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/blkdev.h>
  11 #include <linux/elevator.h>
  12 #include <linux/jiffies.h>
  13 #include <linux/rbtree.h>
  14 #include <linux/ioprio.h>
  15 #include <linux/blktrace_api.h>
  16 #include "blk-cgroup.h"
  17
  18 /*
  19  * tunables
  20  */
  21 /* max queue in one round of service */
  22 static const int cfq_quantum = 4;
  23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
  24 /* maximum backwards seek, in KiB */
  25 static const int cfq_back_max = 16 * 1024;
  26 /* penalty of a backwards seek */
  27 static const int cfq_back_penalty = 2;
  28 static const int cfq_slice_sync = HZ / 10;
  29 static int cfq_slice_async = HZ / 25;
  30 static const int cfq_slice_async_rq = 2;
  31 static int cfq_slice_idle = HZ / 125;
  32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
  33 static const int cfq_hist_divisor = 4;
  34
  35 /*
  36  * offset from end of service tree
  37  */
  38 #define CFQ_IDLE_DELAY          (HZ / 5)
  39
  40 /*
  41  * below this threshold, we consider thinktime immediate
  42  */
  43 #define CFQ_MIN_TT              (2)
  44
  45 /*
  46  * Allow merged cfqqs to perform this amount of seeky I/O before
  47  * deciding to break the queues up again.
  48  */
  49 #define CFQQ_COOP_TOUT          (HZ)
  50
  51 #define CFQ_SLICE_SCALE         (5)
  52 #define CFQ_HW_QUEUE_MIN        (5)
  53 #define CFQ_SERVICE_SHIFT       12
  54
  55 #define RQ_CIC(rq)              \
  56         ((struct cfq_io_context *) (rq)->elevator_private)
  57 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
  58
  59 static struct kmem_cache *cfq_pool;
  60 static struct kmem_cache *cfq_ioc_pool;
  61
  62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
  63 static struct completion *ioc_gone;
  64 static DEFINE_SPINLOCK(ioc_gone_lock);
  65
  66 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
  67 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
  68 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
  69
  70 #define sample_valid(samples)   ((samples) > 80)
  71 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
  72
  73 /*
  74  * Most of our rbtree usage is for sorting with min extraction, so
  75  * if we cache the leftmost node we don't have to walk down the tree
  76  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
  77  * move this into the elevator for the rq sorting as well.
  78  */
  79 struct cfq_rb_root {
  80         struct rb_root rb;
  81         struct rb_node *left;
  82         unsigned count;
  83         u64 min_vdisktime;
  84         struct rb_node *active;
  85         unsigned total_weight;
  86 };
  87 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
  88
  89 /*
  90  * Per process-grouping structure
  91  */
  92 struct cfq_queue {
  93         /* reference count */
  94         atomic_t ref;
  95         /* various state flags, see below */
  96         unsigned int flags;
  97         /* parent cfq_data */
  98         struct cfq_data *cfqd;
  99         /* service_tree member */
 100         struct rb_node rb_node;
 101         /* service_tree key */
 102         unsigned long rb_key;
 103         /* prio tree member */
 104         struct rb_node p_node;
 105         /* prio tree root we belong to, if any */
 106         struct rb_root *p_root;
 107         /* sorted list of pending requests */
 108         struct rb_root sort_list;
 109         /* if fifo isn't expired, next request to serve */
 110         struct request *next_rq;
 111         /* requests queued in sort_list */
 112         int queued[2];
 113         /* currently allocated requests */
 114         int allocated[2];
 115         /* fifo list of requests in sort_list */
 116         struct list_head fifo;
 117
 118         /* time when queue got scheduled in to dispatch first request. */
 119         unsigned long dispatch_start;
 120         unsigned int allocated_slice;
 121         /* time when first request from queue completed and slice started. */
 122         unsigned long slice_start;
 123         unsigned long slice_end;
 124         long slice_resid;
 125         unsigned int slice_dispatch;
 126
 127         /* pending metadata requests */
 128         int meta_pending;
 129         /* number of requests that are on the dispatch list or inside driver */
 130         int dispatched;
 131
 132         /* io prio of this group */
 133         unsigned short ioprio, org_ioprio;
 134         unsigned short ioprio_class, org_ioprio_class;
 135
 136         unsigned int seek_samples;
 137         u64 seek_total;
 138         sector_t seek_mean;
 139         sector_t last_request_pos;
 140         unsigned long seeky_start;
 141
 142         pid_t pid;
 143
 144         struct cfq_rb_root *service_tree;
 145         struct cfq_queue *new_cfqq;
 146         struct cfq_group *cfqg;
 147         struct cfq_group *orig_cfqg;
 148         /* Sectors dispatched in current dispatch round */
 149         unsigned long nr_sectors;
 150 };
 151
 152 /*
 153  * First index in the service_trees.
