3782         sector_t n_sec = blk_rq_sectors(rq);
3783         if (cfqq->last_request_pos) {
3784                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3785                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3786                 else
3787                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3788         }
3789
3790         cfqq->seek_history <<= 1;
3791         if (blk_queue_nonrot(cfqd->queue))
3792                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3793         else
3794                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3795 }
3796
3797 /*
3798  * Disable idle window if the process thinks too long or seeks so much that
3799  * it doesn't matter
3800  */
3801 static void
3802 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3803                        struct cfq_io_cq *cic)
3804 {
3805         int old_idle, enable_idle;
3806
3807         /*
3808          * Don't idle for async or idle io prio class
3809          */
3810         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3811                 return;
3812
3813         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3814
3815         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3816                 cfq_mark_cfqq_deep(cfqq);
3817
3818         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3819                 enable_idle = 0;
3820         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3821                  !cfqd->cfq_slice_idle ||
3822                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3823                 enable_idle = 0;
3824         else if (sample_valid(cic->ttime.ttime_samples)) {
3825                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3826                         enable_idle = 0;
3827                 else
3828                         enable_idle = 1;
3829         }
3830
3831         if (old_idle != enable_idle) {
3832                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3833                 if (enable_idle)
3834                         cfq_mark_cfqq_idle_window(cfqq);
3835                 else
3836                         cfq_clear_cfqq_idle_window(cfqq);
3837         }
3838 }
3839
3840 /*
3841  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3842  * no or if we aren't sure, a 1 will cause a preempt.
3843  */
3844 static bool
3845 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3846                    struct request *rq)
3847 {
3848         struct cfq_queue *cfqq;
3849
3850         cfqq = cfqd->active_queue;
3851         if (!cfqq)
3852                 return false;
3853
3854         if (cfq_class_idle(new_cfqq))
3855                 return false;
3856
3857         if (cfq_class_idle(cfqq))
3858                 return true;
3859
3860         /*
3861          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3862          */
3863         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3864                 return false;
3865
3866         /*
3867          * if the new request is sync, but the currently running queue is
3868          * not, let the sync request have priority.
3869          */
3870         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3871                 return true;
3872
3873         if (new_cfqq->cfqg != cfqq->cfqg)
3874                 return false;
3875
3876         if (cfq_slice_used(cfqq))
3877                 return true;
3878
3879         /* Allow preemption only if we are idling on sync-noidle tree */
3880         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3881             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3882             new_cfqq->service_tree->count == 2 &&
3883             RB_EMPTY_ROOT(&cfqq->sort_list))
3884                 return true;
3885
3886         /*
3887          * So both queues are sync. Let the new request get disk time if
3888          * it's a metadata request and the current queue is doing regular IO.
3889          */
3890         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3891                 return true;
3892
3893         /*
3894          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3895          */
3896         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3897                 return true;
3898
3899         /* An idle queue should not be idle now for some reason */
3900         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3901                 return true;
3902
3903         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3904                 return false;
3905
3906         /*
3907          * if this request is as-good as one we would expect from the
3908          * current cfqq, let it preempt
3909          */
3910         if (cfq_rq_close(cfqd, cfqq, rq))
3911                 return true;
3912
3913         return false;
3914 }
3915
3916 /*
3917  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3918  * let it have half of its nominal slice.
3919  */
3920 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3921 {
3922         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3923
3924         cfq_log_cfqq(cfqd, cfqq, "preempt");
3925         cfq_slice_expired(cfqd, 1);
3926
3927         /*
3928          * workload type is changed, don't save slice, otherwise preempt
3929          * doesn't happen
3930          */
3931         if (old_type != cfqq_type(cfqq))
3932                 cfqq->cfqg->saved_wl_slice = 0;
3933
3934         /*
3935          * Put the new queue at the front of the of the current list,
3936          * so we know that it will be selected next.
