block/cfq-iosched.c

   1 /*
   2  *  CFQ, or complete fairness queueing, disk scheduler.
   3  *
   4  *  Based on ideas from a previously unfinished io
   5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
   6  *
   7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
   8  */
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/elevator.h>
  13 #include <linux/jiffies.h>
  14 #include <linux/rbtree.h>
  15 #include <linux/ioprio.h>
  16 #include <linux/blktrace_api.h>
  17 #include <linux/blk-cgroup.h>
  18 #include "blk.h"
  19
  20 /*
  21  * tunables
  22  */
  23 /* max queue in one round of service */
  24 static const int cfq_quantum = 8;
  25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  26 /* maximum backwards seek, in KiB */
  27 static const int cfq_back_max = 16 * 1024;
  28 /* penalty of a backwards seek */
  29 static const int cfq_back_penalty = 2;
  30 static const int cfq_slice_sync = HZ / 10;
  31 static int cfq_slice_async = HZ / 25;
  32 static const int cfq_slice_async_rq = 2;
  33 static int cfq_slice_idle = HZ / 125;
  34 static int cfq_group_idle = HZ / 125;
  35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
  36 static const int cfq_hist_divisor = 4;
  37
  38 /*
  39  * offset from end of service tree
  40  */
  41 #define CFQ_IDLE_DELAY          (HZ / 5)
  42
  43 /*
  44  * below this threshold, we consider thinktime immediate
  45  */
  46 #define CFQ_MIN_TT              (2)
  47
  48 #define CFQ_SLICE_SCALE         (5)
  49 #define CFQ_HW_QUEUE_MIN        (5)
  50 #define CFQ_SERVICE_SHIFT       12
  51
  52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
  53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
  54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
  55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
  56
  57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
  58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
  59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
  60
  61 static struct kmem_cache *cfq_pool;
  62
  63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  66
  67 #define sample_valid(samples)   ((samples) > 80)
  68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  69
  70 /* blkio-related constants */
  71 #define CFQ_WEIGHT_MIN          10
  72 #define CFQ_WEIGHT_MAX          1000
  73 #define CFQ_WEIGHT_DEFAULT      500
  74
  75 struct cfq_ttime {
  76         unsigned long last_end_request;
  77
  78         unsigned long ttime_total;
  79         unsigned long ttime_samples;
  80         unsigned long ttime_mean;
  81 };
  82
  83 /*
  84  * Most of our rbtree usage is for sorting with min extraction, so
  85  * if we cache the leftmost node we don't have to walk down the tree
  86  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  87  * move this into the elevator for the rq sorting as well.
  88  */
  89 struct cfq_rb_root {
  90         struct rb_root rb;
  91         struct rb_node *left;
  92         unsigned count;
  93         u64 min_vdisktime;
  94         struct cfq_ttime ttime;
  95 };
  96 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
  97                         .ttime = {.last_end_request = jiffies,},}
  98
  99 /*
 100  * Per process-grouping structure
 101  */
 102 struct cfq_queue {
 103         /* reference count */
 104         int ref;
 105         /* various state flags, see below */
 106         unsigned int flags;
 107         /* parent cfq_data */
 108         struct cfq_data *cfqd;
 109         /* service_tree member */
 110         struct rb_node rb_node;
 111         /* service_tree key */
 112         unsigned long rb_key;
 113         /* prio tree member */
 114         struct rb_node p_node;
 115         /* prio tree root we belong to, if any */
 116         struct rb_root *p_root;
 117         /* sorted list of pending requests */
 118         struct rb_root sort_list;
 119         /* if fifo isn't expired, next request to serve */
 120         struct request *next_rq;
 121         /* requests queued in sort_list */
 122         int queued[2];
 123         /* currently allocated requests */
 124         int allocated[2];
 125         /* fifo list of requests in sort_list */
 126         struct list_head fifo;
 127
 128         /* time when queue got scheduled in to dispatch first request. */
 129         unsigned long dispatch_start;
 130         unsigned int allocated_slice;
 131         unsigned int slice_dispatch;
 132         /* time when first request from queue completed and slice started. */
 133         unsigned long slice_start;
 134         unsigned long slice_end;
 135         long slice_resid;
 136
 137         /* pending priority requests */
 138         int prio_pending;
 139         /* number of requests that are on the dispatch list or inside driver */
 140         int dispatched;
 141
 142         /* io prio of this group */
 143         unsigned short ioprio, org_ioprio;
 144         unsigned short ioprio_class;
 145
 146         pid_t pid;
 147
 148         u32 seek_history;
 149         sector_t last_request_pos;
 150
 151         struct cfq_rb_root *service_tree;
 152         struct cfq_queue *new_cfqq;
 153         struct cfq_group *cfqg;
 154         /* Number of sectors dispatched from queue in single dispatch round */
 155         unsigned long nr_sectors;
 156 };
 157
 158 /*
 159  * First index in the service_trees.
 160  * IDLE is handled separately, so it has negative index
 161  */
 162 enum wl_class_t {
 163         BE_WORKLOAD = 0,
 164         RT_WORKLOAD = 1,
 165         IDLE_WORKLOAD = 2,
 166         CFQ_PRIO_NR,
 167 };
 168
 169 /*
 170  * Second index in the service_trees.
 171  */
 172 enum wl_type_t {
 173         ASYNC_WORKLOAD = 0,
 174         SYNC_NOIDLE_WORKLOAD = 1,
 175         SYNC_WORKLOAD = 2
 176 };
 177
 178 struct cfqg_stats {
 179 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 180         /* number of ios merged */
 181         struct blkg_rwstat              merged;
 182         /* total time spent on device in ns, may not be accurate w/ queueing */
 183         struct blkg_rwstat              service_time;
 184         /* total time spent waiting in scheduler queue in ns */
 185         struct blkg_rwstat              wait_time;
 186         /* number of IOs queued up */
 187         struct blkg_rwstat              queued;
 188         /* total sectors transferred */
 189         struct blkg_stat                sectors;
 190         /* total disk time and nr sectors dispatched by this group */
 191         struct blkg_stat                time;
 192 #ifdef CONFIG_DEBUG_BLK_CGROUP
 193         /* time not charged to this cgroup */
 194         struct blkg_stat                unaccounted_time;
 195         /* sum of number of ios queued across all samples */
 196         struct blkg_stat                avg_queue_size_sum;
 197         /* count of samples taken for average */
 198         struct blkg_stat                avg_queue_size_samples;
 199         /* how many times this group has been removed from service tree */
 200         struct blkg_stat                dequeue;
 201         /* total time spent waiting for it to be assigned a timeslice. */
 202         struct blkg_stat                group_wait_time;
 203         /* time spent idling for this blkcg_gq */
 204         struct blkg_stat                idle_time;
 205         /* total time with empty current active q with other requests queued */
 206         struct blkg_stat                empty_time;
 207         /* fields after this shouldn't be cleared on stat reset */
 208         uint64_t                        start_group_wait_time;
 209         uint64_t                        start_idle_time;
 210         uint64_t                        start_empty_time;
 211         uint16_t                        flags;
 212 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
 213 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 214 };
 215
 216 /* Per-cgroup data */
 217 struct cfq_group_data {
 218         /* must be the first member */
 219         struct blkcg_policy_data cpd;
 220
 221         unsigned int weight;
 222         unsigned int leaf_weight;
 223 };
 224
 225 /* This is per cgroup per device grouping structure */
 226 struct cfq_group {
 227         /* must be the first member */
 228         struct blkg_policy_data pd;
 229
 230         /* group service_tree member */
 231         struct rb_node rb_node;
 232
 233         /* group service_tree key */
 234         u64 vdisktime;
 235
 236         /*
 237          * The number of active cfqgs and sum of their weights under this
 238          * cfqg.  This covers this cfqg's leaf_weight and all children's
 239          * weights, but does not cover weights of further descendants.
 240          *
 241          * If a cfqg is on the service tree, it's active.  An active cfqg
 242          * also activates its parent and contributes to the children_weight
 243          * of the parent.
 244          */
 245         int nr_active;
 246         unsigned int children_weight;
 247
 248         /*
 249          * vfraction is the fraction of vdisktime that the tasks in this
 250          * cfqg are entitled to.  This is determined by compounding the
 251          * ratios walking up from this cfqg to the root.
 252          *
 253          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
 254          * vfractions on a service tree is approximately 1.  The sum may
 255          * deviate a bit due to rounding errors and fluctuations caused by
 256          * cfqgs entering and leaving the service tree.
 257          */
 258         unsigned int vfraction;
 259
 260         /*
 261          * There are two weights - (internal) weight is the weight of this
 262          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
 263          * this cfqg against the child cfqgs.  For the root cfqg, both
 264          * weights are kept in sync for backward compatibility.
 265          */
 266         unsigned int weight;
 267         unsigned int new_weight;
 268         unsigned int dev_weight;
 269
 270         unsigned int leaf_weight;
 271         unsigned int new_leaf_weight;
 272         unsigned int dev_leaf_weight;
 273
 274         /* number of cfqq currently on this group */
 275         int nr_cfqq;
 276
 277         /*
 278          * Per group busy queues average. Useful for workload slice calc. We
 279          * create the array for each prio class but at run time it is used
 280          * only for RT and BE class and slot for IDLE class remains unused.
 281          * This is primarily done to avoid confusion and a gcc warning.
 282          */
 283         unsigned int busy_queues_avg[CFQ_PRIO_NR];
 284         /*
 285          * rr lists of queues with requests. We maintain service trees for
 286          * RT and BE classes. These trees are subdivided in subclasses
 287          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
 288          * class there is no subclassification and all the cfq queues go on
 289          * a single tree service_tree_idle.
 290          * Counts are embedded in the cfq_rb_root
 291          */
 292         struct cfq_rb_root service_trees[2][3];
 293         struct cfq_rb_root service_tree_idle;
 294
 295         unsigned long saved_wl_slice;
 296         enum wl_type_t saved_wl_type;
 297         enum wl_class_t saved_wl_class;
 298
 299         /* number of requests that are on the dispatch list or inside driver */
 300         int dispatched;
 301         struct cfq_ttime ttime;
 302         struct cfqg_stats stats;        /* stats for this cfqg */
 303
 304         /* async queue for each priority case */
 305         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 306         struct cfq_queue *async_idle_cfqq;
 307
 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         sector_t last_position;
 374
 375         /*
 376          * tunables, see top of file
 377          */
 378         unsigned int cfq_quantum;
 379         unsigned int cfq_fifo_expire[2];
 380         unsigned int cfq_back_penalty;
 381         unsigned int cfq_back_max;
 382         unsigned int cfq_slice[2];
 383         unsigned int cfq_slice_async_rq;
 384         unsigned int cfq_slice_idle;
 385         unsigned int cfq_group_idle;
 386         unsigned int cfq_latency;
 387         unsigned int cfq_target_latency;
 388
 389         /*
 390          * Fallback dummy cfqq for extreme OOM conditions
 391          */
 392         struct cfq_queue oom_cfqq;
 393
 394         unsigned long last_delayed_sync;
 395 };
 396
 397 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 398 static void cfq_put_queue(struct cfq_queue *cfqq);
 399
 400 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
 401                                             enum wl_class_t class,
 402                                             enum wl_type_t type)
 403 {
 404         if (!cfqg)
 405                 return NULL;
 406
 407         if (class == IDLE_WORKLOAD)
 408                 return &cfqg->service_tree_idle;
 409
 410         return &cfqg->service_trees[class][type];
 411 }
 412
 413 enum cfqq_state_flags {
 414         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 415         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 416         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 417         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 418         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 419         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 420         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 421         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 422         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 423         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 424         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
 425         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 426         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 427 };
 428
 429 #define CFQ_CFQQ_FNS(name)                                              \
 430 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 431 {                                                                       \
 432         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 433 }                                                                       \
 434 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 435 {                                                                       \
 436         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 437 }                                                                       \
 438 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 439 {                                                                       \
 440         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 441 }
 442
 443 CFQ_CFQQ_FNS(on_rr);
 444 CFQ_CFQQ_FNS(wait_request);
 445 CFQ_CFQQ_FNS(must_dispatch);
 446 CFQ_CFQQ_FNS(must_alloc_slice);
 447 CFQ_CFQQ_FNS(fifo_expire);
 448 CFQ_CFQQ_FNS(idle_window);
 449 CFQ_CFQQ_FNS(prio_changed);
 450 CFQ_CFQQ_FNS(slice_new);
 451 CFQ_CFQQ_FNS(sync);
 452 CFQ_CFQQ_FNS(coop);
 453 CFQ_CFQQ_FNS(split_coop);
 454 CFQ_CFQQ_FNS(deep);
 455 CFQ_CFQQ_FNS(wait_busy);
 456 #undef CFQ_CFQQ_FNS
 457
 458 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
 459
 460 /* cfqg stats flags */
 461 enum cfqg_stats_flags {
 462         CFQG_stats_waiting = 0,
 463         CFQG_stats_idling,
 464         CFQG_stats_empty,
 465 };
 466
 467 #define CFQG_FLAG_FNS(name)                                             \
 468 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
 469 {                                                                       \
 470         stats->flags |= (1 << CFQG_stats_##name);                       \
 471 }                                                                       \
 472 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
 473 {                                                                       \
 474         stats->flags &= ~(1 << CFQG_stats_##name);                      \
 475 }                                                                       \
 476 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
 477 {                                                                       \
 478         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
 479 }                                                                       \
 480
 481 CFQG_FLAG_FNS(waiting)
 482 CFQG_FLAG_FNS(idling)
 483 CFQG_FLAG_FNS(empty)
 484 #undef CFQG_FLAG_FNS
 485
 486 /* This should be called with the queue_lock held. */
 487 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
 488 {
 489         unsigned long long now;
 490
 491         if (!cfqg_stats_waiting(stats))
 492                 return;
 493
 494         now = sched_clock();
 495         if (time_after64(now, stats->start_group_wait_time))
 496                 blkg_stat_add(&stats->group_wait_time,
 497                               now - stats->start_group_wait_time);
 498         cfqg_stats_clear_waiting(stats);
 499 }
 500
 501 /* This should be called with the queue_lock held. */
 502 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
 503                                                  struct cfq_group *curr_cfqg)
 504 {
 505         struct cfqg_stats *stats = &cfqg->stats;
 506
 507         if (cfqg_stats_waiting(stats))
 508                 return;
 509         if (cfqg == curr_cfqg)
 510                 return;
 511         stats->start_group_wait_time = sched_clock();
 512         cfqg_stats_mark_waiting(stats);
 513 }
 514
 515 /* This should be called with the queue_lock held. */
 516 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
 517 {
 518         unsigned long long now;
 519
 520         if (!cfqg_stats_empty(stats))
 521                 return;
 522
 523         now = sched_clock();
 524         if (time_after64(now, stats->start_empty_time))
 525                 blkg_stat_add(&stats->empty_time,
 526                               now - stats->start_empty_time);
 527         cfqg_stats_clear_empty(stats);
 528 }
 529
 530 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
 531 {
 532         blkg_stat_add(&cfqg->stats.dequeue, 1);
 533 }
 534
 535 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
 536 {
 537         struct cfqg_stats *stats = &cfqg->stats;
 538
 539         if (blkg_rwstat_total(&stats->queued))
 540                 return;
 541
 542         /*
 543          * group is already marked empty. This can happen if cfqq got new
 544          * request in parent group and moved to this group while being added
 545          * to service tree. Just ignore the event and move on.