 154  * IDLE is handled separately, so it has negative index
 155  */
 156 enum wl_prio_t {
 157         BE_WORKLOAD = 0,
 158         RT_WORKLOAD = 1,
 159         IDLE_WORKLOAD = 2,
 160 };
 161
 162 /*
 163  * Second index in the service_trees.
 164  */
 165 enum wl_type_t {
 166         ASYNC_WORKLOAD = 0,
 167         SYNC_NOIDLE_WORKLOAD = 1,
 168         SYNC_WORKLOAD = 2
 169 };
 170
 171 /* This is per cgroup per device grouping structure */
 172 struct cfq_group {
 173         /* group service_tree member */
 174         struct rb_node rb_node;
 175
 176         /* group service_tree key */
 177         u64 vdisktime;
 178         unsigned int weight;
 179         bool on_st;
 180
 181         /* number of cfqq currently on this group */
 182         int nr_cfqq;
 183
 184         /* Per group busy queus average. Useful for workload slice calc. */
 185         unsigned int busy_queues_avg[2];
 186         /*
 187          * rr lists of queues with requests, onle rr for each priority class.
 188          * Counts are embedded in the cfq_rb_root
 189          */
 190         struct cfq_rb_root service_trees[2][3];
 191         struct cfq_rb_root service_tree_idle;
 192
 193         unsigned long saved_workload_slice;
 194         enum wl_type_t saved_workload;
 195         enum wl_prio_t saved_serving_prio;
 196         struct blkio_group blkg;
 197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 198         struct hlist_node cfqd_node;
 199         atomic_t ref;
 200 #endif
 201 };
 202
 203 /*
 204  * Per block device queue structure
 205  */
 206 struct cfq_data {
 207         struct request_queue *queue;
 208         /* Root service tree for cfq_groups */
 209         struct cfq_rb_root grp_service_tree;
 210         struct cfq_group root_group;
 211         /* Number of active cfq groups on group service tree */
 212         int nr_groups;
 213
 214         /*
 215          * The priority currently being served
 216          */
 217         enum wl_prio_t serving_prio;
 218         enum wl_type_t serving_type;
 219         unsigned long workload_expires;
 220         struct cfq_group *serving_group;
 221         bool noidle_tree_requires_idle;
 222
 223         /*
 224          * Each priority tree is sorted by next_request position.  These
 225          * trees are used when determining if two or more queues are
 226          * interleaving requests (see cfq_close_cooperator).