3937          */
3938         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3939
3940         cfq_service_tree_add(cfqd, cfqq, 1);
3941
3942         cfqq->slice_end = 0;
3943         cfq_mark_cfqq_slice_new(cfqq);
3944 }
3945
3946 /*
3947  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3948  * something we should do about it
3949  */
3950 static void
3951 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3952                 struct request *rq)
3953 {
3954         struct cfq_io_cq *cic = RQ_CIC(rq);
3955
3956         cfqd->rq_queued++;
3957         if (rq->cmd_flags & REQ_PRIO)
3958                 cfqq->prio_pending++;
3959
3960         cfq_update_io_thinktime(cfqd, cfqq, cic);
3961         cfq_update_io_seektime(cfqd, cfqq, rq);
3962         cfq_update_idle_window(cfqd, cfqq, cic);
3963
3964         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3965
3966         if (cfqq == cfqd->active_queue) {
3967                 /*
3968                  * Remember that we saw a request from this process, but
3969                  * don't start queuing just yet. Otherwise we risk seeing lots
3970                  * of tiny requests, because we disrupt the normal plugging
3971                  * and merging. If the request is already larger than a single
3972                  * page, let it rip immediately. For that case we assume that
3973                  * merging is already done. Ditto for a busy system that
3974                  * has other work pending, don't risk delaying until the
3975                  * idle timer unplug to continue working.
3976                  */
3977                 if (cfq_cfqq_wait_request(cfqq)) {
3978                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3979                             cfqd->busy_queues > 1) {
3980                                 cfq_del_timer(cfqd, cfqq);
3981                                 cfq_clear_cfqq_wait_request(cfqq);
3982                                 __blk_run_queue(cfqd->queue);
3983                         } else {
3984                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3985                                 cfq_mark_cfqq_must_dispatch(cfqq);
3986                         }
3987                 }
3988         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3989                 /*
3990                  * not the active queue - expire current slice if it is
3991                  * idle and has expired it's mean thinktime or this new queue
3992                  * has some old slice time left and is of higher priority or
3993                  * this new queue is RT and the current one is BE
3994                  */
3995                 cfq_preempt_queue(cfqd, cfqq);
3996                 __blk_run_queue(cfqd->queue);
3997         }
3998 }
3999
4000 static void cfq_insert_request(struct request_queue *q, struct request *rq)
4001 {
4002         struct cfq_data *cfqd = q->elevator->elevator_data;
4003         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4004
4005         cfq_log_cfqq(cfqd, cfqq, "insert_request");
4006         cfq_init_prio_data(cfqq, RQ_CIC(rq));
4007
4008         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
4009         list_add_tail(&rq->queuelist, &cfqq->fifo);
4010         cfq_add_rq_rb(rq);
4011         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
4012                                  rq->cmd_flags);
4013         cfq_rq_enqueued(cfqd, cfqq, rq);
4014 }
4015
4016 /*
4017  * Update hw_tag based on peak queue depth over 50 samples under
4018  * sufficient load.
4019  */
4020 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4021 {
4022         struct cfq_queue *cfqq = cfqd->active_queue;
4023
4024         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4025                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4026
4027         if (cfqd->hw_tag == 1)
4028                 return;
4029
4030         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4031             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4032                 return;
4033
4034         /*
4035          * If active queue hasn't enough requests and can idle, cfq might not
4036          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4037          * case
4038          */
4039         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4040             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4041             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4042                 return;
4043
4044         if (cfqd->hw_tag_samples++ < 50)
4045                 return;
4046
4047         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4048                 cfqd->hw_tag = 1;
4049         else
4050                 cfqd->hw_tag = 0;
4051 }
4052
4053 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4054 {
4055         struct cfq_io_cq *cic = cfqd->active_cic;
4056
4057         /* If the queue already has requests, don't wait */
4058         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4059                 return false;
4060
4061         /* If there are other queues in the group, don't wait */
4062         if (cfqq->cfqg->nr_cfqq > 1)
4063                 return false;
4064
4065         /* the only queue in the group, but think time is big */
4066         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4067                 return false;
4068
4069         if (cfq_slice_used(cfqq))
4070                 return true;
4071
4072         /* if slice left is less than think time, wait busy */
4073         if (cic && sample_valid(cic->ttime.ttime_samples)
4074             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4075                 return true;
4076
4077         /*
4078          * If think times is less than a jiffy than ttime_mean=0 and above
4079          * will not be true. It might happen that slice has not expired yet
4080          * but will expire soon (4-5 ns) during select_queue(). To cover the
4081          * case where think time is less than a jiffy, mark the queue wait
4082          * busy if only 1 jiffy is left in the slice.