 546          */
 547         if (cfqg_stats_empty(stats))
 548                 return;
 549
 550         stats->start_empty_time = sched_clock();
 551         cfqg_stats_mark_empty(stats);
 552 }
 553
 554 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
 555 {
 556         struct cfqg_stats *stats = &cfqg->stats;
 557
 558         if (cfqg_stats_idling(stats)) {
 559                 unsigned long long now = sched_clock();
 560
 561                 if (time_after64(now, stats->start_idle_time))
 562                         blkg_stat_add(&stats->idle_time,
 563                                       now - stats->start_idle_time);
 564                 cfqg_stats_clear_idling(stats);
 565         }
 566 }
 567
 568 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
 569 {
 570         struct cfqg_stats *stats = &cfqg->stats;
 571
 572         BUG_ON(cfqg_stats_idling(stats));
 573
 574         stats->start_idle_time = sched_clock();
 575         cfqg_stats_mark_idling(stats);
 576 }
 577
 578 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
 579 {
 580         struct cfqg_stats *stats = &cfqg->stats;
 581
 582         blkg_stat_add(&stats->avg_queue_size_sum,
 583                       blkg_rwstat_total(&stats->queued));
 584         blkg_stat_add(&stats->avg_queue_size_samples, 1);
 585         cfqg_stats_update_group_wait_time(stats);
 586 }
 587
 588 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 589
 590 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
 591 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
 592 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
 593 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
 594 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
 595 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
 596 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
 597
 598 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 599
 600 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 601
 602 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
 603 {
 604         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
 605 }
 606
 607 static struct cfq_group_data
 608 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
 609 {
 610         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
 611 }
 612
 613 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
 614 {
 615         return pd_to_blkg(&cfqg->pd);
 616 }
 617
 618 static struct blkcg_policy blkcg_policy_cfq;
 619
 620 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
 621 {
 622         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
 623 }
 624
 625 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
 626 {
 627         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
 628 }
 629
 630 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
 631 {
 632         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
 633
 634         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
 635 }
 636
 637 static inline void cfqg_get(struct cfq_group *cfqg)
 638 {
 639         return blkg_get(cfqg_to_blkg(cfqg));
 640 }
 641
 642 static inline void cfqg_put(struct cfq_group *cfqg)
 643 {
 644         return blkg_put(cfqg_to_blkg(cfqg));
 645 }
 646
 647 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
 648         char __pbuf[128];                                               \
 649                                                                         \
 650         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
 651         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
 652                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 653                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 654                           __pbuf, ##args);                              \
 655 } while (0)
 656
 657 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
 658         char __pbuf[128];                                               \
 659                                                                         \
 660         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
 661         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
 662 } while (0)
 663
 664 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 665                                             struct cfq_group *curr_cfqg, int rw)
 666 {
 667         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
 668         cfqg_stats_end_empty_time(&cfqg->stats);
 669         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
 670 }
 671
 672 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 673                         unsigned long time, unsigned long unaccounted_time)
 674 {
 675         blkg_stat_add(&cfqg->stats.time, time);
 676 #ifdef CONFIG_DEBUG_BLK_CGROUP
 677         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
 678 #endif
 679 }
 680
 681 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
 682 {
 683         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
 684 }
 685
 686 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
 687 {
 688         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
 689 }
 690
 691 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 692                                               uint64_t bytes, int rw)
 693 {
 694         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
 695 }
 696
 697 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 698                         uint64_t start_time, uint64_t io_start_time, int rw)
 699 {
 700         struct cfqg_stats *stats = &cfqg->stats;
 701         unsigned long long now = sched_clock();
 702
 703         if (time_after64(now, io_start_time))
 704                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
 705         if (time_after64(io_start_time, start_time))
 706                 blkg_rwstat_add(&stats->wait_time, rw,
 707                                 io_start_time - start_time);
 708 }
 709
 710 /* @stats = 0 */
 711 static void cfqg_stats_reset(struct cfqg_stats *stats)
 712 {
 713         /* queued stats shouldn't be cleared */
 714         blkg_rwstat_reset(&stats->merged);
 715         blkg_rwstat_reset(&stats->service_time);
 716         blkg_rwstat_reset(&stats->wait_time);
 717         blkg_stat_reset(&stats->time);
 718 #ifdef CONFIG_DEBUG_BLK_CGROUP
 719         blkg_stat_reset(&stats->unaccounted_time);
 720         blkg_stat_reset(&stats->avg_queue_size_sum);
 721         blkg_stat_reset(&stats->avg_queue_size_samples);
 722         blkg_stat_reset(&stats->dequeue);
 723         blkg_stat_reset(&stats->group_wait_time);
 724         blkg_stat_reset(&stats->idle_time);
 725         blkg_stat_reset(&stats->empty_time);
 726 #endif
 727 }
 728
 729 /* @to += @from */
 730 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
 731 {
 732         /* queued stats shouldn't be cleared */
 733         blkg_rwstat_add_aux(&to->merged, &from->merged);
 734         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 735         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 736         blkg_stat_add_aux(&from->time, &from->time);
 737 #ifdef CONFIG_DEBUG_BLK_CGROUP
 738         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
 739         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 740         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
 741         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 742         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 743         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 744         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 745 #endif
 746 }
 747
 748 /*
 749  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
 750  * recursive stats can still account for the amount used by this cfqg after
 751  * it's gone.
 752  */
 753 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
 754 {
 755         struct cfq_group *parent = cfqg_parent(cfqg);
 756
 757         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
 758
 759         if (unlikely(!parent))
 760                 return;
 761
 762         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
 763         cfqg_stats_reset(&cfqg->stats);
 764 }
 765
 766 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
 767
 768 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
 769 static inline void cfqg_get(struct cfq_group *cfqg) { }
 770 static inline void cfqg_put(struct cfq_group *cfqg) { }
 771
 772 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 773         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
 774                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
 775                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
 776                                 ##args)
 777 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
 778
 779 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
 780                         struct cfq_group *curr_cfqg, int rw) { }
 781 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
 782                         unsigned long time, unsigned long unaccounted_time) { }
 783 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
 784 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
 785 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
 786                                               uint64_t bytes, int rw) { }
 787 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
 788                         uint64_t start_time, uint64_t io_start_time, int rw) { }
 789
 790 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
 791
 792 #define cfq_log(cfqd, fmt, args...)     \
 793         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 794
 795 /* Traverses through cfq group service trees */
 796 #define for_each_cfqg_st(cfqg, i, j, st) \
 797         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 798                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 799                         : &cfqg->service_tree_idle; \
 800                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 801                         (i == IDLE_WORKLOAD && j == 0); \
 802                         j++, st = i < IDLE_WORKLOAD ? \
 803                         &cfqg->service_trees[i][j]: NULL) \
 804
 805 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
 806         struct cfq_ttime *ttime, bool group_idle)
 807 {
 808         unsigned long slice;
 809         if (!sample_valid(ttime->ttime_samples))
 810                 return false;
 811         if (group_idle)
 812                 slice = cfqd->cfq_group_idle;
 813         else
 814                 slice = cfqd->cfq_slice_idle;
 815         return ttime->ttime_mean > slice;
 816 }
 817
 818 static inline bool iops_mode(struct cfq_data *cfqd)
 819 {
 820         /*
 821          * If we are not idling on queues and it is a NCQ drive, parallel
 822          * execution of requests is on and measuring time is not possible
 823          * in most of the cases until and unless we drive shallower queue
 824          * depths and that becomes a performance bottleneck. In such cases
 825          * switch to start providing fairness in terms of number of IOs.
 826          */
 827         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
 828                 return true;
 829         else
 830                 return false;
 831 }
 832
 833 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
 834 {
 835         if (cfq_class_idle(cfqq))
 836                 return IDLE_WORKLOAD;
 837         if (cfq_class_rt(cfqq))
 838                 return RT_WORKLOAD;
 839         return BE_WORKLOAD;
 840 }
 841
 842
 843 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 844 {
 845         if (!cfq_cfqq_sync(cfqq))
 846                 return ASYNC_WORKLOAD;
 847         if (!cfq_cfqq_idle_window(cfqq))
 848                 return SYNC_NOIDLE_WORKLOAD;
 849         return SYNC_WORKLOAD;
 850 }
 851
 852 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
 853                                         struct cfq_data *cfqd,
 854                                         struct cfq_group *cfqg)
 855 {
 856         if (wl_class == IDLE_WORKLOAD)
 857                 return cfqg->service_tree_idle.count;
 858
 859         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
 860                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
 861                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
 862 }
 863
 864 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 865                                         struct cfq_group *cfqg)
 866 {
 867         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
 868                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 869 }
 870
 871 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 872 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
 873                                        struct cfq_io_cq *cic, struct bio *bio);
 874
 875 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
 876 {
 877         /* cic->icq is the first member, %NULL will convert to %NULL */
 878         return container_of(icq, struct cfq_io_cq, icq);
 879 }
 880
 881 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
 882                                                struct io_context *ioc)
 883 {
 884         if (ioc)
 885                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
 886         return NULL;
 887 }
 888
 889 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
 890 {
 891         return cic->cfqq[is_sync];
 892 }
 893
 894 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
 895                                 bool is_sync)
 896 {
 897         cic->cfqq[is_sync] = cfqq;
 898 }
 899
 900 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
 901 {
 902         return cic->icq.q->elevator->elevator_data;
 903 }
 904
 905 /*
 906  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 907  * set (in which case it could also be direct WRITE).
 908  */
 909 static inline bool cfq_bio_sync(struct bio *bio)
 910 {
 911         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
 912 }
 913
 914 /*
 915  * scheduler run of queue, if there are requests pending and no one in the
 916  * driver that will restart queueing
 917  */
 918 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 919 {
 920         if (cfqd->busy_queues) {
 921                 cfq_log(cfqd, "schedule dispatch");
 922                 kblockd_schedule_work(&cfqd->unplug_work);
 923         }
 924 }
 925
 926 /*
 927  * Scale schedule slice based on io priority. Use the sync time slice only
 928  * if a queue is marked sync and has sync io queued. A sync queue with async
 929  * io only, should not get full sync slice length.
 930  */
 931 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 932                                  unsigned short prio)
 933 {
 934         const int base_slice = cfqd->cfq_slice[sync];
 935
 936         WARN_ON(prio >= IOPRIO_BE_NR);
 937
 938         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 939 }
 940
 941 static inline int
 942 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 943 {
 944         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 945 }
 946
 947 /**
 948  * cfqg_scale_charge - scale disk time charge according to cfqg weight
 949  * @charge: disk time being charged
 950  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
 951  *
 952  * Scale @charge according to @vfraction, which is in range (0, 1].  The
 953  * scaling is inversely proportional.
 954  *
 955  * scaled = charge / vfraction
 956  *
 957  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
 958  */
 959 static inline u64 cfqg_scale_charge(unsigned long charge,
 960                                     unsigned int vfraction)
 961 {
 962         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
 963
 964         /* charge / vfraction */
 965         c <<= CFQ_SERVICE_SHIFT;
 966         do_div(c, vfraction);
 967         return c;
 968 }
 969
 970 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 971 {
 972         s64 delta = (s64)(vdisktime - min_vdisktime);
 973         if (delta > 0)
 974                 min_vdisktime = vdisktime;
 975
 976         return min_vdisktime;
 977 }
 978
 979 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 980 {
 981         s64 delta = (s64)(vdisktime - min_vdisktime);
 982         if (delta < 0)
 983                 min_vdisktime = vdisktime;
 984
 985         return min_vdisktime;
 986 }
 987
 988 static void update_min_vdisktime(struct cfq_rb_root *st)
 989 {
 990         struct cfq_group *cfqg;
 991
 992         if (st->left) {
 993                 cfqg = rb_entry_cfqg(st->left);
 994                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
 995                                                   cfqg->vdisktime);
 996         }
 997 }
 998
 999 /*
1000  * get averaged number of queues of RT/BE priority.
1001  * average is updated, with a formula that gives more weight to higher numbers,
1002  * to quickly follows sudden increases and decrease slowly
1003  */
1004
1005 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1006                                         struct cfq_group *cfqg, bool rt)
1007 {
1008         unsigned min_q, max_q;
1009         unsigned mult  = cfq_hist_divisor - 1;
1010         unsigned round = cfq_hist_divisor / 2;
1011         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1012
1013         min_q = min(cfqg->busy_queues_avg[rt], busy);
1014         max_q = max(cfqg->busy_queues_avg[rt], busy);
1015         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1016                 cfq_hist_divisor;
1017         return cfqg->busy_queues_avg[rt];
1018 }
1019
1020 static inline unsigned
1021 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1022 {
1023         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1024 }
1025
1026 static inline unsigned
1027 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1028 {
1029         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
1030         if (cfqd->cfq_latency) {
1031                 /*
1032                  * interested queues (we consider only the ones with the same
1033                  * priority class in the cfq group)
1034                  */
1035                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1036                                                 cfq_class_rt(cfqq));
1037                 unsigned sync_slice = cfqd->cfq_slice[1];
1038                 unsigned expect_latency = sync_slice * iq;
1039                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1040
1041                 if (expect_latency > group_slice) {
1042                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
1043                         /* scale low_slice according to IO priority
1044                          * and sync vs async */
1045                         unsigned low_slice =
1046                                 min(slice, base_low_slice * slice / sync_slice);
1047                         /* the adapted slice value is scaled to fit all iqs
1048                          * into the target latency */
1049                         slice = max(slice * group_slice / expect_latency,
1050                                     low_slice);
1051                 }
1052         }
1053         return slice;
1054 }
1055
1056 static inline void
1057 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1058 {
1059         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1060
1061         cfqq->slice_start = jiffies;
1062         cfqq->slice_end = jiffies + slice;
1063         cfqq->allocated_slice = slice;
1064         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1065 }
1066
1067 /*
1068  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1069  * isn't valid until the first request from the dispatch is activated
1070  * and the slice time set.
1071  */
1072 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1073 {
1074         if (cfq_cfqq_slice_new(cfqq))
1075                 return false;
1076         if (time_before(jiffies, cfqq->slice_end))
1077                 return false;
1078
1079         return true;
1080 }
1081
1082 /*
1083  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1084  * We choose the request that is closest to the head right now. Distance
1085  * behind the head is penalized and only allowed to a certain extent.
1086  */
1087 static struct request *
1088 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1089 {
1090         sector_t s1, s2, d1 = 0, d2 = 0;
1091         unsigned long back_max;
1092 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1093 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1094         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1095
1096         if (rq1 == NULL || rq1 == rq2)
1097                 return rq2;
1098         if (rq2 == NULL)
1099                 return rq1;
1100
1101         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1102                 return rq_is_sync(rq1) ? rq1 : rq2;
1103
1104         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1105                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1106
1107         s1 = blk_rq_pos(rq1);
1108         s2 = blk_rq_pos(rq2);
1109
1110         /*
1111          * by definition, 1KiB is 2 sectors
1112          */
1113         back_max = cfqd->cfq_back_max * 2;
1114
1115         /*
1116          * Strict one way elevator _except_ in the case where we allow
1117          * short backward seeks which are biased as twice the cost of a
1118          * similar forward seek.
1119          */
1120         if (s1 >= last)
1121                 d1 = s1 - last;
1122         else if (s1 + back_max >= last)
1123                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1124         else
1125                 wrap |= CFQ_RQ1_WRAP;
1126
1127         if (s2 >= last)
1128                 d2 = s2 - last;
1129         else if (s2 + back_max >= last)
1130                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1131         else
1132                 wrap |= CFQ_RQ2_WRAP;
1133
1134         /* Found required data */
1135
1136         /*
1137          * By doing switch() on the bit mask "wrap" we avoid having to
1138          * check two variables for all permutations: --> faster!
1139          */
1140         switch (wrap) {
1141         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1142                 if (d1 < d2)
1143                         return rq1;
1144                 else if (d2 < d1)
1145                         return rq2;
1146                 else {
1147                         if (s1 >= s2)
1148                                 return rq1;
1149                         else
1150                                 return rq2;
1151                 }
1152
1153         case CFQ_RQ2_WRAP:
1154                 return rq1;
1155         case CFQ_RQ1_WRAP:
1156                 return rq2;
1157         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1158         default:
1159                 /*
1160                  * Since both rqs are wrapped,
1161                  * start with the one that's further behind head
1162                  * (--> only *one* back seek required),
1163                  * since back seek takes more time than forward.
1164                  */
1165                 if (s1 <= s2)
1166                         return rq1;
1167                 else
1168                         return rq2;
1169         }
1170 }
1171
1172 /*
1173  * The below is leftmost cache rbtree addon
1174  */
1175 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1176 {
1177         /* Service tree is empty */
1178         if (!root->count)
1179                 return NULL;
1180
1181         if (!root->left)
1182                 root->left = rb_first(&root->rb);
1183
1184         if (root->left)
1185                 return rb_entry(root->left, struct cfq_queue, rb_node);
1186
1187         return NULL;
1188 }
1189
1190 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1191 {
1192         if (!root->left)
1193                 root->left = rb_first(&root->rb);
1194
1195         if (root->left)
1196                 return rb_entry_cfqg(root->left);
1197
1198         return NULL;
1199 }
1200
1201 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1202 {
1203         rb_erase(n, root);
1204         RB_CLEAR_NODE(n);
1205 }
1206
1207 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1208 {
1209         if (root->left == n)
1210                 root->left = NULL;
1211         rb_erase_init(n, &root->rb);
1212         --root->count;
1213 }
1214
1215 /*
1216  * would be nice to take fifo expire time into account as well
1217  */
1218 static struct request *
1219 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1220                   struct request *last)
1221 {
1222         struct rb_node *rbnext = rb_next(&last->rb_node);
1223         struct rb_node *rbprev = rb_prev(&last->rb_node);
1224         struct request *next = NULL, *prev = NULL;
1225
1226         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1227
1228         if (rbprev)
1229                 prev = rb_entry_rq(rbprev);
1230
1231         if (rbnext)
1232                 next = rb_entry_rq(rbnext);
1233         else {
1234                 rbnext = rb_first(&cfqq->sort_list);
1235                 if (rbnext && rbnext != &last->rb_node)
1236                         next = rb_entry_rq(rbnext);
1237         }
1238
1239         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1240 }
1241
1242 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1243                                       struct cfq_queue *cfqq)
1244 {
1245         /*
1246          * just an approximation, should be ok.