 227          */
 228         struct rb_root prio_trees[CFQ_PRIO_LISTS];
 229
 230         unsigned int busy_queues;
 231
 232         int rq_in_driver[2];
 233         int sync_flight;
 234
 235         /*
 236          * queue-depth detection
 237          */
 238         int rq_queued;
 239         int hw_tag;
 240         /*
 241          * hw_tag can be
 242          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
 243          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
 244          *  0 => no NCQ
 245          */
 246         int hw_tag_est_depth;
 247         unsigned int hw_tag_samples;
 248
 249         /*
 250          * idle window management
 251          */
 252         struct timer_list idle_slice_timer;
 253         struct work_struct unplug_work;
 254
 255         struct cfq_queue *active_queue;
 256         struct cfq_io_context *active_cic;
 257
 258         /*
 259          * async queue for each priority case
 260          */
 261         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
 262         struct cfq_queue *async_idle_cfqq;
 263
 264         sector_t last_position;
 265
 266         /*
 267          * tunables, see top of file
 268          */
 269         unsigned int cfq_quantum;
 270         unsigned int cfq_fifo_expire[2];
 271         unsigned int cfq_back_penalty;
 272         unsigned int cfq_back_max;
 273         unsigned int cfq_slice[2];
 274         unsigned int cfq_slice_async_rq;
 275         unsigned int cfq_slice_idle;
 276         unsigned int cfq_latency;
 277         unsigned int cfq_group_isolation;
 278
 279         struct list_head cic_list;
 280
 281         /*
 282          * Fallback dummy cfqq for extreme OOM conditions
 283          */
 284         struct cfq_queue oom_cfqq;
 285
 286         unsigned long last_delayed_sync;
 287
 288         /* List of cfq groups being managed on this device*/
 289         struct hlist_head cfqg_list;
 290         struct rcu_head rcu;
 291 };
 292
 293 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
 294
 295 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
 296                                             enum wl_prio_t prio,
 297                                             enum wl_type_t type,
 298                                             struct cfq_data *cfqd)
 299 {
 300         if (!cfqg)
 301                 return NULL;
 302
 303         if (prio == IDLE_WORKLOAD)
 304                 return &cfqg->service_tree_idle;
 305
 306         return &cfqg->service_trees[prio][type];
 307 }
 308
 309 enum cfqq_state_flags {
 310         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
 311         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
 312         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
 313         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
 314         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
 315         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
 316         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
 317         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
 318         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
 319         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
 320         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
 321         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
 322 };
 323
 324 #define CFQ_CFQQ_FNS(name)                                              \
 325 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
 326 {                                                                       \
 327         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
 328 }                                                                       \
 329 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
 330 {                                                                       \
 331         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
 332 }                                                                       \
 333 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
 334 {                                                                       \
 335         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
 336 }
 337
 338 CFQ_CFQQ_FNS(on_rr);
 339 CFQ_CFQQ_FNS(wait_request);
 340 CFQ_CFQQ_FNS(must_dispatch);
 341 CFQ_CFQQ_FNS(must_alloc_slice);
 342 CFQ_CFQQ_FNS(fifo_expire);
 343 CFQ_CFQQ_FNS(idle_window);
 344 CFQ_CFQQ_FNS(prio_changed);
 345 CFQ_CFQQ_FNS(slice_new);
 346 CFQ_CFQQ_FNS(sync);
 347 CFQ_CFQQ_FNS(coop);
 348 CFQ_CFQQ_FNS(deep);
 349 CFQ_CFQQ_FNS(wait_busy);
 350 #undef CFQ_CFQQ_FNS
 351
 352 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
 353 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 354         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
 355                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
 356                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
 357
 358 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
 359         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
 360                                 blkg_path(&(cfqg)->blkg), ##args);      \
 361
 362 #else
 363 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
 364         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
 365 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
 366 #endif
 367 #define cfq_log(cfqd, fmt, args...)     \
 368         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
 369
 370 /* Traverses through cfq group service trees */
 371 #define for_each_cfqg_st(cfqg, i, j, st) \
 372         for (i = 0; i <= IDLE_WORKLOAD; i++) \