4083          */
4084         if (cfqq->slice_end - jiffies == 1)
4085                 return true;
4086
4087         return false;
4088 }
4089
4090 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4091 {
4092         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4093         struct cfq_data *cfqd = cfqq->cfqd;
4094         const int sync = rq_is_sync(rq);
4095         unsigned long now;
4096
4097         now = jiffies;
4098         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4099                      !!(rq->cmd_flags & REQ_NOIDLE));
4100
4101         cfq_update_hw_tag(cfqd);
4102
4103         WARN_ON(!cfqd->rq_in_driver);
4104         WARN_ON(!cfqq->dispatched);
4105         cfqd->rq_in_driver--;
4106         cfqq->dispatched--;
4107         (RQ_CFQG(rq))->dispatched--;
4108         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4109                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4110
4111         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4112
4113         if (sync) {
4114                 struct cfq_rb_root *st;
4115
4116                 RQ_CIC(rq)->ttime.last_end_request = now;
4117
4118                 if (cfq_cfqq_on_rr(cfqq))
4119                         st = cfqq->service_tree;
4120                 else
4121                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4122                                         cfqq_type(cfqq));
4123
4124                 st->ttime.last_end_request = now;
4125                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4126                         cfqd->last_delayed_sync = now;
4127         }
4128
4129 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4130         cfqq->cfqg->ttime.last_end_request = now;
4131 #endif
4132
4133         /*
4134          * If this is the active queue, check if it needs to be expired,
4135          * or if we want to idle in case it has no pending requests.
4136          */
4137         if (cfqd->active_queue == cfqq) {
4138                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4139
4140                 if (cfq_cfqq_slice_new(cfqq)) {
4141                         cfq_set_prio_slice(cfqd, cfqq);
4142                         cfq_clear_cfqq_slice_new(cfqq);
4143                 }
4144
4145                 /*
4146                  * Should we wait for next request to come in before we expire
4147                  * the queue.
4148                  */
4149                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4150                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4151                         if (!cfqd->cfq_slice_idle)
4152                                 extend_sl = cfqd->cfq_group_idle;
4153                         cfqq->slice_end = jiffies + extend_sl;
4154                         cfq_mark_cfqq_wait_busy(cfqq);
4155                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4156                 }
4157
4158                 /*
4159                  * Idling is not enabled on:
4160                  * - expired queues
4161                  * - idle-priority queues
4162                  * - async queues
4163                  * - queues with still some requests queued
4164                  * - when there is a close cooperator
4165                  */
4166                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4167                         cfq_slice_expired(cfqd, 1);
4168                 else if (sync && cfqq_empty &&
4169                          !cfq_close_cooperator(cfqd, cfqq)) {
4170                         cfq_arm_slice_timer(cfqd);
4171                 }
4172         }
4173
4174         if (!cfqd->rq_in_driver)
4175                 cfq_schedule_dispatch(cfqd);
4176 }
4177
4178 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4179 {
4180         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4181                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4182                 return ELV_MQUEUE_MUST;
4183         }
4184
4185         return ELV_MQUEUE_MAY;
4186 }
4187
4188 static int cfq_may_queue(struct request_queue *q, int rw)
4189 {
4190         struct cfq_data *cfqd = q->elevator->elevator_data;
4191         struct task_struct *tsk = current;
4192         struct cfq_io_cq *cic;
4193         struct cfq_queue *cfqq;
4194
4195         /*
4196          * don't force setup of a queue from here, as a call to may_queue
4197          * does not necessarily imply that a request actually will be queued.