1247          */
1248         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1249                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1250 }
1251
1252 static inline s64
1253 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1254 {
1255         return cfqg->vdisktime - st->min_vdisktime;
1256 }
1257
1258 static void
1259 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1260 {
1261         struct rb_node **node = &st->rb.rb_node;
1262         struct rb_node *parent = NULL;
1263         struct cfq_group *__cfqg;
1264         s64 key = cfqg_key(st, cfqg);
1265         int left = 1;
1266
1267         while (*node != NULL) {
1268                 parent = *node;
1269                 __cfqg = rb_entry_cfqg(parent);
1270
1271                 if (key < cfqg_key(st, __cfqg))
1272                         node = &parent->rb_left;
1273                 else {
1274                         node = &parent->rb_right;
1275                         left = 0;
1276                 }
1277         }
1278
1279         if (left)
1280                 st->left = &cfqg->rb_node;
1281
1282         rb_link_node(&cfqg->rb_node, parent, node);
1283         rb_insert_color(&cfqg->rb_node, &st->rb);
1284 }
1285
1286 /*
1287  * This has to be called only on activation of cfqg
1288  */
1289 static void
1290 cfq_update_group_weight(struct cfq_group *cfqg)
1291 {
1292         if (cfqg->new_weight) {
1293                 cfqg->weight = cfqg->new_weight;
1294                 cfqg->new_weight = 0;
1295         }
1296 }
1297
1298 static void
1299 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1300 {
1301         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1302
1303         if (cfqg->new_leaf_weight) {
1304                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1305                 cfqg->new_leaf_weight = 0;
1306         }
1307 }
1308
1309 static void
1310 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1311 {
1312         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1313         struct cfq_group *pos = cfqg;
1314         struct cfq_group *parent;
1315         bool propagate;
1316
1317         /* add to the service tree */
1318         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1319
1320         /*
1321          * Update leaf_weight.  We cannot update weight at this point
1322          * because cfqg might already have been activated and is
1323          * contributing its current weight to the parent's child_weight.
1324          */
1325         cfq_update_group_leaf_weight(cfqg);
1326         __cfq_group_service_tree_add(st, cfqg);
1327
1328         /*
1329          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1330          * entitled to.  vfraction is calculated by walking the tree
1331          * towards the root calculating the fraction it has at each level.
1332          * The compounded ratio is how much vfraction @cfqg owns.
1333          *
1334          * Start with the proportion tasks in this cfqg has against active
1335          * children cfqgs - its leaf_weight against children_weight.
1336          */
1337         propagate = !pos->nr_active++;
1338         pos->children_weight += pos->leaf_weight;
1339         vfr = vfr * pos->leaf_weight / pos->children_weight;
1340
1341         /*
1342          * Compound ->weight walking up the tree.  Both activation and
1343          * vfraction calculation are done in the same loop.  Propagation
1344          * stops once an already activated node is met.  vfraction
1345          * calculation should always continue to the root.
1346          */
1347         while ((parent = cfqg_parent(pos))) {
1348                 if (propagate) {
1349                         cfq_update_group_weight(pos);
1350                         propagate = !parent->nr_active++;
1351                         parent->children_weight += pos->weight;
1352                 }
1353                 vfr = vfr * pos->weight / parent->children_weight;
1354                 pos = parent;
1355         }
1356
1357         cfqg->vfraction = max_t(unsigned, vfr, 1);
1358 }
1359
1360 static void
1361 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1362 {
1363         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1364         struct cfq_group *__cfqg;
1365         struct rb_node *n;
1366
1367         cfqg->nr_cfqq++;
1368         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1369                 return;
1370
1371         /*
1372          * Currently put the group at the end. Later implement something
1373          * so that groups get lesser vtime based on their weights, so that
1374          * if group does not loose all if it was not continuously backlogged.
1375          */
1376         n = rb_last(&st->rb);
1377         if (n) {
1378                 __cfqg = rb_entry_cfqg(n);
1379                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1380         } else
1381                 cfqg->vdisktime = st->min_vdisktime;
1382         cfq_group_service_tree_add(st, cfqg);
1383 }
1384
1385 static void
1386 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1387 {
1388         struct cfq_group *pos = cfqg;
1389         bool propagate;
1390
1391         /*
1392          * Undo activation from cfq_group_service_tree_add().  Deactivate
1393          * @cfqg and propagate deactivation upwards.
1394          */
1395         propagate = !--pos->nr_active;
1396         pos->children_weight -= pos->leaf_weight;
1397
1398         while (propagate) {
1399                 struct cfq_group *parent = cfqg_parent(pos);
1400
1401                 /* @pos has 0 nr_active at this point */
1402                 WARN_ON_ONCE(pos->children_weight);
1403                 pos->vfraction = 0;
1404
1405                 if (!parent)
1406                         break;
1407
1408                 propagate = !--parent->nr_active;
1409                 parent->children_weight -= pos->weight;
1410                 pos = parent;
1411         }
1412
1413         /* remove from the service tree */
1414         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1415                 cfq_rb_erase(&cfqg->rb_node, st);
1416 }
1417
1418 static void
1419 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1420 {
1421         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1422
1423         BUG_ON(cfqg->nr_cfqq < 1);
1424         cfqg->nr_cfqq--;
1425
1426         /* If there are other cfq queues under this group, don't delete it */
1427         if (cfqg->nr_cfqq)
1428                 return;
1429
1430         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1431         cfq_group_service_tree_del(st, cfqg);
1432         cfqg->saved_wl_slice = 0;
1433         cfqg_stats_update_dequeue(cfqg);
1434 }
1435
1436 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1437                                                 unsigned int *unaccounted_time)
1438 {
1439         unsigned int slice_used;
1440
1441         /*
1442          * Queue got expired before even a single request completed or
1443          * got expired immediately after first request completion.
1444          */
1445         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1446                 /*
1447                  * Also charge the seek time incurred to the group, otherwise
1448                  * if there are mutiple queues in the group, each can dispatch
1449                  * a single request on seeky media and cause lots of seek time
1450                  * and group will never know it.
1451                  */
1452                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1453                                         1);
1454         } else {
1455                 slice_used = jiffies - cfqq->slice_start;
1456                 if (slice_used > cfqq->allocated_slice) {
1457                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1458                         slice_used = cfqq->allocated_slice;
1459                 }
1460                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1461                         *unaccounted_time += cfqq->slice_start -
1462                                         cfqq->dispatch_start;
1463         }
1464
1465         return slice_used;
1466 }
1467
1468 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1469                                 struct cfq_queue *cfqq)
1470 {
1471         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1472         unsigned int used_sl, charge, unaccounted_sl = 0;
1473         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1474                         - cfqg->service_tree_idle.count;
1475         unsigned int vfr;
1476
1477         BUG_ON(nr_sync < 0);
1478         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1479
1480         if (iops_mode(cfqd))
1481                 charge = cfqq->slice_dispatch;
1482         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1483                 charge = cfqq->allocated_slice;
1484
1485         /*
1486          * Can't update vdisktime while on service tree and cfqg->vfraction
1487          * is valid only while on it.  Cache vfr, leave the service tree,
1488          * update vdisktime and go back on.  The re-addition to the tree
1489          * will also update the weights as necessary.
1490          */
1491         vfr = cfqg->vfraction;
1492         cfq_group_service_tree_del(st, cfqg);
1493         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1494         cfq_group_service_tree_add(st, cfqg);
1495
1496         /* This group is being expired. Save the context */
1497         if (time_after(cfqd->workload_expires, jiffies)) {
1498                 cfqg->saved_wl_slice = cfqd->workload_expires
1499                                                 - jiffies;
1500                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1501                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1502         } else
1503                 cfqg->saved_wl_slice = 0;
1504
1505         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1506                                         st->min_vdisktime);
1507         cfq_log_cfqq(cfqq->cfqd, cfqq,
1508                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1509                      used_sl, cfqq->slice_dispatch, charge,
1510                      iops_mode(cfqd), cfqq->nr_sectors);
1511         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1512         cfqg_stats_set_start_empty_time(cfqg);
1513 }
1514
1515 /**
1516  * cfq_init_cfqg_base - initialize base part of a cfq_group
1517  * @cfqg: cfq_group to initialize
1518  *
1519  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1520  * is enabled or not.
1521  */
1522 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1523 {
1524         struct cfq_rb_root *st;
1525         int i, j;
1526
1527         for_each_cfqg_st(cfqg, i, j, st)
1528                 *st = CFQ_RB_ROOT;
1529         RB_CLEAR_NODE(&cfqg->rb_node);
1530
1531         cfqg->ttime.last_end_request = jiffies;
1532 }
1533
1534 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1535 static void cfqg_stats_exit(struct cfqg_stats *stats)
1536 {
1537         blkg_rwstat_exit(&stats->merged);
1538         blkg_rwstat_exit(&stats->service_time);
1539         blkg_rwstat_exit(&stats->wait_time);
1540         blkg_rwstat_exit(&stats->queued);
1541
1542         blkg_stat_exit(&stats->sectors);
1543         blkg_stat_exit(&stats->time);
1544 #ifdef CONFIG_DEBUG_BLK_CGROUP
1545         blkg_stat_exit(&stats->unaccounted_time);
1546         blkg_stat_exit(&stats->avg_queue_size_sum);
1547         blkg_stat_exit(&stats->avg_queue_size_samples);
1548         blkg_stat_exit(&stats->dequeue);
1549         blkg_stat_exit(&stats->group_wait_time);
1550         blkg_stat_exit(&stats->idle_time);
1551         blkg_stat_exit(&stats->empty_time);
1552 #endif
1553 }
1554
1555 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1556 {
1557         if (blkg_rwstat_init(&stats->merged, gfp) ||
1558             blkg_rwstat_init(&stats->service_time, gfp) ||
1559             blkg_rwstat_init(&stats->wait_time, gfp) ||
1560             blkg_rwstat_init(&stats->queued, gfp) ||
1561
1562             blkg_stat_init(&stats->sectors, gfp) ||
1563             blkg_stat_init(&stats->time, gfp))
1564                 goto err;
1565
1566 #ifdef CONFIG_DEBUG_BLK_CGROUP
1567         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1568             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1569             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1570             blkg_stat_init(&stats->dequeue, gfp) ||
1571             blkg_stat_init(&stats->group_wait_time, gfp) ||
1572             blkg_stat_init(&stats->idle_time, gfp) ||
1573             blkg_stat_init(&stats->empty_time, gfp))
1574                 goto err;
1575 #endif
1576         return 0;
1577 err:
1578         cfqg_stats_exit(stats);
1579         return -ENOMEM;
1580 }
1581
1582 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1583 {
1584         struct cfq_group_data *cgd;
1585
1586         cgd = kzalloc(sizeof(*cgd), GFP_KERNEL);
1587         if (!cgd)
1588                 return NULL;
1589         return &cgd->cpd;
1590 }
1591
1592 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1593 {
1594         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1595
1596         if (cpd_to_blkcg(cpd) == &blkcg_root) {
1597                 cgd->weight = 2 * CFQ_WEIGHT_DEFAULT;
1598                 cgd->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
1599         } else {
1600                 cgd->weight = CFQ_WEIGHT_DEFAULT;
1601                 cgd->leaf_weight = CFQ_WEIGHT_DEFAULT;
1602         }
1603 }
1604
1605 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1606 {
1607         kfree(cpd_to_cfqgd(cpd));
1608 }
1609
1610 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1611 {
1612         struct cfq_group *cfqg;
1613
1614         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1615         if (!cfqg)
1616                 return NULL;
1617
1618         cfq_init_cfqg_base(cfqg);
1619         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1620                 kfree(cfqg);
1621                 return NULL;
1622         }
1623
1624         return &cfqg->pd;
1625 }
1626
1627 static void cfq_pd_init(struct blkg_policy_data *pd)
1628 {
1629         struct cfq_group *cfqg = pd_to_cfqg(pd);
1630         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1631
1632         cfqg->weight = cgd->weight;
1633         cfqg->leaf_weight = cgd->leaf_weight;
1634 }
1635
1636 static void cfq_pd_offline(struct blkg_policy_data *pd)
1637 {
1638         struct cfq_group *cfqg = pd_to_cfqg(pd);
1639         int i;
1640
1641         for (i = 0; i < IOPRIO_BE_NR; i++) {
1642                 if (cfqg->async_cfqq[0][i])
1643                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1644                 if (cfqg->async_cfqq[1][i])
1645                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1646         }
1647
1648         if (cfqg->async_idle_cfqq)
1649                 cfq_put_queue(cfqg->async_idle_cfqq);
1650
1651         /*
1652          * @blkg is going offline and will be ignored by
1653          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1654          * that they don't get lost.  If IOs complete after this point, the
1655          * stats for them will be lost.  Oh well...