 373                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
 374                         : &cfqg->service_tree_idle; \
 375                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
 376                         (i == IDLE_WORKLOAD && j == 0); \
 377                         j++, st = i < IDLE_WORKLOAD ? \
 378                         &cfqg->service_trees[i][j]: NULL) \
 379
 380
 381 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
 382 {
 383         if (cfq_class_idle(cfqq))
 384                 return IDLE_WORKLOAD;
 385         if (cfq_class_rt(cfqq))
 386                 return RT_WORKLOAD;
 387         return BE_WORKLOAD;
 388 }
 389
 390
 391 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
 392 {
 393         if (!cfq_cfqq_sync(cfqq))
 394                 return ASYNC_WORKLOAD;
 395         if (!cfq_cfqq_idle_window(cfqq))
 396                 return SYNC_NOIDLE_WORKLOAD;
 397         return SYNC_WORKLOAD;
 398 }
 399
 400 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
 401                                         struct cfq_data *cfqd,
 402                                         struct cfq_group *cfqg)
 403 {
 404         if (wl == IDLE_WORKLOAD)
 405                 return cfqg->service_tree_idle.count;
 406
 407         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
 408                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
 409                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
 410 }
 411
 412 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
 413                                         struct cfq_group *cfqg)
 414 {
 415         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
 416                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
 417 }
 418
 419 static void cfq_dispatch_insert(struct request_queue *, struct request *);
 420 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
 421                                        struct io_context *, gfp_t);
 422 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
 423                                                 struct io_context *);
 424
 425 static inline int rq_in_driver(struct cfq_data *cfqd)
 426 {
 427         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
 428 }
 429
 430 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
 431                                             bool is_sync)
 432 {
 433         return cic->cfqq[is_sync];
 434 }
 435
 436 static inline void cic_set_cfqq(struct cfq_io_context *cic,
 437                                 struct cfq_queue *cfqq, bool is_sync)
 438 {
 439         cic->cfqq[is_sync] = cfqq;
 440 }
 441
 442 /*
 443  * We regard a request as SYNC, if it's either a read or has the SYNC bit
 444  * set (in which case it could also be direct WRITE).
 445  */
 446 static inline bool cfq_bio_sync(struct bio *bio)
 447 {
 448         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
 449 }
 450
 451 /*
 452  * scheduler run of queue, if there are requests pending and no one in the
 453  * driver that will restart queueing
 454  */
 455 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
 456 {
 457         if (cfqd->busy_queues) {
 458                 cfq_log(cfqd, "schedule dispatch");
 459                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
 460         }
 461 }
 462
 463 static int cfq_queue_empty(struct request_queue *q)
 464 {
 465         struct cfq_data *cfqd = q->elevator->elevator_data;
 466
 467         return !cfqd->rq_queued;
 468 }
 469
 470 /*
 471  * Scale schedule slice based on io priority. Use the sync time slice only
 472  * if a queue is marked sync and has sync io queued. A sync queue with async
 473  * io only, should not get full sync slice length.
 474  */
 475 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
 476                                  unsigned short prio)
 477 {
 478         const int base_slice = cfqd->cfq_slice[sync];
 479
 480         WARN_ON(prio >= IOPRIO_BE_NR);
 481
 482         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
 483 }
 484
 485 static inline int
 486 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 487 {
 488         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
 489 }
 490
 491 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
 492 {
 493         u64 d = delta << CFQ_SERVICE_SHIFT;
 494
 495         d = d * BLKIO_WEIGHT_DEFAULT;
 496         do_div(d, cfqg->weight);
 497         return d;
 498 }
 499
 500 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
 501 {
 502         s64 delta = (s64)(vdisktime - min_vdisktime);
 503         if (delta > 0)
 504                 min_vdisktime = vdisktime;
 505
 506         return min_vdisktime;
 507 }
 508
 509 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
 510 {
 511         s64 delta = (s64)(vdisktime - min_vdisktime);
 512         if (delta < 0)
 513                 min_vdisktime = vdisktime;
 514
 515         return min_vdisktime;
 516 }
 517
 518 static void update_min_vdisktime(struct cfq_rb_root *st)
 519 {
 520         u64 vdisktime = st->min_vdisktime;
 521         struct cfq_group *cfqg;
 522
 523         if (st->active) {
 524                 cfqg = rb_entry_cfqg(st->active);
 525                 vdisktime = cfqg->vdisktime;
 526         }
 527
 528         if (st->left) {
 529                 cfqg = rb_entry_cfqg(st->left);
 530                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
 531         }
 532
 533         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
 534 }
 535
 536 /*
 537  * get averaged number of queues of RT/BE priority.