4198          * so just lookup a possibly existing queue, or return 'may queue'
4199          * if that fails
4200          */
4201         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4202         if (!cic)
4203                 return ELV_MQUEUE_MAY;
4204
4205         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4206         if (cfqq) {
4207                 cfq_init_prio_data(cfqq, cic);
4208
4209                 return __cfq_may_queue(cfqq);
4210         }
4211
4212         return ELV_MQUEUE_MAY;
4213 }
4214
4215 /*
4216  * queue lock held here
4217  */
4218 static void cfq_put_request(struct request *rq)
4219 {
4220         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4221
4222         if (cfqq) {
4223                 const int rw = rq_data_dir(rq);
4224
4225                 BUG_ON(!cfqq->allocated[rw]);
4226                 cfqq->allocated[rw]--;
4227
4228                 /* Put down rq reference on cfqg */
4229                 cfqg_put(RQ_CFQG(rq));
4230                 rq->elv.priv[0] = NULL;
4231                 rq->elv.priv[1] = NULL;
4232
4233                 cfq_put_queue(cfqq);
4234         }
4235 }
4236
4237 static struct cfq_queue *
4238 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4239                 struct cfq_queue *cfqq)
4240 {
4241         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4242         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4243         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4244         cfq_put_queue(cfqq);
4245         return cic_to_cfqq(cic, 1);
4246 }
4247
4248 /*
4249  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4250  * was the last process referring to said cfqq.
4251  */
4252 static struct cfq_queue *
4253 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4254 {
4255         if (cfqq_process_refs(cfqq) == 1) {
4256                 cfqq->pid = current->pid;
4257                 cfq_clear_cfqq_coop(cfqq);
4258                 cfq_clear_cfqq_split_coop(cfqq);
4259                 return cfqq;
4260         }
4261
4262         cic_set_cfqq(cic, NULL, 1);
4263
4264         cfq_put_cooperator(cfqq);
4265
4266         cfq_put_queue(cfqq);
4267         return NULL;
4268 }
4269 /*
4270  * Allocate cfq data structures associated with this request.
4271  */
4272 static int
4273 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4274                 gfp_t gfp_mask)
4275 {
4276         struct cfq_data *cfqd = q->elevator->elevator_data;
4277         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4278         const int rw = rq_data_dir(rq);
4279         const bool is_sync = rq_is_sync(rq);
4280         struct cfq_queue *cfqq;
4281
4282         might_sleep_if(gfp_mask & __GFP_WAIT);
4283
4284         spin_lock_irq(q->queue_lock);
4285
4286         check_ioprio_changed(cic, bio);
4287         check_blkcg_changed(cic, bio);
4288 new_queue:
4289         cfqq = cic_to_cfqq(cic, is_sync);
4290         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4291                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4292                 cic_set_cfqq(cic, cfqq, is_sync);
4293         } else {
4294                 /*
4295                  * If the queue was seeky for too long, break it apart.
4296                  */
4297                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4298                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4299                         cfqq = split_cfqq(cic, cfqq);
4300                         if (!cfqq)
4301                                 goto new_queue;
4302                 }
4303
4304                 /*
4305                  * Check to see if this queue is scheduled to merge with
4306                  * another, closely cooperating queue.  The merging of
4307                  * queues happens here as it must be done in process context.
4308                  * The reference on new_cfqq was taken in merge_cfqqs.
4309                  */
4310                 if (cfqq->new_cfqq)
4311                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4312         }
4313
4314         cfqq->allocated[rw]++;
4315
4316         cfqq->ref++;
4317         cfqg_get(cfqq->cfqg);
4318         rq->elv.priv[0] = cfqq;
4319         rq->elv.priv[1] = cfqq->cfqg;
4320         spin_unlock_irq(q->queue_lock);
4321         return 0;
4322 }
4323
4324 static void cfq_kick_queue(struct work_struct *work)
4325 {
4326         struct cfq_data *cfqd =
4327                 container_of(work, struct cfq_data, unplug_work);
4328         struct request_queue *q = cfqd->queue;
4329
4330         spin_lock_irq(q->queue_lock);
4331         __blk_run_queue(cfqd->queue);
4332         spin_unlock_irq(q->queue_lock);
4333 }
4334
4335 /*
4336  * Timer running if the active_queue is currently idling inside its time slice
4337  */
4338 static void cfq_idle_slice_timer(unsigned long data)
4339 {
4340         struct cfq_data *cfqd = (struct cfq_data *) data;
4341         struct cfq_queue *cfqq;
4342         unsigned long flags;
4343         int timed_out = 1;
4344
4345         cfq_log(cfqd, "idle timer fired");
4346
4347         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4348