1656          */
1657         cfqg_stats_xfer_dead(cfqg);
1658 }
1659
1660 static void cfq_pd_free(struct blkg_policy_data *pd)
1661 {
1662         struct cfq_group *cfqg = pd_to_cfqg(pd);
1663
1664         cfqg_stats_exit(&cfqg->stats);
1665         return kfree(cfqg);
1666 }
1667
1668 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1669 {
1670         struct cfq_group *cfqg = pd_to_cfqg(pd);
1671
1672         cfqg_stats_reset(&cfqg->stats);
1673 }
1674
1675 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1676                                          struct blkcg *blkcg)
1677 {
1678         struct blkcg_gq *blkg;
1679
1680         blkg = blkg_lookup(blkcg, cfqd->queue);
1681         if (likely(blkg))
1682                 return blkg_to_cfqg(blkg);
1683         return NULL;
1684 }
1685
1686 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1687 {
1688         cfqq->cfqg = cfqg;
1689         /* cfqq reference on cfqg */
1690         cfqg_get(cfqg);
1691 }
1692
1693 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1694                                      struct blkg_policy_data *pd, int off)
1695 {
1696         struct cfq_group *cfqg = pd_to_cfqg(pd);
1697
1698         if (!cfqg->dev_weight)
1699                 return 0;
1700         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1701 }
1702
1703 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1704 {
1705         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1706                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1707                           0, false);
1708         return 0;
1709 }
1710
1711 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1712                                           struct blkg_policy_data *pd, int off)
1713 {
1714         struct cfq_group *cfqg = pd_to_cfqg(pd);
1715
1716         if (!cfqg->dev_leaf_weight)
1717                 return 0;
1718         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1719 }
1720
1721 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1722 {
1723         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1724                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1725                           0, false);
1726         return 0;
1727 }
1728
1729 static int cfq_print_weight(struct seq_file *sf, void *v)
1730 {
1731         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1732         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1733         unsigned int val = 0;
1734
1735         if (cgd)
1736                 val = cgd->weight;
1737
1738         seq_printf(sf, "%u\n", val);
1739         return 0;
1740 }
1741
1742 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1743 {
1744         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1745         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1746         unsigned int val = 0;
1747
1748         if (cgd)
1749                 val = cgd->leaf_weight;
1750
1751         seq_printf(sf, "%u\n", val);
1752         return 0;
1753 }
1754
1755 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1756                                         char *buf, size_t nbytes, loff_t off,
1757                                         bool is_leaf_weight)
1758 {
1759         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1760         struct blkg_conf_ctx ctx;
1761         struct cfq_group *cfqg;
1762         struct cfq_group_data *cfqgd;
1763         int ret;
1764
1765         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1766         if (ret)
1767                 return ret;
1768
1769         ret = -EINVAL;
1770         cfqg = blkg_to_cfqg(ctx.blkg);
1771         cfqgd = blkcg_to_cfqgd(blkcg);
1772         if (!cfqg || !cfqgd)
1773                 goto err;
1774
1775         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1776                 if (!is_leaf_weight) {
1777                         cfqg->dev_weight = ctx.v;
1778                         cfqg->new_weight = ctx.v ?: cfqgd->weight;
1779                 } else {
1780                         cfqg->dev_leaf_weight = ctx.v;
1781                         cfqg->new_leaf_weight = ctx.v ?: cfqgd->leaf_weight;
1782                 }
1783                 ret = 0;
1784         }
1785
1786 err:
1787         blkg_conf_finish(&ctx);
1788         return ret ?: nbytes;
1789 }
1790
1791 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1792                                       char *buf, size_t nbytes, loff_t off)
1793 {
1794         return __cfqg_set_weight_device(of, buf, nbytes, off, false);
1795 }
1796
1797 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1798                                            char *buf, size_t nbytes, loff_t off)
1799 {
1800         return __cfqg_set_weight_device(of, buf, nbytes, off, true);
1801 }
1802
1803 static int __cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1804                             u64 val, bool is_leaf_weight)
1805 {
1806         struct blkcg *blkcg = css_to_blkcg(css);
1807         struct blkcg_gq *blkg;
1808         struct cfq_group_data *cfqgd;
1809         int ret = 0;
1810
1811         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1812                 return -EINVAL;
1813
1814         spin_lock_irq(&blkcg->lock);
1815         cfqgd = blkcg_to_cfqgd(blkcg);
1816         if (!cfqgd) {
1817                 ret = -EINVAL;
1818                 goto out;
1819         }
1820
1821         if (!is_leaf_weight)
1822                 cfqgd->weight = val;
1823         else
1824                 cfqgd->leaf_weight = val;
1825
1826         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1827                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1828
1829                 if (!cfqg)
1830                         continue;
1831
1832                 if (!is_leaf_weight) {
1833                         if (!cfqg->dev_weight)
1834                                 cfqg->new_weight = cfqgd->weight;
1835                 } else {
1836                         if (!cfqg->dev_leaf_weight)
1837                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1838                 }
1839         }
1840
1841 out:
1842         spin_unlock_irq(&blkcg->lock);
1843         return ret;
1844 }
1845
1846 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1847                           u64 val)
1848 {
1849         return __cfq_set_weight(css, cft, val, false);
1850 }
1851
1852 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1853                                struct cftype *cft, u64 val)
1854 {
1855         return __cfq_set_weight(css, cft, val, true);
1856 }
1857
1858 static int cfqg_print_stat(struct seq_file *sf, void *v)
1859 {
1860         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1861                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1862         return 0;
1863 }
1864
1865 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1866 {
1867         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1868                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1869         return 0;
1870 }
1871
1872 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1873                                       struct blkg_policy_data *pd, int off)
1874 {
1875         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1876                                           &blkcg_policy_cfq, off);
1877         return __blkg_prfill_u64(sf, pd, sum);
1878 }
1879
1880 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1881                                         struct blkg_policy_data *pd, int off)
1882 {
1883         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1884                                                         &blkcg_policy_cfq, off);
1885         return __blkg_prfill_rwstat(sf, pd, &sum);
1886 }
1887
1888 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1889 {
1890         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1891                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1892                           seq_cft(sf)->private, false);
1893         return 0;
1894 }
1895
1896 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1897 {
1898         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1899                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1900                           seq_cft(sf)->private, true);
1901         return 0;
1902 }
1903
1904 #ifdef CONFIG_DEBUG_BLK_CGROUP
1905 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1906                                       struct blkg_policy_data *pd, int off)
1907 {
1908         struct cfq_group *cfqg = pd_to_cfqg(pd);
1909         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1910         u64 v = 0;
1911
1912         if (samples) {
1913                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1914                 v = div64_u64(v, samples);
1915         }
1916         __blkg_prfill_u64(sf, pd, v);
1917         return 0;
1918 }
1919
1920 /* print avg_queue_size */
1921 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1922 {
1923         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1924                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1925                           0, false);
1926         return 0;
1927 }
1928 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1929
1930 static struct cftype cfq_blkcg_files[] = {
1931         /* on root, weight is mapped to leaf_weight */
1932         {
1933                 .name = "weight_device",
1934                 .flags = CFTYPE_ONLY_ON_ROOT,
1935                 .seq_show = cfqg_print_leaf_weight_device,
1936                 .write = cfqg_set_leaf_weight_device,
1937         },
1938         {
1939                 .name = "weight",
1940                 .flags = CFTYPE_ONLY_ON_ROOT,
1941                 .seq_show = cfq_print_leaf_weight,
1942                 .write_u64 = cfq_set_leaf_weight,
1943         },
1944
1945         /* no such mapping necessary for !roots */
1946         {
1947                 .name = "weight_device",
1948                 .flags = CFTYPE_NOT_ON_ROOT,
1949                 .seq_show = cfqg_print_weight_device,
1950                 .write = cfqg_set_weight_device,
1951         },
1952         {
1953                 .name = "weight",
1954                 .flags = CFTYPE_NOT_ON_ROOT,
1955                 .seq_show = cfq_print_weight,
1956                 .write_u64 = cfq_set_weight,
1957         },
1958
1959         {
1960                 .name = "leaf_weight_device",
1961                 .seq_show = cfqg_print_leaf_weight_device,
1962                 .write = cfqg_set_leaf_weight_device,
1963         },
1964         {
1965                 .name = "leaf_weight",
1966                 .seq_show = cfq_print_leaf_weight,
1967                 .write_u64 = cfq_set_leaf_weight,
1968         },
1969
1970         /* statistics, covers only the tasks in the cfqg */
1971         {
1972                 .name = "time",
1973                 .private = offsetof(struct cfq_group, stats.time),
1974                 .seq_show = cfqg_print_stat,
1975         },
1976         {
1977                 .name = "sectors",
1978                 .private = offsetof(struct cfq_group, stats.sectors),
1979                 .seq_show = cfqg_print_stat,
1980         },
1981         {
1982                 .name = "io_service_bytes",
1983                 .private = (unsigned long)&blkcg_policy_cfq,
1984                 .seq_show = blkg_print_stat_bytes,
1985         },
1986         {
1987                 .name = "io_serviced",
1988                 .private = (unsigned long)&blkcg_policy_cfq,
1989                 .seq_show = blkg_print_stat_ios,
1990         },
1991         {
1992                 .name = "io_service_time",
1993                 .private = offsetof(struct cfq_group, stats.service_time),
1994                 .seq_show = cfqg_print_rwstat,
1995         },
1996         {
1997                 .name = "io_wait_time",
1998                 .private = offsetof(struct cfq_group, stats.wait_time),
1999                 .seq_show = cfqg_print_rwstat,
2000         },
2001         {
2002                 .name = "io_merged",
2003                 .private = offsetof(struct cfq_group, stats.merged),
2004                 .seq_show = cfqg_print_rwstat,
2005         },
2006         {
2007                 .name = "io_queued",
2008                 .private = offsetof(struct cfq_group, stats.queued),
2009                 .seq_show = cfqg_print_rwstat,
2010         },
2011
2012         /* the same statictics which cover the cfqg and its descendants */
2013         {
2014                 .name = "time_recursive",
2015                 .private = offsetof(struct cfq_group, stats.time),
2016                 .seq_show = cfqg_print_stat_recursive,
2017         },
2018         {
2019                 .name = "sectors_recursive",
2020                 .private = offsetof(struct cfq_group, stats.sectors),
2021                 .seq_show = cfqg_print_stat_recursive,
2022         },
2023         {
2024                 .name = "io_service_bytes_recursive",
2025                 .private = (unsigned long)&blkcg_policy_cfq,
2026                 .seq_show = blkg_print_stat_bytes_recursive,
2027         },
2028         {
2029                 .name = "io_serviced_recursive",
2030                 .private = (unsigned long)&blkcg_policy_cfq,
2031                 .seq_show = blkg_print_stat_ios_recursive,
2032         },
2033         {
2034                 .name = "io_service_time_recursive",
2035                 .private = offsetof(struct cfq_group, stats.service_time),
2036                 .seq_show = cfqg_print_rwstat_recursive,
2037         },
2038         {
2039                 .name = "io_wait_time_recursive",
2040                 .private = offsetof(struct cfq_group, stats.wait_time),
2041                 .seq_show = cfqg_print_rwstat_recursive,
2042         },
2043         {
2044                 .name = "io_merged_recursive",
2045                 .private = offsetof(struct cfq_group, stats.merged),
2046                 .seq_show = cfqg_print_rwstat_recursive,
2047         },
2048         {
2049                 .name = "io_queued_recursive",
2050                 .private = offsetof(struct cfq_group, stats.queued),
2051                 .seq_show = cfqg_print_rwstat_recursive,
2052         },
2053 #ifdef CONFIG_DEBUG_BLK_CGROUP
2054         {
2055                 .name = "avg_queue_size",
2056                 .seq_show = cfqg_print_avg_queue_size,
2057         },
2058         {
2059                 .name = "group_wait_time",
2060                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2061                 .seq_show = cfqg_print_stat,
2062         },
2063         {
2064                 .name = "idle_time",
2065                 .private = offsetof(struct cfq_group, stats.idle_time),
2066                 .seq_show = cfqg_print_stat,
2067         },
2068         {
2069                 .name = "empty_time",
2070                 .private = offsetof(struct cfq_group, stats.empty_time),
2071                 .seq_show = cfqg_print_stat,
2072         },
2073         {
2074                 .name = "dequeue",
2075                 .private = offsetof(struct cfq_group, stats.dequeue),
2076                 .seq_show = cfqg_print_stat,
2077         },
2078         {
2079                 .name = "unaccounted_time",
2080                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2081                 .seq_show = cfqg_print_stat,
2082         },
2083 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2084         { }     /* terminate */
2085 };
2086 #else /* GROUP_IOSCHED */
2087 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2088                                          struct blkcg *blkcg)
2089 {
2090         return cfqd->root_group;
2091 }
2092
2093 static inline void
2094 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2095         cfqq->cfqg = cfqg;
2096 }
2097
2098 #endif /* GROUP_IOSCHED */
2099
2100 /*
2101  * The cfqd->service_trees holds all pending cfq_queue's that have
2102  * requests waiting to be processed. It is sorted in the order that
2103  * we will service the queues.
2104  */
2105 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2106                                  bool add_front)
2107 {
2108         struct rb_node **p, *parent;
2109         struct cfq_queue *__cfqq;
2110         unsigned long rb_key;
2111         struct cfq_rb_root *st;
2112         int left;
2113         int new_cfqq = 1;
2114
2115         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2116         if (cfq_class_idle(cfqq)) {
2117                 rb_key = CFQ_IDLE_DELAY;
2118                 parent = rb_last(&st->rb);
2119                 if (parent && parent != &cfqq->rb_node) {
2120                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2121                         rb_key += __cfqq->rb_key;
2122                 } else
2123                         rb_key += jiffies;
2124         } else if (!add_front) {
2125                 /*
2126                  * Get our rb key offset. Subtract any residual slice
2127                  * value carried from last service. A negative resid
2128                  * count indicates slice overrun, and this should position
2129                  * the next service time further away in the tree.
2130                  */
2131                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
2132                 rb_key -= cfqq->slice_resid;
2133                 cfqq->slice_resid = 0;
2134         } else {
2135                 rb_key = -HZ;
2136                 __cfqq = cfq_rb_first(st);
2137                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
2138         }
2139
2140         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2141                 new_cfqq = 0;
2142                 /*
2143                  * same position, nothing more to do
2144                  */
2145                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2146                         return;
2147
2148                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2149                 cfqq->service_tree = NULL;
2150         }
2151
2152         left = 1;
2153         parent = NULL;
2154         cfqq->service_tree = st;
2155         p = &st->rb.rb_node;
2156         while (*p) {
2157                 parent = *p;
2158                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2159
2160                 /*
2161                  * sort by key, that represents service time.
2162                  */
2163                 if (time_before(rb_key, __cfqq->rb_key))
2164                         p = &parent->rb_left;
2165                 else {
2166                         p = &parent->rb_right;
2167                         left = 0;
2168                 }
2169         }
2170
2171         if (left)
2172                 st->left = &cfqq->rb_node;
2173
2174         cfqq->rb_key = rb_key;
2175         rb_link_node(&cfqq->rb_node, parent, p);
2176         rb_insert_color(&cfqq->rb_node, &st->rb);
2177         st->count++;
2178         if (add_front || !new_cfqq)
2179                 return;
2180         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2181 }
2182
2183 static struct cfq_queue *
2184 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2185                      sector_t sector, struct rb_node **ret_parent,
2186                      struct rb_node ***rb_link)
2187 {
2188         struct rb_node **p, *parent;
2189         struct cfq_queue *cfqq = NULL;
2190
2191         parent = NULL;
2192         p = &root->rb_node;
2193         while (*p) {
2194                 struct rb_node **n;
2195
2196                 parent = *p;
2197                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2198
2199                 /*
2200                  * Sort strictly based on sector.  Smallest to the left,
2201                  * largest to the right.
2202                  */
2203                 if (sector > blk_rq_pos(cfqq->next_rq))
2204                         n = &(*p)->rb_right;
2205                 else if (sector < blk_rq_pos(cfqq->next_rq))
2206                         n = &(*p)->rb_left;
2207                 else
2208                         break;
2209                 p = n;
2210                 cfqq = NULL;
2211         }
2212
2213         *ret_parent = parent;
2214         if (rb_link)
2215                 *rb_link = p;
2216         return cfqq;
2217 }
2218
2219 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2220 {
2221         struct rb_node **p, *parent;
2222         struct cfq_queue *__cfqq;
2223
2224         if (cfqq->p_root) {
2225                 rb_erase(&cfqq->p_node, cfqq->p_root);
2226                 cfqq->p_root = NULL;
2227         }
2228
2229         if (cfq_class_idle(cfqq))
2230                 return;
2231         if (!cfqq->next_rq)
2232                 return;
2233
2234         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2235         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2236                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2237         if (!__cfqq) {
2238                 rb_link_node(&cfqq->p_node, parent, p);
2239                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2240         } else
2241                 cfqq->p_root = NULL;
2242 }
2243
2244 /*
2245  * Update cfqq's position in the service tree.
2246  */
2247 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2248 {
2249         /*
2250          * Resorting requires the cfqq to be on the RR list already.
2251          */
2252         if (cfq_cfqq_on_rr(cfqq)) {
2253                 cfq_service_tree_add(cfqd, cfqq, 0);
2254                 cfq_prio_tree_add(cfqd, cfqq);
2255         }
2256 }
2257
2258 /*
2259  * add to busy list of queues for service, trying to be fair in ordering
2260  * the pending list according to last request service
2261  */
2262 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2263 {
2264         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2265         BUG_ON(cfq_cfqq_on_rr(cfqq));
2266         cfq_mark_cfqq_on_rr(cfqq);
2267         cfqd->busy_queues++;
2268         if (cfq_cfqq_sync(cfqq))
2269                 cfqd->busy_sync_queues++;
2270
2271         cfq_resort_rr_list(cfqd, cfqq);
2272 }
2273
2274 /*
2275  * Called when the cfqq no longer has requests pending, remove it from
2276  * the service tree.
2277  */
2278 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2279 {
2280         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2281         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2282         cfq_clear_cfqq_on_rr(cfqq);
2283
2284         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2285                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2286                 cfqq->service_tree = NULL;
2287         }
2288         if (cfqq->p_root) {
2289                 rb_erase(&cfqq->p_node, cfqq->p_root);
2290                 cfqq->p_root = NULL;
2291         }
2292
2293         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2294         BUG_ON(!cfqd->busy_queues);
2295         cfqd->busy_queues--;
2296         if (cfq_cfqq_sync(cfqq))
2297                 cfqd->busy_sync_queues--;
2298 }
2299
2300 /*
2301  * rb tree support functions
2302  */
2303 static void cfq_del_rq_rb(struct request *rq)
2304 {
2305         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2306         const int sync = rq_is_sync(rq);
2307
2308         BUG_ON(!cfqq->queued[sync]);
2309         cfqq->queued[sync]--;
2310
2311         elv_rb_del(&cfqq->sort_list, rq);
2312
2313         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2314                 /*
2315                  * Queue will be deleted from service tree when we actually
2316                  * expire it later. Right now just remove it from prio tree
2317                  * as it is empty.