 538  * average is updated, with a formula that gives more weight to higher numbers,
 539  * to quickly follows sudden increases and decrease slowly
 540  */
 541
 542 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
 543                                         struct cfq_group *cfqg, bool rt)
 544 {
 545         unsigned min_q, max_q;
 546         unsigned mult  = cfq_hist_divisor - 1;
 547         unsigned round = cfq_hist_divisor / 2;
 548         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
 549
 550         min_q = min(cfqg->busy_queues_avg[rt], busy);
 551         max_q = max(cfqg->busy_queues_avg[rt], busy);
 552         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
 553                 cfq_hist_divisor;
 554         return cfqg->busy_queues_avg[rt];
 555 }
 556
 557 static inline unsigned
 558 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
 559 {
 560         struct cfq_rb_root *st = &cfqd->grp_service_tree;
 561
 562         return cfq_target_latency * cfqg->weight / st->total_weight;
 563 }
 564
 565 static inline void
 566 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 567 {
 568         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
 569         if (cfqd->cfq_latency) {
 570                 /*
 571                  * interested queues (we consider only the ones with the same
 572                  * priority class in the cfq group)
 573                  */
 574                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
 575                                                 cfq_class_rt(cfqq));
 576                 unsigned sync_slice = cfqd->cfq_slice[1];
 577                 unsigned expect_latency = sync_slice * iq;
 578                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
 579
 580                 if (expect_latency > group_slice) {
 581                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
 582                         /* scale low_slice according to IO priority
 583                          * and sync vs async */
 584                         unsigned low_slice =
 585                                 min(slice, base_low_slice * slice / sync_slice);
 586                         /* the adapted slice value is scaled to fit all iqs
 587                          * into the target latency */
 588                         slice = max(slice * group_slice / expect_latency,
 589                                     low_slice);
 590                 }
 591         }
 592         cfqq->slice_start = jiffies;
 593         cfqq->slice_end = jiffies + slice;
 594         cfqq->allocated_slice = slice;
 595         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 596 }
 597
 598 /*
 599  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
 600  * isn't valid until the first request from the dispatch is activated
 601  * and the slice time set.
 602  */
 603 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
 604 {
 605         if (cfq_cfqq_slice_new(cfqq))
 606                 return 0;
 607         if (time_before(jiffies, cfqq->slice_end))
 608                 return 0;
 609
 610         return 1;
 611 }
 612
 613 /*
 614  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
 615  * We choose the request that is closest to the head right now. Distance
 616  * behind the head is penalized and only allowed to a certain extent.
 617  */
 618 static struct request *
 619 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
 620 {
 621         sector_t s1, s2, d1 = 0, d2 = 0;
 622         unsigned long back_max;
 623 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
 624 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
 625         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
 626
 627         if (rq1 == NULL || rq1 == rq2)
 628                 return rq2;
 629         if (rq2 == NULL)
 630                 return rq1;
 631
 632         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
 633                 return rq1;
 634         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
 635                 return rq2;
 636         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
 637                 return rq1;
 638         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
 639                 return rq2;
 640
 641         s1 = blk_rq_pos(rq1);
 642         s2 = blk_rq_pos(rq2);
 643
 644         /*
 645          * by definition, 1KiB is 2 sectors
 646          */
 647         back_max = cfqd->cfq_back_max * 2;
 648
 649         /*
 650          * Strict one way elevator _except_ in the case where we allow
 651          * short backward seeks which are biased as twice the cost of a
 652          * similar forward seek.