4349         cfqq = cfqd->active_queue;
4350         if (cfqq) {
4351                 timed_out = 0;
4352
4353                 /*
4354                  * We saw a request before the queue expired, let it through
4355                  */
4356                 if (cfq_cfqq_must_dispatch(cfqq))
4357                         goto out_kick;
4358
4359                 /*
4360                  * expired
4361                  */
4362                 if (cfq_slice_used(cfqq))
4363                         goto expire;
4364
4365                 /*
4366                  * only expire and reinvoke request handler, if there are
4367                  * other queues with pending requests
4368                  */
4369                 if (!cfqd->busy_queues)
4370                         goto out_cont;
4371
4372                 /*
4373                  * not expired and it has a request pending, let it dispatch
4374                  */
4375                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4376                         goto out_kick;
4377
4378                 /*
4379                  * Queue depth flag is reset only when the idle didn't succeed
4380                  */
4381                 cfq_clear_cfqq_deep(cfqq);
4382         }
4383 expire:
4384         cfq_slice_expired(cfqd, timed_out);
4385 out_kick:
4386         cfq_schedule_dispatch(cfqd);
4387 out_cont:
4388         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4389 }
4390
4391 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4392 {
4393         del_timer_sync(&cfqd->idle_slice_timer);
4394         cancel_work_sync(&cfqd->unplug_work);
4395 }
4396
4397 static void cfq_put_async_queues(struct cfq_data *cfqd)
4398 {
4399         int i;
4400
4401         for (i = 0; i < IOPRIO_BE_NR; i++) {
4402                 if (cfqd->async_cfqq[0][i])
4403                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4404                 if (cfqd->async_cfqq[1][i])
4405                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4406         }
4407
4408         if (cfqd->async_idle_cfqq)
4409                 cfq_put_queue(cfqd->async_idle_cfqq);
4410 }
4411
4412 static void cfq_exit_queue(struct elevator_queue *e)
4413 {
4414         struct cfq_data *cfqd = e->elevator_data;
4415         struct request_queue *q = cfqd->queue;
4416
4417         cfq_shutdown_timer_wq(cfqd);
4418
4419         spin_lock_irq(q->queue_lock);
4420
4421         if (cfqd->active_queue)
4422                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4423
4424         cfq_put_async_queues(cfqd);
4425
4426         spin_unlock_irq(q->queue_lock);
4427
4428         cfq_shutdown_timer_wq(cfqd);
4429
4430 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4431         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4432 #else
4433         kfree(cfqd->root_group);
4434 #endif
4435         kfree(cfqd);
4436 }
4437
4438 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4439 {
4440         struct cfq_data *cfqd;
4441         struct blkcg_gq *blkg __maybe_unused;
4442         int i, ret;
4443         struct elevator_queue *eq;
4444
4445         eq = elevator_alloc(q, e);
4446         if (!eq)
4447                 return -ENOMEM;
4448
4449         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4450         if (!cfqd) {
4451                 kobject_put(&eq->kobj);
4452                 return -ENOMEM;
4453         }
4454         eq->elevator_data = cfqd;
4455
4456         cfqd->queue = q;
4457         spin_lock_irq(q->queue_lock);
4458         q->elevator = eq;
4459         spin_unlock_irq(q->queue_lock);
4460
4461         /* Init root service tree */
4462         cfqd->grp_service_tree = CFQ_RB_ROOT;
4463
4464         /* Init root group and prefer root group over other groups by default */
4465 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4466         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4467         if (ret)
4468                 goto out_free;
4469
4470         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4471 #else
4472         ret = -ENOMEM;
4473         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4474                                         GFP_KERNEL, cfqd->queue->node);
4475         if (!cfqd->root_group)
4476                 goto out_free;
4477
4478         cfq_init_cfqg_base(cfqd->root_group);
4479 #endif
4480         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4481         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4482
4483         /*
4484          * Not strictly needed (since RB_ROOT just clears the node and we
4485          * zeroed cfqd on alloc), but better be safe in case someone decides
4486          * to add magic to the rb code
4487          */
4488         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4489                 cfqd->prio_trees[i] = RB_ROOT;
4490
4491         /*
4492          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4493          * Grab a permanent reference to it, so that the normal code flow
4494          * will not attempt to free it.  oom_cfqq is linked to root_group
4495          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4496          * the reference from linking right away.