2318                  */
2319                 if (cfqq->p_root) {
2320                         rb_erase(&cfqq->p_node, cfqq->p_root);
2321                         cfqq->p_root = NULL;
2322                 }
2323         }
2324 }
2325
2326 static void cfq_add_rq_rb(struct request *rq)
2327 {
2328         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2329         struct cfq_data *cfqd = cfqq->cfqd;
2330         struct request *prev;
2331
2332         cfqq->queued[rq_is_sync(rq)]++;
2333
2334         elv_rb_add(&cfqq->sort_list, rq);
2335
2336         if (!cfq_cfqq_on_rr(cfqq))
2337                 cfq_add_cfqq_rr(cfqd, cfqq);
2338
2339         /*
2340          * check if this request is a better next-serve candidate
2341          */
2342         prev = cfqq->next_rq;
2343         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2344
2345         /*
2346          * adjust priority tree position, if ->next_rq changes
2347          */
2348         if (prev != cfqq->next_rq)
2349                 cfq_prio_tree_add(cfqd, cfqq);
2350
2351         BUG_ON(!cfqq->next_rq);
2352 }
2353
2354 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2355 {
2356         elv_rb_del(&cfqq->sort_list, rq);
2357         cfqq->queued[rq_is_sync(rq)]--;
2358         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2359         cfq_add_rq_rb(rq);
2360         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2361                                  rq->cmd_flags);
2362 }
2363
2364 static struct request *
2365 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2366 {
2367         struct task_struct *tsk = current;
2368         struct cfq_io_cq *cic;
2369         struct cfq_queue *cfqq;
2370
2371         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2372         if (!cic)
2373                 return NULL;
2374
2375         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2376         if (cfqq)
2377                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2378
2379         return NULL;
2380 }
2381
2382 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2383 {
2384         struct cfq_data *cfqd = q->elevator->elevator_data;
2385
2386         cfqd->rq_in_driver++;
2387         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2388                                                 cfqd->rq_in_driver);
2389
2390         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2391 }
2392
2393 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2394 {
2395         struct cfq_data *cfqd = q->elevator->elevator_data;
2396
2397         WARN_ON(!cfqd->rq_in_driver);
2398         cfqd->rq_in_driver--;
2399         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2400                                                 cfqd->rq_in_driver);
2401 }
2402
2403 static void cfq_remove_request(struct request *rq)
2404 {
2405         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2406
2407         if (cfqq->next_rq == rq)
2408                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2409
2410         list_del_init(&rq->queuelist);
2411         cfq_del_rq_rb(rq);
2412
2413         cfqq->cfqd->rq_queued--;
2414         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2415         if (rq->cmd_flags & REQ_PRIO) {
2416                 WARN_ON(!cfqq->prio_pending);
2417                 cfqq->prio_pending--;
2418         }
2419 }
2420
2421 static int cfq_merge(struct request_queue *q, struct request **req,
2422                      struct bio *bio)
2423 {
2424         struct cfq_data *cfqd = q->elevator->elevator_data;
2425         struct request *__rq;
2426
2427         __rq = cfq_find_rq_fmerge(cfqd, bio);
2428         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2429                 *req = __rq;
2430                 return ELEVATOR_FRONT_MERGE;
2431         }
2432
2433         return ELEVATOR_NO_MERGE;
2434 }
2435
2436 static void cfq_merged_request(struct request_queue *q, struct request *req,
2437                                int type)
2438 {
2439         if (type == ELEVATOR_FRONT_MERGE) {
2440                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2441
2442                 cfq_reposition_rq_rb(cfqq, req);
2443         }
2444 }
2445
2446 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2447                                 struct bio *bio)
2448 {
2449         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2450 }
2451
2452 static void
2453 cfq_merged_requests(struct request_queue *q, struct request *rq,
2454                     struct request *next)
2455 {
2456         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2457         struct cfq_data *cfqd = q->elevator->elevator_data;
2458
2459         /*
2460          * reposition in fifo if next is older than rq
2461          */
2462         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2463             time_before(next->fifo_time, rq->fifo_time) &&
2464             cfqq == RQ_CFQQ(next)) {
2465                 list_move(&rq->queuelist, &next->queuelist);
2466                 rq->fifo_time = next->fifo_time;
2467         }
2468
2469         if (cfqq->next_rq == next)
2470                 cfqq->next_rq = rq;
2471         cfq_remove_request(next);
2472         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2473
2474         cfqq = RQ_CFQQ(next);
2475         /*
2476          * all requests of this queue are merged to other queues, delete it
2477          * from the service tree. If it's the active_queue,
2478          * cfq_dispatch_requests() will choose to expire it or do idle
2479          */
2480         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2481             cfqq != cfqd->active_queue)
2482                 cfq_del_cfqq_rr(cfqd, cfqq);
2483 }
2484
2485 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2486                            struct bio *bio)
2487 {
2488         struct cfq_data *cfqd = q->elevator->elevator_data;
2489         struct cfq_io_cq *cic;
2490         struct cfq_queue *cfqq;
2491
2492         /*
2493          * Disallow merge of a sync bio into an async request.
2494          */
2495         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2496                 return false;
2497
2498         /*
2499          * Lookup the cfqq that this bio will be queued with and allow
2500          * merge only if rq is queued there.
2501          */
2502         cic = cfq_cic_lookup(cfqd, current->io_context);
2503         if (!cic)
2504                 return false;
2505
2506         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2507         return cfqq == RQ_CFQQ(rq);
2508 }
2509
2510 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2511 {
2512         del_timer(&cfqd->idle_slice_timer);
2513         cfqg_stats_update_idle_time(cfqq->cfqg);
2514 }
2515
2516 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2517                                    struct cfq_queue *cfqq)
2518 {
2519         if (cfqq) {
2520                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2521                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2522                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2523                 cfqq->slice_start = 0;
2524                 cfqq->dispatch_start = jiffies;
2525                 cfqq->allocated_slice = 0;
2526                 cfqq->slice_end = 0;
2527                 cfqq->slice_dispatch = 0;
2528                 cfqq->nr_sectors = 0;
2529
2530                 cfq_clear_cfqq_wait_request(cfqq);
2531                 cfq_clear_cfqq_must_dispatch(cfqq);
2532                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2533                 cfq_clear_cfqq_fifo_expire(cfqq);
2534                 cfq_mark_cfqq_slice_new(cfqq);
2535
2536                 cfq_del_timer(cfqd, cfqq);
2537         }
2538
2539         cfqd->active_queue = cfqq;
2540 }
2541
2542 /*
2543  * current cfqq expired its slice (or was too idle), select new one
2544  */
2545 static void
2546 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2547                     bool timed_out)
2548 {
2549         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2550
2551         if (cfq_cfqq_wait_request(cfqq))
2552                 cfq_del_timer(cfqd, cfqq);
2553
2554         cfq_clear_cfqq_wait_request(cfqq);
2555         cfq_clear_cfqq_wait_busy(cfqq);
2556
2557         /*
2558          * If this cfqq is shared between multiple processes, check to
2559          * make sure that those processes are still issuing I/Os within
2560          * the mean seek distance.  If not, it may be time to break the
2561          * queues apart again.
2562          */
2563         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2564                 cfq_mark_cfqq_split_coop(cfqq);
2565
2566         /*
2567          * store what was left of this slice, if the queue idled/timed out
2568          */
2569         if (timed_out) {
2570                 if (cfq_cfqq_slice_new(cfqq))
2571                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2572                 else
2573                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2574                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2575         }
2576
2577         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2578
2579         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2580                 cfq_del_cfqq_rr(cfqd, cfqq);
2581
2582         cfq_resort_rr_list(cfqd, cfqq);
2583
2584         if (cfqq == cfqd->active_queue)
2585                 cfqd->active_queue = NULL;
2586
2587         if (cfqd->active_cic) {
2588                 put_io_context(cfqd->active_cic->icq.ioc);
2589                 cfqd->active_cic = NULL;
2590         }
2591 }
2592
2593 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2594 {
2595         struct cfq_queue *cfqq = cfqd->active_queue;
2596
2597         if (cfqq)
2598                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2599 }
2600
2601 /*
2602  * Get next queue for service. Unless we have a queue preemption,
2603  * we'll simply select the first cfqq in the service tree.
2604  */
2605 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2606 {
2607         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2608                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2609
2610         if (!cfqd->rq_queued)
2611                 return NULL;
2612
2613         /* There is nothing to dispatch */
2614         if (!st)
2615                 return NULL;
2616         if (RB_EMPTY_ROOT(&st->rb))
2617                 return NULL;
2618         return cfq_rb_first(st);
2619 }
2620
2621 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2622 {
2623         struct cfq_group *cfqg;
2624         struct cfq_queue *cfqq;
2625         int i, j;
2626         struct cfq_rb_root *st;
2627
2628         if (!cfqd->rq_queued)
2629                 return NULL;
2630
2631         cfqg = cfq_get_next_cfqg(cfqd);
2632         if (!cfqg)
2633                 return NULL;
2634
2635         for_each_cfqg_st(cfqg, i, j, st)
2636                 if ((cfqq = cfq_rb_first(st)) != NULL)
2637                         return cfqq;
2638         return NULL;
2639 }
2640
2641 /*
2642  * Get and set a new active queue for service.
2643  */
2644 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2645                                               struct cfq_queue *cfqq)
2646 {
2647         if (!cfqq)
2648                 cfqq = cfq_get_next_queue(cfqd);
2649
2650         __cfq_set_active_queue(cfqd, cfqq);
2651         return cfqq;
2652 }
2653
2654 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2655                                           struct request *rq)
2656 {
2657         if (blk_rq_pos(rq) >= cfqd->last_position)
2658                 return blk_rq_pos(rq) - cfqd->last_position;
2659         else
2660                 return cfqd->last_position - blk_rq_pos(rq);
2661 }
2662
2663 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2664                                struct request *rq)
2665 {
2666         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2667 }
2668
2669 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2670                                     struct cfq_queue *cur_cfqq)
2671 {
2672         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2673         struct rb_node *parent, *node;
2674         struct cfq_queue *__cfqq;
2675         sector_t sector = cfqd->last_position;
2676
2677         if (RB_EMPTY_ROOT(root))
2678                 return NULL;
2679
2680         /*
2681          * First, if we find a request starting at the end of the last
2682          * request, choose it.
2683          */
2684         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2685         if (__cfqq)
2686                 return __cfqq;
2687
2688         /*
2689          * If the exact sector wasn't found, the parent of the NULL leaf
2690          * will contain the closest sector.
2691          */
2692         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2693         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2694                 return __cfqq;
2695
2696         if (blk_rq_pos(__cfqq->next_rq) < sector)
2697                 node = rb_next(&__cfqq->p_node);
2698         else
2699                 node = rb_prev(&__cfqq->p_node);
2700         if (!node)
2701                 return NULL;
2702
2703         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2704         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2705                 return __cfqq;
2706
2707         return NULL;
2708 }
2709
2710 /*
2711  * cfqd - obvious
2712  * cur_cfqq - passed in so that we don't decide that the current queue is
2713  *            closely cooperating with itself.
2714  *
2715  * So, basically we're assuming that that cur_cfqq has dispatched at least
2716  * one request, and that cfqd->last_position reflects a position on the disk
2717  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2718  * assumption.
2719  */
2720 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2721                                               struct cfq_queue *cur_cfqq)
2722 {
2723         struct cfq_queue *cfqq;
2724
2725         if (cfq_class_idle(cur_cfqq))
2726                 return NULL;
2727         if (!cfq_cfqq_sync(cur_cfqq))
2728                 return NULL;
2729         if (CFQQ_SEEKY(cur_cfqq))
2730                 return NULL;
2731
2732         /*
2733          * Don't search priority tree if it's the only queue in the group.
2734          */
2735         if (cur_cfqq->cfqg->nr_cfqq == 1)
2736                 return NULL;
2737
2738         /*
2739          * We should notice if some of the queues are cooperating, eg
2740          * working closely on the same area of the disk. In that case,
2741          * we can group them together and don't waste time idling.
2742          */
2743         cfqq = cfqq_close(cfqd, cur_cfqq);
2744         if (!cfqq)
2745                 return NULL;
2746
2747         /* If new queue belongs to different cfq_group, don't choose it */
2748         if (cur_cfqq->cfqg != cfqq->cfqg)
2749                 return NULL;
2750
2751         /*
2752          * It only makes sense to merge sync queues.
2753          */
2754         if (!cfq_cfqq_sync(cfqq))
2755                 return NULL;
2756         if (CFQQ_SEEKY(cfqq))
2757                 return NULL;
2758
2759         /*
2760          * Do not merge queues of different priority classes
2761          */
2762         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2763                 return NULL;
2764
2765         return cfqq;
2766 }
2767
2768 /*
2769  * Determine whether we should enforce idle window for this queue.
2770  */
2771
2772 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2773 {
2774         enum wl_class_t wl_class = cfqq_class(cfqq);
2775         struct cfq_rb_root *st = cfqq->service_tree;
2776
2777         BUG_ON(!st);
2778         BUG_ON(!st->count);
2779
2780         if (!cfqd->cfq_slice_idle)
2781                 return false;
2782
2783         /* We never do for idle class queues. */
2784         if (wl_class == IDLE_WORKLOAD)
2785                 return false;
2786
2787         /* We do for queues that were marked with idle window flag. */
2788         if (cfq_cfqq_idle_window(cfqq) &&
2789            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2790                 return true;
2791
2792         /*
2793          * Otherwise, we do only if they are the last ones
2794          * in their service tree.
2795          */
2796         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2797            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2798                 return true;
2799         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2800         return false;
2801 }
2802
2803 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2804 {
2805         struct cfq_queue *cfqq = cfqd->active_queue;
2806         struct cfq_io_cq *cic;
2807         unsigned long sl, group_idle = 0;
2808
2809         /*
2810          * SSD device without seek penalty, disable idling. But only do so
2811          * for devices that support queuing, otherwise we still have a problem
2812          * with sync vs async workloads.
2813          */
2814         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2815                 return;
2816
2817         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2818         WARN_ON(cfq_cfqq_slice_new(cfqq));
2819
2820         /*
2821          * idle is disabled, either manually or by past process history
2822          */
2823         if (!cfq_should_idle(cfqd, cfqq)) {
2824                 /* no queue idling. Check for group idling */
2825                 if (cfqd->cfq_group_idle)
2826                         group_idle = cfqd->cfq_group_idle;
2827                 else
2828                         return;
2829         }
2830
2831         /*
2832          * still active requests from this queue, don't idle
2833          */
2834         if (cfqq->dispatched)
2835                 return;
2836
2837         /*
2838          * task has exited, don't wait
2839          */
2840         cic = cfqd->active_cic;
2841         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2842                 return;
2843
2844         /*
2845          * If our average think time is larger than the remaining time
2846          * slice, then don't idle. This avoids overrunning the allotted
2847          * time slice.
2848          */
2849         if (sample_valid(cic->ttime.ttime_samples) &&
2850             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2851                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2852                              cic->ttime.ttime_mean);
2853                 return;
2854         }
2855
2856         /* There are other queues in the group, don't do group idle */
2857         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2858                 return;
2859
2860         cfq_mark_cfqq_wait_request(cfqq);
2861
2862         if (group_idle)
2863                 sl = cfqd->cfq_group_idle;
2864         else
2865                 sl = cfqd->cfq_slice_idle;
2866
2867         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2868         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2869         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2870                         group_idle ? 1 : 0);
2871 }
2872
2873 /*
2874  * Move request from internal lists to the request queue dispatch list.
2875  */
2876 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2877 {
2878         struct cfq_data *cfqd = q->elevator->elevator_data;
2879         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2880
2881         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2882
2883         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2884         cfq_remove_request(rq);
2885         cfqq->dispatched++;
2886         (RQ_CFQG(rq))->dispatched++;
2887         elv_dispatch_sort(q, rq);
2888
2889         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2890         cfqq->nr_sectors += blk_rq_sectors(rq);
2891         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2892 }
2893
2894 /*
2895  * return expired entry, or NULL to just start from scratch in rbtree
2896  */
2897 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2898 {
2899         struct request *rq = NULL;
2900
2901         if (cfq_cfqq_fifo_expire(cfqq))
2902                 return NULL;
2903
2904         cfq_mark_cfqq_fifo_expire(cfqq);
2905
2906         if (list_empty(&cfqq->fifo))
2907                 return NULL;
2908
2909         rq = rq_entry_fifo(cfqq->fifo.next);
2910         if (time_before(jiffies, rq->fifo_time))
2911                 rq = NULL;
2912
2913         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2914         return rq;
2915 }
2916
2917 static inline int
2918 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2919 {
2920         const int base_rq = cfqd->cfq_slice_async_rq;
2921
2922         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2923
2924         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2925 }
2926
2927 /*
2928  * Must be called with the queue_lock held.
2929  */
2930 static int cfqq_process_refs(struct cfq_queue *cfqq)
2931 {
2932         int process_refs, io_refs;
2933
2934         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2935         process_refs = cfqq->ref - io_refs;
2936         BUG_ON(process_refs < 0);
2937         return process_refs;
2938 }
2939
2940 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2941 {
2942         int process_refs, new_process_refs;
2943         struct cfq_queue *__cfqq;
2944
2945         /*
2946          * If there are no process references on the new_cfqq, then it is
2947          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2948          * chain may have dropped their last reference (not just their
2949          * last process reference).
2950          */
2951         if (!cfqq_process_refs(new_cfqq))
2952                 return;
2953
2954         /* Avoid a circular list and skip interim queue merges */
2955         while ((__cfqq = new_cfqq->new_cfqq)) {
2956                 if (__cfqq == cfqq)
2957                         return;
2958                 new_cfqq = __cfqq;
2959         }
2960
2961         process_refs = cfqq_process_refs(cfqq);
2962         new_process_refs = cfqq_process_refs(new_cfqq);
2963         /*
2964          * If the process for the cfqq has gone away, there is no
2965          * sense in merging the queues.
2966          */
2967         if (process_refs == 0 || new_process_refs == 0)
2968                 return;
2969
2970         /*
2971          * Merge in the direction of the lesser amount of work.
2972          */
2973         if (new_process_refs >= process_refs) {
2974                 cfqq->new_cfqq = new_cfqq;
2975                 new_cfqq->ref += process_refs;
2976         } else {
2977                 new_cfqq->new_cfqq = cfqq;
2978                 cfqq->ref += new_process_refs;
2979         }
2980 }
2981
2982 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2983                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2984 {
2985         struct cfq_queue *queue;
2986         int i;
2987         bool key_valid = false;
2988         unsigned long lowest_key = 0;
2989         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2990
2991         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2992                 /* select the one with lowest rb_key */
2993                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2994                 if (queue &&
2995                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2996                         lowest_key = queue->rb_key;
2997                         cur_best = i;
2998                         key_valid = true;
2999                 }
3000         }
3001
3002         return cur_best;
3003 }
3004
3005 static void
3006 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3007 {
3008         unsigned slice;
3009         unsigned count;
3010         struct cfq_rb_root *st;
3011         unsigned group_slice;
3012         enum wl_class_t original_class = cfqd->serving_wl_class;
3013
3014         /* Choose next priority. RT > BE > IDLE */
3015         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3016                 cfqd->serving_wl_class = RT_WORKLOAD;
3017         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3018                 cfqd->serving_wl_class = BE_WORKLOAD;
3019         else {
3020                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3021                 cfqd->workload_expires = jiffies + 1;
3022                 return;
3023         }
3024
3025         if (original_class != cfqd->serving_wl_class)
3026                 goto new_workload;
3027
3028         /*
3029          * For RT and BE, we have to choose also the type
3030          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3031          * expiration time
3032          */
3033         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3034         count = st->count;
3035
3036         /*
3037          * check workload expiration, and that we still have other queues ready
3038          */
3039         if (count && !time_after(jiffies, cfqd->workload_expires))
3040                 return;
3041
3042 new_workload:
3043         /* otherwise select new workload type */
3044         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3045                                         cfqd->serving_wl_class);
3046         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3047         count = st->count;
3048
3049         /*
3050          * the workload slice is computed as a fraction of target latency
3051          * proportional to the number of queues in that workload, over
3052          * all the queues in the same priority class
3053          */
3054         group_slice = cfq_group_slice(cfqd, cfqg);
3055
3056         slice = group_slice * count /
3057                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3058                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3059                                         cfqg));
3060
3061         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3062                 unsigned int tmp;
3063
3064                 /*
3065                  * Async queues are currently system wide. Just taking
3066                  * proportion of queues with-in same group will lead to higher
3067                  * async ratio system wide as generally root group is going
3068                  * to have higher weight. A more accurate thing would be to
3069                  * calculate system wide asnc/sync ratio.
3070                  */
3071                 tmp = cfqd->cfq_target_latency *
3072                         cfqg_busy_async_queues(cfqd, cfqg);
3073                 tmp = tmp/cfqd->busy_queues;
3074                 slice = min_t(unsigned, slice, tmp);
3075
3076                 /* async workload slice is scaled down according to
3077                  * the sync/async slice ratio. */
3078                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
3079         } else
3080                 /* sync workload slice is at least 2 * cfq_slice_idle */
3081                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3082
3083         slice = max_t(unsigned, slice, CFQ_MIN_TT);
3084         cfq_log(cfqd, "workload slice:%d", slice);
3085         cfqd->workload_expires = jiffies + slice;
3086 }
3087
3088 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3089 {
3090         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3091         struct cfq_group *cfqg;
3092
3093         if (RB_EMPTY_ROOT(&st->rb))
3094                 return NULL;
3095         cfqg = cfq_rb_first_group(st);
3096         update_min_vdisktime(st);
3097         return cfqg;
3098 }
3099
3100 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3101 {
3102         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3103
3104         cfqd->serving_group = cfqg;
3105
3106         /* Restore the workload type data */
3107         if (cfqg->saved_wl_slice) {
3108                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
3109                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3110                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3111         } else
3112                 cfqd->workload_expires = jiffies - 1;
3113
3114         choose_wl_class_and_type(cfqd, cfqg);
3115 }
3116
3117 /*
3118  * Select a queue for service. If we have a current active queue,
3119  * check whether to continue servicing it, or retrieve and set a new one.
3120  */
3121 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3122 {
3123         struct cfq_queue *cfqq, *new_cfqq = NULL;
3124
3125         cfqq = cfqd->active_queue;
3126         if (!cfqq)
3127                 goto new_queue;
3128
3129         if (!cfqd->rq_queued)
3130                 return NULL;
3131
3132         /*
3133          * We were waiting for group to get backlogged. Expire the queue
3134          */
3135         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3136                 goto expire;
3137
3138         /*
3139          * The active queue has run out of time, expire it and select new.
3140          */
3141         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3142                 /*
3143                  * If slice had not expired at the completion of last request
3144                  * we might not have turned on wait_busy flag. Don't expire
3145                  * the queue yet. Allow the group to get backlogged.
3146                  *
3147                  * The very fact that we have used the slice, that means we
3148                  * have been idling all along on this queue and it should be
3149                  * ok to wait for this request to complete.
3150                  */
3151                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3152                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3153                         cfqq = NULL;
3154                         goto keep_queue;
3155                 } else
3156                         goto check_group_idle;
3157         }
3158
3159         /*
3160          * The active queue has requests and isn't expired, allow it to
3161          * dispatch.
3162          */
3163         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3164                 goto keep_queue;
3165
3166         /*
3167          * If another queue has a request waiting within our mean seek
3168          * distance, let it run.  The expire code will check for close
3169          * cooperators and put the close queue at the front of the service
3170          * tree.  If possible, merge the expiring queue with the new cfqq.
3171          */
3172         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3173         if (new_cfqq) {
3174                 if (!cfqq->new_cfqq)
3175                         cfq_setup_merge(cfqq, new_cfqq);
3176                 goto expire;
3177         }
3178
3179         /*
3180          * No requests pending. If the active queue still has requests in
3181          * flight or is idling for a new request, allow either of these
3182          * conditions to happen (or time out) before selecting a new queue.
3183          */
3184         if (timer_pending(&cfqd->idle_slice_timer)) {
3185                 cfqq = NULL;
3186                 goto keep_queue;
3187         }
3188
3189         /*
3190          * This is a deep seek queue, but the device is much faster than
3191          * the queue can deliver, don't idle
3192          **/
3193         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3194             (cfq_cfqq_slice_new(cfqq) ||
3195             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
3196                 cfq_clear_cfqq_deep(cfqq);
3197                 cfq_clear_cfqq_idle_window(cfqq);
3198         }
3199
3200         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3201                 cfqq = NULL;
3202                 goto keep_queue;
3203         }
3204
3205         /*
3206          * If group idle is enabled and there are requests dispatched from
3207          * this group, wait for requests to complete.
3208          */
3209 check_group_idle:
3210         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3211             cfqq->cfqg->dispatched &&
3212             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3213                 cfqq = NULL;
3214                 goto keep_queue;
3215         }
3216
3217 expire:
3218         cfq_slice_expired(cfqd, 0);
3219 new_queue:
3220         /*
3221          * Current queue expired. Check if we have to switch to a new
3222          * service tree
3223          */
3224         if (!new_cfqq)
3225                 cfq_choose_cfqg(cfqd);
3226
3227         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3228 keep_queue:
3229         return cfqq;
3230 }
3231
3232 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3233 {
3234         int dispatched = 0;
3235
3236         while (cfqq->next_rq) {
3237                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3238                 dispatched++;
3239         }
3240
3241         BUG_ON(!list_empty(&cfqq->fifo));
3242
3243         /* By default cfqq is not expired if it is empty. Do it explicitly */
3244         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3245         return dispatched;
3246 }
3247
3248 /*
3249  * Drain our current requests. Used for barriers and when switching
3250  * io schedulers on-the-fly.
3251  */
3252 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3253 {
3254         struct cfq_queue *cfqq;
3255         int dispatched = 0;
3256
3257         /* Expire the timeslice of the current active queue first */
3258         cfq_slice_expired(cfqd, 0);
3259         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
3260                 __cfq_set_active_queue(cfqd, cfqq);
3261                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
3262         }
3263
3264         BUG_ON(cfqd->busy_queues);
3265
3266         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3267         return dispatched;
3268 }
3269
3270 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3271         struct cfq_queue *cfqq)
3272 {
3273         /* the queue hasn't finished any request, can't estimate */
3274         if (cfq_cfqq_slice_new(cfqq))
3275                 return true;
3276         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3277                 cfqq->slice_end))
3278                 return true;
3279
3280         return false;
3281 }
3282
3283 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3284 {
3285         unsigned int max_dispatch;
3286
3287         /*
3288          * Drain async requests before we start sync IO
3289          */
3290         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3291                 return false;
3292
3293         /*
3294          * If this is an async queue and we have sync IO in flight, let it wait
3295          */
3296         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3297                 return false;
3298
3299         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3300         if (cfq_class_idle(cfqq))
3301                 max_dispatch = 1;
3302
3303         /*
3304          * Does this cfqq already have too much IO in flight?
3305          */
3306         if (cfqq->dispatched >= max_dispatch) {
3307                 bool promote_sync = false;
3308                 /*
3309                  * idle queue must always only have a single IO in flight
3310                  */
3311                 if (cfq_class_idle(cfqq))
3312                         return false;
3313
3314                 /*
3315                  * If there is only one sync queue
3316                  * we can ignore async queue here and give the sync
3317                  * queue no dispatch limit. The reason is a sync queue can
3318                  * preempt async queue, limiting the sync queue doesn't make
3319                  * sense. This is useful for aiostress test.
3320                  */
3321                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3322                         promote_sync = true;
3323
3324                 /*
3325                  * We have other queues, don't allow more IO from this one
3326                  */
3327                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3328                                 !promote_sync)
3329                         return false;
3330
3331                 /*
3332                  * Sole queue user, no limit
3333                  */
3334                 if (cfqd->busy_queues == 1 || promote_sync)
3335                         max_dispatch = -1;
3336                 else
3337                         /*
3338                          * Normally we start throttling cfqq when cfq_quantum/2
3339                          * requests have been dispatched. But we can drive
3340                          * deeper queue depths at the beginning of slice
3341                          * subjected to upper limit of cfq_quantum.
3342                          * */
3343                         max_dispatch = cfqd->cfq_quantum;
3344         }
3345
3346         /*
3347          * Async queues must wait a bit before being allowed dispatch.
3348          * We also ramp up the dispatch depth gradually for async IO,
3349          * based on the last sync IO we serviced
3350          */
3351         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3352                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3353                 unsigned int depth;
3354
3355                 depth = last_sync / cfqd->cfq_slice[1];
3356                 if (!depth && !cfqq->dispatched)
3357                         depth = 1;
3358                 if (depth < max_dispatch)
3359                         max_dispatch = depth;
3360         }
3361
3362         /*
3363          * If we're below the current max, allow a dispatch
3364          */
3365         return cfqq->dispatched < max_dispatch;
3366 }
3367
3368 /*
3369  * Dispatch a request from cfqq, moving them to the request queue
3370  * dispatch list.
3371  */
3372 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3373 {
3374         struct request *rq;
3375
3376         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3377
3378         if (!cfq_may_dispatch(cfqd, cfqq))
3379                 return false;
3380
3381         /*
3382          * follow expired path, else get first next available
3383          */
3384         rq = cfq_check_fifo(cfqq);
3385         if (!rq)
3386                 rq = cfqq->next_rq;
3387
3388         /*
3389          * insert request into driver dispatch list
3390          */
3391         cfq_dispatch_insert(cfqd->queue, rq);
3392
3393         if (!cfqd->active_cic) {
3394                 struct cfq_io_cq *cic = RQ_CIC(rq);
3395
3396                 atomic_long_inc(&cic->icq.ioc->refcount);
3397                 cfqd->active_cic = cic;
3398         }
3399
3400         return true;
3401 }
3402
3403 /*
3404  * Find the cfqq that we need to service and move a request from that to the
3405  * dispatch list
3406  */
3407 static int cfq_dispatch_requests(struct request_queue *q, int force)
3408 {
3409         struct cfq_data *cfqd = q->elevator->elevator_data;
3410         struct cfq_queue *cfqq;
3411
3412         if (!cfqd->busy_queues)
3413                 return 0;
3414
3415         if (unlikely(force))
3416                 return cfq_forced_dispatch(cfqd);
3417
3418         cfqq = cfq_select_queue(cfqd);
3419         if (!cfqq)
3420                 return 0;
3421
3422         /*
3423          * Dispatch a request from this cfqq, if it is allowed
3424          */
3425         if (!cfq_dispatch_request(cfqd, cfqq))
3426                 return 0;
3427
3428         cfqq->slice_dispatch++;
3429         cfq_clear_cfqq_must_dispatch(cfqq);
3430
3431         /*
3432          * expire an async queue immediately if it has used up its slice. idle
3433          * queue always expire after 1 dispatch round.
3434          */
3435         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3436             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3437             cfq_class_idle(cfqq))) {
3438                 cfqq->slice_end = jiffies + 1;
3439                 cfq_slice_expired(cfqd, 0);
3440         }
3441
3442         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3443         return 1;
3444 }
3445
3446 /*
3447  * task holds one reference to the queue, dropped when task exits. each rq
3448  * in-flight on this queue also holds a reference, dropped when rq is freed.
3449  *
3450  * Each cfq queue took a reference on the parent group. Drop it now.
3451  * queue lock must be held here.
3452  */
3453 static void cfq_put_queue(struct cfq_queue *cfqq)
3454 {
3455         struct cfq_data *cfqd = cfqq->cfqd;
3456         struct cfq_group *cfqg;
3457
3458         BUG_ON(cfqq->ref <= 0);
3459
3460         cfqq->ref--;
3461         if (cfqq->ref)
3462                 return;
3463
3464         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3465         BUG_ON(rb_first(&cfqq->sort_list));
3466         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3467         cfqg = cfqq->cfqg;
3468
3469         if (unlikely(cfqd->active_queue == cfqq)) {
3470                 __cfq_slice_expired(cfqd, cfqq, 0);
3471                 cfq_schedule_dispatch(cfqd);
3472         }
3473
3474         BUG_ON(cfq_cfqq_on_rr(cfqq));
3475         kmem_cache_free(cfq_pool, cfqq);
3476         cfqg_put(cfqg);
3477 }
3478
3479 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3480 {
3481         struct cfq_queue *__cfqq, *next;
3482
3483         /*
3484          * If this queue was scheduled to merge with another queue, be
3485          * sure to drop the reference taken on that queue (and others in
3486          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3487          */
3488         __cfqq = cfqq->new_cfqq;
3489         while (__cfqq) {
3490                 if (__cfqq == cfqq) {
3491                         WARN(1, "cfqq->new_cfqq loop detected\n");
3492                         break;
3493                 }
3494                 next = __cfqq->new_cfqq;
3495                 cfq_put_queue(__cfqq);
3496                 __cfqq = next;
3497         }
3498 }
3499
3500 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3501 {
3502         if (unlikely(cfqq == cfqd->active_queue)) {
3503                 __cfq_slice_expired(cfqd, cfqq, 0);
3504                 cfq_schedule_dispatch(cfqd);
3505         }
3506
3507         cfq_put_cooperator(cfqq);
3508
3509         cfq_put_queue(cfqq);
3510 }
3511
3512 static void cfq_init_icq(struct io_cq *icq)
3513 {
3514         struct cfq_io_cq *cic = icq_to_cic(icq);
3515
3516         cic->ttime.last_end_request = jiffies;
3517 }
3518
3519 static void cfq_exit_icq(struct io_cq *icq)
3520 {
3521         struct cfq_io_cq *cic = icq_to_cic(icq);
3522         struct cfq_data *cfqd = cic_to_cfqd(cic);
3523
3524         if (cic_to_cfqq(cic, false)) {
3525                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false));
3526                 cic_set_cfqq(cic, NULL, false);
3527         }
3528
3529         if (cic_to_cfqq(cic, true)) {
3530                 cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true));
3531                 cic_set_cfqq(cic, NULL, true);
3532         }
3533 }
3534
3535 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3536 {
3537         struct task_struct *tsk = current;
3538         int ioprio_class;
3539
3540         if (!cfq_cfqq_prio_changed(cfqq))
3541                 return;
3542
3543         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3544         switch (ioprio_class) {
3545         default:
3546                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3547         case IOPRIO_CLASS_NONE:
3548                 /*
3549                  * no prio set, inherit CPU scheduling settings
3550                  */
3551                 cfqq->ioprio = task_nice_ioprio(tsk);
3552                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3553                 break;
3554         case IOPRIO_CLASS_RT:
3555                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3556                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3557                 break;
3558         case IOPRIO_CLASS_BE:
3559                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3560                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3561                 break;
3562         case IOPRIO_CLASS_IDLE:
3563                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3564                 cfqq->ioprio = 7;
3565                 cfq_clear_cfqq_idle_window(cfqq);
3566                 break;
3567         }
3568
3569         /*
3570          * keep track of original prio settings in case we have to temporarily
3571          * elevate the priority of this queue
3572          */
3573         cfqq->org_ioprio = cfqq->ioprio;
3574         cfq_clear_cfqq_prio_changed(cfqq);
3575 }
3576
3577 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3578 {
3579         int ioprio = cic->icq.ioc->ioprio;
3580         struct cfq_data *cfqd = cic_to_cfqd(cic);
3581         struct cfq_queue *cfqq;
3582
3583         /*
3584          * Check whether ioprio has changed.  The condition may trigger
3585          * spuriously on a newly created cic but there's no harm.