4497          */
4498         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4499         cfqd->oom_cfqq.ref++;
4500
4501         spin_lock_irq(q->queue_lock);
4502         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4503         cfqg_put(cfqd->root_group);
4504         spin_unlock_irq(q->queue_lock);
4505
4506         init_timer(&cfqd->idle_slice_timer);
4507         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4508         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4509
4510         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4511
4512         cfqd->cfq_quantum = cfq_quantum;
4513         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4514         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4515         cfqd->cfq_back_max = cfq_back_max;
4516         cfqd->cfq_back_penalty = cfq_back_penalty;
4517         cfqd->cfq_slice[0] = cfq_slice_async;
4518         cfqd->cfq_slice[1] = cfq_slice_sync;
4519         cfqd->cfq_target_latency = cfq_target_latency;
4520         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4521         cfqd->cfq_slice_idle = cfq_slice_idle;
4522         cfqd->cfq_group_idle = cfq_group_idle;
4523         cfqd->cfq_latency = 1;
4524         cfqd->hw_tag = -1;
4525         /*
4526          * we optimistically start assuming sync ops weren't delayed in last
4527          * second, in order to have larger depth for async operations.
4528          */
4529         cfqd->last_delayed_sync = jiffies - HZ;
4530         return 0;
4531
4532 out_free:
4533         kfree(cfqd);
4534         kobject_put(&eq->kobj);
4535         return ret;
4536 }
4537
4538 static void cfq_registered_queue(struct request_queue *q)
4539 {
4540         struct elevator_queue *e = q->elevator;
4541         struct cfq_data *cfqd = e->elevator_data;
4542
4543         /*
4544          * Default to IOPS mode with no idling for SSDs
4545          */
4546         if (blk_queue_nonrot(q))
4547                 cfqd->cfq_slice_idle = 0;
4548 }
4549
4550 /*
4551  * sysfs parts below -->
4552  */
4553 static ssize_t
4554 cfq_var_show(unsigned int var, char *page)
4555 {
4556         return sprintf(page, "%u\n", var);
4557 }
4558
4559 static ssize_t
4560 cfq_var_store(unsigned int *var, const char *page, size_t count)
4561 {
4562         char *p = (char *) page;
4563
4564         *var = simple_strtoul(p, &p, 10);
4565         return count;
4566 }
4567
4568 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4569 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4570 {                                                                       \
4571         struct cfq_data *cfqd = e->elevator_data;                       \
4572         unsigned int __data = __VAR;                                    \
4573         if (__CONV)                                                     \
4574                 __data = jiffies_to_msecs(__data);                      \
4575         return cfq_var_show(__data, (page));                            \
4576 }
4577 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4578 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4579 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4580 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4581 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4582 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4583 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4584 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4585 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4586 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4587 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4588 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4589 #undef SHOW_FUNCTION
4590
4591 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4592 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4593 {                                                                       \
4594         struct cfq_data *cfqd = e->elevator_data;                       \
4595         unsigned int __data;                                            \
4596         int ret = cfq_var_store(&__data, (page), count);                \
4597         if (__data < (MIN))                                             \
4598                 __data = (MIN);                                         \
4599         else if (__data > (MAX))                                        \
4600                 __data = (MAX);                                         \
4601         if (__CONV)                                                     \
4602                 *(__PTR) = msecs_to_jiffies(__data);                    \
4603         else                                                            \
4604                 *(__PTR) = __data;                                      \
4605         return ret;                                                     \
4606 }
4607 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4608 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4609                 UINT_MAX, 1);
4610 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4611                 UINT_MAX, 1);