3586          */
3587         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3588                 return;
3589
3590         cfqq = cic_to_cfqq(cic, false);
3591         if (cfqq) {
3592                 cfq_put_queue(cfqq);
3593                 cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio);
3594                 cic_set_cfqq(cic, cfqq, false);
3595         }
3596
3597         cfqq = cic_to_cfqq(cic, true);
3598         if (cfqq)
3599                 cfq_mark_cfqq_prio_changed(cfqq);
3600
3601         cic->ioprio = ioprio;
3602 }
3603
3604 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3605                           pid_t pid, bool is_sync)
3606 {
3607         RB_CLEAR_NODE(&cfqq->rb_node);
3608         RB_CLEAR_NODE(&cfqq->p_node);
3609         INIT_LIST_HEAD(&cfqq->fifo);
3610
3611         cfqq->ref = 0;
3612         cfqq->cfqd = cfqd;
3613
3614         cfq_mark_cfqq_prio_changed(cfqq);
3615
3616         if (is_sync) {
3617                 if (!cfq_class_idle(cfqq))
3618                         cfq_mark_cfqq_idle_window(cfqq);
3619                 cfq_mark_cfqq_sync(cfqq);
3620         }
3621         cfqq->pid = pid;
3622 }
3623
3624 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3625 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3626 {
3627         struct cfq_data *cfqd = cic_to_cfqd(cic);
3628         struct cfq_queue *cfqq;
3629         uint64_t serial_nr;
3630
3631         rcu_read_lock();
3632         serial_nr = bio_blkcg(bio)->css.serial_nr;
3633         rcu_read_unlock();
3634
3635         /*
3636          * Check whether blkcg has changed.  The condition may trigger
3637          * spuriously on a newly created cic but there's no harm.
3638          */
3639         if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr))
3640                 return;
3641
3642         /*
3643          * Drop reference to queues.  New queues will be assigned in new
3644          * group upon arrival of fresh requests.
3645          */
3646         cfqq = cic_to_cfqq(cic, false);
3647         if (cfqq) {
3648                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3649                 cic_set_cfqq(cic, NULL, false);
3650                 cfq_put_queue(cfqq);
3651         }
3652
3653         cfqq = cic_to_cfqq(cic, true);
3654         if (cfqq) {
3655                 cfq_log_cfqq(cfqd, cfqq, "changed cgroup");
3656                 cic_set_cfqq(cic, NULL, true);
3657                 cfq_put_queue(cfqq);
3658         }
3659
3660         cic->blkcg_serial_nr = serial_nr;
3661 }
3662 #else
3663 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3664 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3665
3666 static struct cfq_queue **
3667 cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio)
3668 {
3669         switch (ioprio_class) {
3670         case IOPRIO_CLASS_RT:
3671                 return &cfqg->async_cfqq[0][ioprio];
3672         case IOPRIO_CLASS_NONE:
3673                 ioprio = IOPRIO_NORM;
3674                 /* fall through */
3675         case IOPRIO_CLASS_BE:
3676                 return &cfqg->async_cfqq[1][ioprio];
3677         case IOPRIO_CLASS_IDLE:
3678                 return &cfqg->async_idle_cfqq;
3679         default:
3680                 BUG();
3681         }
3682 }
3683
3684 static struct cfq_queue *
3685 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3686               struct bio *bio)
3687 {
3688         int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3689         int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3690         struct cfq_queue **async_cfqq = NULL;
3691         struct cfq_queue *cfqq;
3692         struct cfq_group *cfqg;
3693
3694         rcu_read_lock();
3695         cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio));
3696         if (!cfqg) {
3697                 cfqq = &cfqd->oom_cfqq;
3698                 goto out;
3699         }
3700
3701         if (!is_sync) {
3702                 if (!ioprio_valid(cic->ioprio)) {
3703                         struct task_struct *tsk = current;
3704                         ioprio = task_nice_ioprio(tsk);
3705                         ioprio_class = task_nice_ioclass(tsk);
3706                 }
3707                 async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio);
3708                 cfqq = *async_cfqq;
3709                 if (cfqq)
3710                         goto out;
3711         }
3712
3713         cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO,
3714                                      cfqd->queue->node);
3715         if (!cfqq) {
3716                 cfqq = &cfqd->oom_cfqq;
3717                 goto out;
3718         }
3719
3720         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3721         cfq_init_prio_data(cfqq, cic);
3722         cfq_link_cfqq_cfqg(cfqq, cfqg);
3723         cfq_log_cfqq(cfqd, cfqq, "alloced");
3724
3725         if (async_cfqq) {
3726                 /* a new async queue is created, pin and remember */
3727                 cfqq->ref++;
3728                 *async_cfqq = cfqq;
3729         }
3730 out:
3731         cfqq->ref++;
3732         rcu_read_unlock();
3733         return cfqq;
3734 }
3735
3736 static void
3737 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3738 {
3739         unsigned long elapsed = jiffies - ttime->last_end_request;
3740         elapsed = min(elapsed, 2UL * slice_idle);
3741
3742         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3743         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3744         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3745 }
3746
3747 static void
3748 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3749                         struct cfq_io_cq *cic)
3750 {
3751         if (cfq_cfqq_sync(cfqq)) {
3752                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3753                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3754                         cfqd->cfq_slice_idle);
3755         }
3756 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3757         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3758 #endif
3759 }
3760
3761 static void
3762 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3763                        struct request *rq)
3764 {
3765         sector_t sdist = 0;
3766         sector_t n_sec = blk_rq_sectors(rq);
3767         if (cfqq->last_request_pos) {
3768                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3769                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3770                 else
3771                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3772         }
3773
3774         cfqq->seek_history <<= 1;
3775         if (blk_queue_nonrot(cfqd->queue))
3776                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3777         else
3778                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3779 }
3780
3781 /*
3782  * Disable idle window if the process thinks too long or seeks so much that
3783  * it doesn't matter
3784  */
3785 static void
3786 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3787                        struct cfq_io_cq *cic)
3788 {
3789         int old_idle, enable_idle;
3790
3791         /*
3792          * Don't idle for async or idle io prio class
3793          */
3794         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3795                 return;
3796
3797         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3798
3799         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3800                 cfq_mark_cfqq_deep(cfqq);
3801
3802         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3803                 enable_idle = 0;
3804         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3805                  !cfqd->cfq_slice_idle ||
3806                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3807                 enable_idle = 0;
3808         else if (sample_valid(cic->ttime.ttime_samples)) {
3809                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3810                         enable_idle = 0;
3811                 else
3812                         enable_idle = 1;
3813         }
3814
3815         if (old_idle != enable_idle) {
3816                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3817                 if (enable_idle)
3818                         cfq_mark_cfqq_idle_window(cfqq);
3819                 else
3820                         cfq_clear_cfqq_idle_window(cfqq);
3821         }
3822 }
3823
3824 /*
3825  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3826  * no or if we aren't sure, a 1 will cause a preempt.
3827  */
3828 static bool
3829 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3830                    struct request *rq)
3831 {
3832         struct cfq_queue *cfqq;
3833
3834         cfqq = cfqd->active_queue;
3835         if (!cfqq)
3836                 return false;
3837
3838         if (cfq_class_idle(new_cfqq))
3839                 return false;
3840
3841         if (cfq_class_idle(cfqq))
3842                 return true;
3843
3844         /*
3845          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3846          */
3847         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3848                 return false;
3849
3850         /*
3851          * if the new request is sync, but the currently running queue is
3852          * not, let the sync request have priority.
3853          */
3854         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3855                 return true;
3856
3857         if (new_cfqq->cfqg != cfqq->cfqg)
3858                 return false;
3859
3860         if (cfq_slice_used(cfqq))
3861                 return true;
3862
3863         /* Allow preemption only if we are idling on sync-noidle tree */
3864         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3865             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3866             new_cfqq->service_tree->count == 2 &&
3867             RB_EMPTY_ROOT(&cfqq->sort_list))
3868                 return true;
3869
3870         /*
3871          * So both queues are sync. Let the new request get disk time if
3872          * it's a metadata request and the current queue is doing regular IO.
3873          */
3874         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3875                 return true;
3876
3877         /*
3878          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3879          */
3880         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3881                 return true;
3882
3883         /* An idle queue should not be idle now for some reason */
3884         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3885                 return true;
3886
3887         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3888                 return false;
3889
3890         /*
3891          * if this request is as-good as one we would expect from the
3892          * current cfqq, let it preempt
3893          */
3894         if (cfq_rq_close(cfqd, cfqq, rq))
3895                 return true;
3896
3897         return false;
3898 }
3899
3900 /*
3901  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3902  * let it have half of its nominal slice.
3903  */
3904 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3905 {
3906         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3907
3908         cfq_log_cfqq(cfqd, cfqq, "preempt");
3909         cfq_slice_expired(cfqd, 1);
3910
3911         /*
3912          * workload type is changed, don't save slice, otherwise preempt
3913          * doesn't happen
3914          */
3915         if (old_type != cfqq_type(cfqq))
3916                 cfqq->cfqg->saved_wl_slice = 0;
3917
3918         /*
3919          * Put the new queue at the front of the of the current list,
3920          * so we know that it will be selected next.
3921          */
3922         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3923
3924         cfq_service_tree_add(cfqd, cfqq, 1);
3925
3926         cfqq->slice_end = 0;
3927         cfq_mark_cfqq_slice_new(cfqq);
3928 }
3929
3930 /*
3931  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3932  * something we should do about it
3933  */
3934 static void
3935 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3936                 struct request *rq)
3937 {
3938         struct cfq_io_cq *cic = RQ_CIC(rq);
3939
3940         cfqd->rq_queued++;
3941         if (rq->cmd_flags & REQ_PRIO)
3942                 cfqq->prio_pending++;
3943
3944         cfq_update_io_thinktime(cfqd, cfqq, cic);
3945         cfq_update_io_seektime(cfqd, cfqq, rq);
3946         cfq_update_idle_window(cfqd, cfqq, cic);
3947
3948         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3949
3950         if (cfqq == cfqd->active_queue) {
3951                 /*
3952                  * Remember that we saw a request from this process, but
3953                  * don't start queuing just yet. Otherwise we risk seeing lots
3954                  * of tiny requests, because we disrupt the normal plugging
3955                  * and merging. If the request is already larger than a single
3956                  * page, let it rip immediately. For that case we assume that
3957                  * merging is already done. Ditto for a busy system that
3958                  * has other work pending, don't risk delaying until the
3959                  * idle timer unplug to continue working.
3960                  */
3961                 if (cfq_cfqq_wait_request(cfqq)) {
3962                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3963                             cfqd->busy_queues > 1) {
3964                                 cfq_del_timer(cfqd, cfqq);
3965                                 cfq_clear_cfqq_wait_request(cfqq);
3966                                 __blk_run_queue(cfqd->queue);
3967                         } else {
3968                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3969                                 cfq_mark_cfqq_must_dispatch(cfqq);
3970                         }
3971                 }
3972         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3973                 /*
3974                  * not the active queue - expire current slice if it is
3975                  * idle and has expired it's mean thinktime or this new queue
3976                  * has some old slice time left and is of higher priority or
3977                  * this new queue is RT and the current one is BE
3978                  */
3979                 cfq_preempt_queue(cfqd, cfqq);
3980                 __blk_run_queue(cfqd->queue);
3981         }
3982 }
3983
3984 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3985 {
3986         struct cfq_data *cfqd = q->elevator->elevator_data;
3987         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3988
3989         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3990         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3991
3992         rq->fifo_time = jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)];
3993         list_add_tail(&rq->queuelist, &cfqq->fifo);
3994         cfq_add_rq_rb(rq);
3995         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3996                                  rq->cmd_flags);
3997         cfq_rq_enqueued(cfqd, cfqq, rq);
3998 }
3999
4000 /*
4001  * Update hw_tag based on peak queue depth over 50 samples under
4002  * sufficient load.
4003  */
4004 static void cfq_update_hw_tag(struct cfq_data *cfqd)
4005 {
4006         struct cfq_queue *cfqq = cfqd->active_queue;
4007
4008         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
4009                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
4010
4011         if (cfqd->hw_tag == 1)
4012                 return;
4013
4014         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
4015             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
4016                 return;
4017
4018         /*
4019          * If active queue hasn't enough requests and can idle, cfq might not
4020          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
4021          * case
4022          */
4023         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
4024             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
4025             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
4026                 return;
4027
4028         if (cfqd->hw_tag_samples++ < 50)
4029                 return;
4030
4031         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
4032                 cfqd->hw_tag = 1;
4033         else
4034                 cfqd->hw_tag = 0;
4035 }
4036
4037 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
4038 {
4039         struct cfq_io_cq *cic = cfqd->active_cic;
4040
4041         /* If the queue already has requests, don't wait */
4042         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4043                 return false;
4044
4045         /* If there are other queues in the group, don't wait */
4046         if (cfqq->cfqg->nr_cfqq > 1)
4047                 return false;
4048
4049         /* the only queue in the group, but think time is big */
4050         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
4051                 return false;
4052
4053         if (cfq_slice_used(cfqq))
4054                 return true;
4055
4056         /* if slice left is less than think time, wait busy */
4057         if (cic && sample_valid(cic->ttime.ttime_samples)
4058             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
4059                 return true;
4060
4061         /*
4062          * If think times is less than a jiffy than ttime_mean=0 and above
4063          * will not be true. It might happen that slice has not expired yet
4064          * but will expire soon (4-5 ns) during select_queue(). To cover the
4065          * case where think time is less than a jiffy, mark the queue wait
4066          * busy if only 1 jiffy is left in the slice.
4067          */
4068         if (cfqq->slice_end - jiffies == 1)
4069                 return true;
4070
4071         return false;
4072 }
4073
4074 static void cfq_completed_request(struct request_queue *q, struct request *rq)
4075 {
4076         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4077         struct cfq_data *cfqd = cfqq->cfqd;
4078         const int sync = rq_is_sync(rq);
4079         unsigned long now;
4080
4081         now = jiffies;
4082         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
4083                      !!(rq->cmd_flags & REQ_NOIDLE));
4084
4085         cfq_update_hw_tag(cfqd);
4086
4087         WARN_ON(!cfqd->rq_in_driver);
4088         WARN_ON(!cfqq->dispatched);
4089         cfqd->rq_in_driver--;
4090         cfqq->dispatched--;
4091         (RQ_CFQG(rq))->dispatched--;
4092         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
4093                                      rq_io_start_time_ns(rq), rq->cmd_flags);
4094
4095         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
4096
4097         if (sync) {
4098                 struct cfq_rb_root *st;
4099
4100                 RQ_CIC(rq)->ttime.last_end_request = now;
4101
4102                 if (cfq_cfqq_on_rr(cfqq))
4103                         st = cfqq->service_tree;
4104                 else
4105                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
4106                                         cfqq_type(cfqq));
4107
4108                 st->ttime.last_end_request = now;
4109                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
4110                         cfqd->last_delayed_sync = now;
4111         }
4112
4113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4114         cfqq->cfqg->ttime.last_end_request = now;
4115 #endif
4116
4117         /*
4118          * If this is the active queue, check if it needs to be expired,
4119          * or if we want to idle in case it has no pending requests.
4120          */
4121         if (cfqd->active_queue == cfqq) {
4122                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
4123
4124                 if (cfq_cfqq_slice_new(cfqq)) {
4125                         cfq_set_prio_slice(cfqd, cfqq);
4126                         cfq_clear_cfqq_slice_new(cfqq);
4127                 }
4128
4129                 /*
4130                  * Should we wait for next request to come in before we expire
4131                  * the queue.
4132                  */
4133                 if (cfq_should_wait_busy(cfqd, cfqq)) {
4134                         unsigned long extend_sl = cfqd->cfq_slice_idle;
4135                         if (!cfqd->cfq_slice_idle)
4136                                 extend_sl = cfqd->cfq_group_idle;
4137                         cfqq->slice_end = jiffies + extend_sl;
4138                         cfq_mark_cfqq_wait_busy(cfqq);
4139                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
4140                 }
4141
4142                 /*
4143                  * Idling is not enabled on:
4144                  * - expired queues
4145                  * - idle-priority queues
4146                  * - async queues
4147                  * - queues with still some requests queued
4148                  * - when there is a close cooperator
4149                  */
4150                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
4151                         cfq_slice_expired(cfqd, 1);
4152                 else if (sync && cfqq_empty &&
4153                          !cfq_close_cooperator(cfqd, cfqq)) {
4154                         cfq_arm_slice_timer(cfqd);
4155                 }
4156         }
4157
4158         if (!cfqd->rq_in_driver)
4159                 cfq_schedule_dispatch(cfqd);
4160 }
4161
4162 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
4163 {
4164         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
4165                 cfq_mark_cfqq_must_alloc_slice(cfqq);
4166                 return ELV_MQUEUE_MUST;
4167         }
4168
4169         return ELV_MQUEUE_MAY;
4170 }
4171
4172 static int cfq_may_queue(struct request_queue *q, int rw)
4173 {
4174         struct cfq_data *cfqd = q->elevator->elevator_data;
4175         struct task_struct *tsk = current;
4176         struct cfq_io_cq *cic;
4177         struct cfq_queue *cfqq;
4178
4179         /*
4180          * don't force setup of a queue from here, as a call to may_queue
4181          * does not necessarily imply that a request actually will be queued.
4182          * so just lookup a possibly existing queue, or return 'may queue'
4183          * if that fails
4184          */
4185         cic = cfq_cic_lookup(cfqd, tsk->io_context);
4186         if (!cic)
4187                 return ELV_MQUEUE_MAY;
4188
4189         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
4190         if (cfqq) {
4191                 cfq_init_prio_data(cfqq, cic);
4192
4193                 return __cfq_may_queue(cfqq);
4194         }
4195
4196         return ELV_MQUEUE_MAY;
4197 }
4198
4199 /*
4200  * queue lock held here
4201  */
4202 static void cfq_put_request(struct request *rq)
4203 {
4204         struct cfq_queue *cfqq = RQ_CFQQ(rq);
4205
4206         if (cfqq) {
4207                 const int rw = rq_data_dir(rq);
4208
4209                 BUG_ON(!cfqq->allocated[rw]);
4210                 cfqq->allocated[rw]--;
4211
4212                 /* Put down rq reference on cfqg */
4213                 cfqg_put(RQ_CFQG(rq));
4214                 rq->elv.priv[0] = NULL;
4215                 rq->elv.priv[1] = NULL;
4216
4217                 cfq_put_queue(cfqq);
4218         }
4219 }
4220
4221 static struct cfq_queue *
4222 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
4223                 struct cfq_queue *cfqq)
4224 {
4225         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
4226         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
4227         cfq_mark_cfqq_coop(cfqq->new_cfqq);
4228         cfq_put_queue(cfqq);
4229         return cic_to_cfqq(cic, 1);
4230 }
4231
4232 /*
4233  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4234  * was the last process referring to said cfqq.
4235  */
4236 static struct cfq_queue *
4237 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4238 {
4239         if (cfqq_process_refs(cfqq) == 1) {
4240                 cfqq->pid = current->pid;
4241                 cfq_clear_cfqq_coop(cfqq);
4242                 cfq_clear_cfqq_split_coop(cfqq);
4243                 return cfqq;
4244         }
4245
4246         cic_set_cfqq(cic, NULL, 1);
4247
4248         cfq_put_cooperator(cfqq);
4249
4250         cfq_put_queue(cfqq);
4251         return NULL;
4252 }
4253 /*
4254  * Allocate cfq data structures associated with this request.
4255  */
4256 static int
4257 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4258                 gfp_t gfp_mask)
4259 {
4260         struct cfq_data *cfqd = q->elevator->elevator_data;
4261         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4262         const int rw = rq_data_dir(rq);
4263         const bool is_sync = rq_is_sync(rq);
4264         struct cfq_queue *cfqq;
4265
4266         spin_lock_irq(q->queue_lock);
4267
4268         check_ioprio_changed(cic, bio);
4269         check_blkcg_changed(cic, bio);
4270 new_queue:
4271         cfqq = cic_to_cfqq(cic, is_sync);
4272         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4273                 if (cfqq)
4274                         cfq_put_queue(cfqq);
4275                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio);
4276                 cic_set_cfqq(cic, cfqq, is_sync);
4277         } else {
4278                 /*
4279                  * If the queue was seeky for too long, break it apart.
4280                  */
4281                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4282                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4283                         cfqq = split_cfqq(cic, cfqq);
4284                         if (!cfqq)
4285                                 goto new_queue;
4286                 }
4287
4288                 /*
4289                  * Check to see if this queue is scheduled to merge with
4290                  * another, closely cooperating queue.  The merging of
4291                  * queues happens here as it must be done in process context.
4292                  * The reference on new_cfqq was taken in merge_cfqqs.
4293                  */
4294                 if (cfqq->new_cfqq)
4295                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4296         }
4297
4298         cfqq->allocated[rw]++;
4299
4300         cfqq->ref++;
4301         cfqg_get(cfqq->cfqg);
4302         rq->elv.priv[0] = cfqq;
4303         rq->elv.priv[1] = cfqq->cfqg;
4304         spin_unlock_irq(q->queue_lock);
4305         return 0;
4306 }
4307
4308 static void cfq_kick_queue(struct work_struct *work)
4309 {
4310         struct cfq_data *cfqd =
4311                 container_of(work, struct cfq_data, unplug_work);
4312         struct request_queue *q = cfqd->queue;
4313
4314         spin_lock_irq(q->queue_lock);
4315         __blk_run_queue(cfqd->queue);
4316         spin_unlock_irq(q->queue_lock);
4317 }
4318
4319 /*
4320  * Timer running if the active_queue is currently idling inside its time slice
4321  */
4322 static void cfq_idle_slice_timer(unsigned long data)
4323 {
4324         struct cfq_data *cfqd = (struct cfq_data *) data;
4325         struct cfq_queue *cfqq;
4326         unsigned long flags;
4327         int timed_out = 1;
4328
4329         cfq_log(cfqd, "idle timer fired");
4330
4331         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4332
4333         cfqq = cfqd->active_queue;
4334         if (cfqq) {
4335                 timed_out = 0;
4336
4337                 /*
4338                  * We saw a request before the queue expired, let it through
4339                  */
4340                 if (cfq_cfqq_must_dispatch(cfqq))
4341                         goto out_kick;
4342
4343                 /*
4344                  * expired
4345                  */
4346                 if (cfq_slice_used(cfqq))
4347                         goto expire;
4348
4349                 /*
4350                  * only expire and reinvoke request handler, if there are
4351                  * other queues with pending requests
4352                  */
4353                 if (!cfqd->busy_queues)
4354                         goto out_cont;
4355
4356                 /*
4357                  * not expired and it has a request pending, let it dispatch
4358                  */
4359                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4360                         goto out_kick;
4361
4362                 /*
4363                  * Queue depth flag is reset only when the idle didn't succeed
4364                  */
4365                 cfq_clear_cfqq_deep(cfqq);
4366         }
4367 expire:
4368         cfq_slice_expired(cfqd, timed_out);
4369 out_kick:
4370         cfq_schedule_dispatch(cfqd);
4371 out_cont:
4372         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4373 }
4374
4375 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4376 {
4377         del_timer_sync(&cfqd->idle_slice_timer);
4378         cancel_work_sync(&cfqd->unplug_work);
4379 }
4380
4381 static void cfq_exit_queue(struct elevator_queue *e)
4382 {
4383         struct cfq_data *cfqd = e->elevator_data;
4384         struct request_queue *q = cfqd->queue;
4385
4386         cfq_shutdown_timer_wq(cfqd);
4387
4388         spin_lock_irq(q->queue_lock);
4389
4390         if (cfqd->active_queue)
4391                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4392
4393         spin_unlock_irq(q->queue_lock);
4394
4395         cfq_shutdown_timer_wq(cfqd);
4396
4397 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4398         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4399 #else
4400         kfree(cfqd->root_group);
4401 #endif
4402         kfree(cfqd);
4403 }
4404
4405 static int cfq_init_queue(struct request_queue *q, struct elevator_type *e)
4406 {
4407         struct cfq_data *cfqd;
4408         struct blkcg_gq *blkg __maybe_unused;
4409         int i, ret;
4410         struct elevator_queue *eq;
4411
4412         eq = elevator_alloc(q, e);
4413         if (!eq)
4414                 return -ENOMEM;
4415
4416         cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
4417         if (!cfqd) {
4418                 kobject_put(&eq->kobj);
4419                 return -ENOMEM;
4420         }
4421         eq->elevator_data = cfqd;
4422
4423         cfqd->queue = q;
4424         spin_lock_irq(q->queue_lock);
4425         q->elevator = eq;
4426         spin_unlock_irq(q->queue_lock);
4427
4428         /* Init root service tree */
4429         cfqd->grp_service_tree = CFQ_RB_ROOT;
4430
4431         /* Init root group and prefer root group over other groups by default */
4432 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4433         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4434         if (ret)
4435                 goto out_free;
4436
4437         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4438 #else
4439         ret = -ENOMEM;
4440         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4441                                         GFP_KERNEL, cfqd->queue->node);
4442         if (!cfqd->root_group)
4443                 goto out_free;
4444
4445         cfq_init_cfqg_base(cfqd->root_group);
4446 #endif
4447         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4448         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4449
4450         /*
4451          * Not strictly needed (since RB_ROOT just clears the node and we
4452          * zeroed cfqd on alloc), but better be safe in case someone decides
4453          * to add magic to the rb code
4454          */
4455         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4456                 cfqd->prio_trees[i] = RB_ROOT;
4457
4458         /*
4459          * Our fallback cfqq if cfq_get_queue() runs into OOM issues.
4460          * Grab a permanent reference to it, so that the normal code flow
4461          * will not attempt to free it.  oom_cfqq is linked to root_group
4462          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4463          * the reference from linking right away.
4464          */
4465         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4466         cfqd->oom_cfqq.ref++;
4467
4468         spin_lock_irq(q->queue_lock);
4469         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4470         cfqg_put(cfqd->root_group);
4471         spin_unlock_irq(q->queue_lock);
4472
4473         init_timer(&cfqd->idle_slice_timer);
4474         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4475         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4476
4477         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4478
4479         cfqd->cfq_quantum = cfq_quantum;
4480         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4481         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4482         cfqd->cfq_back_max = cfq_back_max;
4483         cfqd->cfq_back_penalty = cfq_back_penalty;
4484         cfqd->cfq_slice[0] = cfq_slice_async;
4485         cfqd->cfq_slice[1] = cfq_slice_sync;
4486         cfqd->cfq_target_latency = cfq_target_latency;
4487         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4488         cfqd->cfq_slice_idle = cfq_slice_idle;
4489         cfqd->cfq_group_idle = cfq_group_idle;
4490         cfqd->cfq_latency = 1;
4491         cfqd->hw_tag = -1;
4492         /*
4493          * we optimistically start assuming sync ops weren't delayed in last
4494          * second, in order to have larger depth for async operations.
4495          */
4496         cfqd->last_delayed_sync = jiffies - HZ;
4497         return 0;
4498
4499 out_free:
4500         kfree(cfqd);
4501         kobject_put(&eq->kobj);
4502         return ret;
4503 }
4504
4505 static void cfq_registered_queue(struct request_queue *q)
4506 {
4507         struct elevator_queue *e = q->elevator;
4508         struct cfq_data *cfqd = e->elevator_data;
4509
4510         /*
4511          * Default to IOPS mode with no idling for SSDs
4512          */
4513         if (blk_queue_nonrot(q))
4514                 cfqd->cfq_slice_idle = 0;
4515 }
4516
4517 /*
4518  * sysfs parts below -->
4519  */
4520 static ssize_t
4521 cfq_var_show(unsigned int var, char *page)
4522 {
4523         return sprintf(page, "%u\n", var);
4524 }
4525
4526 static ssize_t
4527 cfq_var_store(unsigned int *var, const char *page, size_t count)
4528 {
4529         char *p = (char *) page;
4530
4531         *var = simple_strtoul(p, &p, 10);
4532         return count;
4533 }
4534
4535 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4536 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4537 {                                                                       \
4538         struct cfq_data *cfqd = e->elevator_data;                       \
4539         unsigned int __data = __VAR;                                    \
4540         if (__CONV)                                                     \
4541                 __data = jiffies_to_msecs(__data);                      \
4542         return cfq_var_show(__data, (page));                            \
4543 }
4544 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4545 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4546 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4547 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4548 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4549 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4550 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4551 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4552 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4553 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4554 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4555 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4556 #undef SHOW_FUNCTION
4557
4558 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4559 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4560 {                                                                       \
4561         struct cfq_data *cfqd = e->elevator_data;                       \
4562         unsigned int __data;                                            \
4563         int ret = cfq_var_store(&__data, (page), count);                \
4564         if (__data < (MIN))                                             \
4565                 __data = (MIN);                                         \
4566         else if (__data > (MAX))                                        \
4567                 __data = (MAX);                                         \
4568         if (__CONV)                                                     \
4569                 *(__PTR) = msecs_to_jiffies(__data);                    \
4570         else                                                            \
4571                 *(__PTR) = __data;                                      \
4572         return ret;                                                     \
4573 }
4574 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4575 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4576                 UINT_MAX, 1);
4577 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4578                 UINT_MAX, 1);
4579 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4580 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4581                 UINT_MAX, 0);
4582 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4583 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4584 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4585 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4586 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4587                 UINT_MAX, 0);
4588 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4589 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4590 #undef STORE_FUNCTION
4591
4592 #define CFQ_ATTR(name) \
4593         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4594
4595 static struct elv_fs_entry cfq_attrs[] = {
4596         CFQ_ATTR(quantum),
4597         CFQ_ATTR(fifo_expire_sync),
4598         CFQ_ATTR(fifo_expire_async),
4599         CFQ_ATTR(back_seek_max),
4600         CFQ_ATTR(back_seek_penalty),
4601         CFQ_ATTR(slice_sync),
4602         CFQ_ATTR(slice_async),
4603         CFQ_ATTR(slice_async_rq),
4604         CFQ_ATTR(slice_idle),
4605         CFQ_ATTR(group_idle),
4606         CFQ_ATTR(low_latency),
4607         CFQ_ATTR(target_latency),
4608         __ATTR_NULL
4609 };
4610
4611 static struct elevator_type iosched_cfq = {
4612         .ops = {
4613                 .elevator_merge_fn =            cfq_merge,
4614                 .elevator_merged_fn =           cfq_merged_request,
4615                 .elevator_merge_req_fn =        cfq_merged_requests,
4616                 .elevator_allow_merge_fn =      cfq_allow_merge,
4617                 .elevator_bio_merged_fn =       cfq_bio_merged,
4618                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4619                 .elevator_add_req_fn =          cfq_insert_request,
4620                 .elevator_activate_req_fn =     cfq_activate_request,
4621                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4622                 .elevator_completed_req_fn =    cfq_completed_request,
4623                 .elevator_former_req_fn =       elv_rb_former_request,
4624                 .elevator_latter_req_fn =       elv_rb_latter_request,
4625                 .elevator_init_icq_fn =         cfq_init_icq,
4626                 .elevator_exit_icq_fn =         cfq_exit_icq,
4627                 .elevator_set_req_fn =          cfq_set_request,
4628                 .elevator_put_req_fn =          cfq_put_request,
4629                 .elevator_may_queue_fn =        cfq_may_queue,
4630                 .elevator_init_fn =             cfq_init_queue,
4631                 .elevator_exit_fn =             cfq_exit_queue,
4632                 .elevator_registered_fn =       cfq_registered_queue,
4633         },
4634         .icq_size       =       sizeof(struct cfq_io_cq),
4635         .icq_align      =       __alignof__(struct cfq_io_cq),
4636         .elevator_attrs =       cfq_attrs,
4637         .elevator_name  =       "cfq",
4638         .elevator_owner =       THIS_MODULE,
4639 };
4640
4641 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4642 static struct blkcg_policy blkcg_policy_cfq = {
4643         .cftypes                = cfq_blkcg_files,
4644
4645         .cpd_alloc_fn           = cfq_cpd_alloc,
4646         .cpd_init_fn            = cfq_cpd_init,
4647         .cpd_free_fn            = cfq_cpd_free,
4648
4649         .pd_alloc_fn            = cfq_pd_alloc,
4650         .pd_init_fn             = cfq_pd_init,
4651         .pd_offline_fn          = cfq_pd_offline,
4652         .pd_free_fn             = cfq_pd_free,
4653         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4654 };
4655 #endif
4656
4657 static int __init cfq_init(void)
4658 {
4659         int ret;
4660
4661         /*
4662          * could be 0 on HZ < 1000 setups
4663          */
4664         if (!cfq_slice_async)
4665                 cfq_slice_async = 1;
4666         if (!cfq_slice_idle)
4667                 cfq_slice_idle = 1;
4668
4669 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4670         if (!cfq_group_idle)
4671                 cfq_group_idle = 1;
4672
4673         ret = blkcg_policy_register(&blkcg_policy_cfq);
4674         if (ret)
4675                 return ret;
4676 #else
4677         cfq_group_idle = 0;
4678 #endif
4679
4680         ret = -ENOMEM;
4681         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4682         if (!cfq_pool)
4683                 goto err_pol_unreg;
4684
4685         ret = elv_register(&iosched_cfq);
4686         if (ret)
4687                 goto err_free_pool;
4688
4689         return 0;
4690
4691 err_free_pool:
4692         kmem_cache_destroy(cfq_pool);
4693 err_pol_unreg:
4694 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4695         blkcg_policy_unregister(&blkcg_policy_cfq);
4696 #endif
4697         return ret;
4698 }
4699
4700 static void __exit cfq_exit(void)
4701 {
4702 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4703         blkcg_policy_unregister(&blkcg_policy_cfq);
4704 #endif
4705         elv_unregister(&iosched_cfq);
4706         kmem_cache_destroy(cfq_pool);
4707 }
4708
4709 module_init(cfq_init);
4710 module_exit(cfq_exit);
4711
4712 MODULE_AUTHOR("Jens Axboe");
4713 MODULE_LICENSE("GPL");
4714 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");