4612 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4613 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4614                 UINT_MAX, 0);
4615 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4616 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4617 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4618 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4619 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4620                 UINT_MAX, 0);
4621 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4622 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4623 #undef STORE_FUNCTION
4624
4625 #define CFQ_ATTR(name) \
4626         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4627
4628 static struct elv_fs_entry cfq_attrs[] = {
4629         CFQ_ATTR(quantum),
4630         CFQ_ATTR(fifo_expire_sync),
4631         CFQ_ATTR(fifo_expire_async),
4632         CFQ_ATTR(back_seek_max),
4633         CFQ_ATTR(back_seek_penalty),
4634         CFQ_ATTR(slice_sync),
4635         CFQ_ATTR(slice_async),
4636         CFQ_ATTR(slice_async_rq),
4637         CFQ_ATTR(slice_idle),
4638         CFQ_ATTR(group_idle),
4639         CFQ_ATTR(low_latency),
4640         CFQ_ATTR(target_latency),
4641         __ATTR_NULL
4642 };
4643
4644 static struct elevator_type iosched_cfq = {
4645         .ops = {
4646                 .elevator_merge_fn =            cfq_merge,
4647                 .elevator_merged_fn =           cfq_merged_request,
4648                 .elevator_merge_req_fn =        cfq_merged_requests,
4649                 .elevator_allow_merge_fn =      cfq_allow_merge,
4650                 .elevator_bio_merged_fn =       cfq_bio_merged,
4651                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4652                 .elevator_add_req_fn =          cfq_insert_request,
4653                 .elevator_activate_req_fn =     cfq_activate_request,
4654                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4655                 .elevator_completed_req_fn =    cfq_completed_request,
4656                 .elevator_former_req_fn =       elv_rb_former_request,
4657                 .elevator_latter_req_fn =       elv_rb_latter_request,
4658                 .elevator_init_icq_fn =         cfq_init_icq,
4659                 .elevator_exit_icq_fn =         cfq_exit_icq,
4660                 .elevator_set_req_fn =          cfq_set_request,
4661                 .elevator_put_req_fn =          cfq_put_request,
4662                 .elevator_may_queue_fn =        cfq_may_queue,
4663                 .elevator_init_fn =             cfq_init_queue,
4664                 .elevator_exit_fn =             cfq_exit_queue,
4665                 .elevator_registered_fn =       cfq_registered_queue,
4666         },
4667         .icq_size       =       sizeof(struct cfq_io_cq),
4668         .icq_align      =       __alignof__(struct cfq_io_cq),
4669         .elevator_attrs =       cfq_attrs,
4670         .elevator_name  =       "cfq",
4671         .elevator_owner =       THIS_MODULE,
4672 };
4673
4674 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4675 static struct blkcg_policy blkcg_policy_cfq = {
4676         .pd_size                = sizeof(struct cfq_group),
4677         .cpd_size               = sizeof(struct cfq_group_data),
4678         .cftypes                = cfq_blkcg_files,
4679
4680         .cpd_init_fn            = cfq_cpd_init,
4681         .pd_init_fn             = cfq_pd_init,
4682         .pd_offline_fn          = cfq_pd_offline,
4683         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4684 };
4685 #endif
4686
4687 static int __init cfq_init(void)
4688 {
4689         int ret;
4690
4691         /*
4692          * could be 0 on HZ < 1000 setups
4693          */
4694         if (!cfq_slice_async)
4695                 cfq_slice_async = 1;
4696         if (!cfq_slice_idle)
4697                 cfq_slice_idle = 1;
4698
4699 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4700         if (!cfq_group_idle)
4701                 cfq_group_idle = 1;
4702
4703         ret = blkcg_policy_register(&blkcg_policy_cfq);
4704         if (ret)
4705                 return ret;
4706 #else
4707         cfq_group_idle = 0;
4708 #endif
4709
4710         ret = -ENOMEM;
4711         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4712         if (!cfq_pool)
4713                 goto err_pol_unreg;
4714
4715         ret = elv_register(&iosched_cfq);
4716         if (ret)
4717                 goto err_free_pool;
4718
4719         return 0;
4720
4721 err_free_pool:
4722         kmem_cache_destroy(cfq_pool);
4723 err_pol_unreg:
4724 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4725         blkcg_policy_unregister(&blkcg_policy_cfq);
4726 #endif
4727         return ret;
4728 }
4729
4730 static void __exit cfq_exit(void)
4731 {
4732 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4733         blkcg_policy_unregister(&blkcg_policy_cfq);
4734 #endif
4735         elv_unregister(&iosched_cfq);
4736         kmem_cache_destroy(cfq_pool);
4737 }
4738
4739 module_init(cfq_init);
4740 module_exit(cfq_exit);
4741
4742 MODULE_AUTHOR("Jens Axboe");
4743 MODULE_LICENSE("GPL");
4744 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");