 653          */
 654         if (s1 >= last)
 655                 d1 = s1 - last;
 656         else if (s1 + back_max >= last)
 657                 d1 = (last - s1) * cfqd->cfq_back_penalty;
 658         else
 659                 wrap |= CFQ_RQ1_WRAP;
 660
 661         if (s2 >= last)
 662                 d2 = s2 - last;
 663         else if (s2 + back_max >= last)
 664                 d2 = (last - s2) * cfqd->cfq_back_penalty;
 665         else
 666                 wrap |= CFQ_RQ2_WRAP;
 667
 668         /* Found required data */
 669
 670         /*
 671          * By doing switch() on the bit mask "wrap" we avoid having to
 672          * check two variables for all permutations: --> faster!
 673          */
 674         switch (wrap) {
 675         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
 676                 if (d1 < d2)
 677                         return rq1;
 678                 else if (d2 < d1)
 679                         return rq2;
 680                 else {
 681                         if (s1 >= s2)
 682                                 return rq1;
 683                         else
 684                                 return rq2;
 685                 }
 686
 687         case CFQ_RQ2_WRAP:
 688                 return rq1;
 689         case CFQ_RQ1_WRAP:
 690                 return rq2;
 691         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
 692         default:
 693                 /*
 694                  * Since both rqs are wrapped,
 695                  * start with the one that's further behind head
 696                  * (--> only *one* back seek required),
 697                  * since back seek takes more time than forward.
 698                  */
 699                 if (s1 <= s2)
 700                         return rq1;
 701                 else
 702                         return rq2;
 703         }
 704 }
 705
 706 /*
 707  * The below is leftmost cache rbtree addon
 708  */
 709 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
 710 {
 711         /* Service tree is empty */
 712         if (!root->count)
 713                 return NULL;
 714
 715         if (!root->left)
 716                 root->left = rb_first(&root->rb);
 717
 718         if (root->left)
 719                 return rb_entry(root->left, struct cfq_queue, rb_node);
 720
 721         return NULL;
 722 }
 723
 724 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
 725 {
 726         if (!root->left)
 727                 root->left = rb_first(&root->rb);
 728
 729         if (root->left)
 730                 return rb_entry_cfqg(root->left);
 731
 732         return NULL;
 733 }
 734
 735 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 736 {
 737         rb_erase(n, root);
 738         RB_CLEAR_NODE(n);
 739 }
 740
 741 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
 742 {
 743         if (root->left == n)
 744                 root->left = NULL;
 745         rb_erase_init(n, &root->rb);
 746         --root->count;
 747 }
 748
 749 /*
 750  * would be nice to take fifo expire time into account as well
 751  */
 752 static struct request *
 753 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
 754                   struct request *last)
 755 {
 756         struct rb_node *rbnext = rb_next(&last->rb_node);
 757         struct rb_node *rbprev = rb_prev(&last->rb_node);
 758         struct request *next = NULL, *prev = NULL;
 759
 760         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
 761
 762         if (rbprev)
 763                 prev = rb_entry_rq(rbprev);
 764
 765         if (rbnext)
 766                 next = rb_entry_rq(rbnext);
 767         else {
 768                 rbnext = rb_first(&cfqq->sort_list);
 769                 if (rbnext && rbnext != &last->rb_node)
 770                         next = rb_entry_rq(rbnext);
 771         }
 772
 773         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
 774 }
 775
 776 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
 777                                       struct cfq_queue *cfqq)
 778 {
 779         /*
 780          * just an approximation, should be ok.
 781          */
 782         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
 783                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
 784 }
 785
 786 static inline s64
 787 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
 788 {
 789         return cfqg->vdisktime - st->min_vdisktime;
 790 }
 791
 792 static void
 793 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
 794 {
 795         struct rb_node **node = &st->rb.rb_node;
 796         struct rb_node *parent = NULL;
 797         struct cfq_group *__cfqg;
 798         s64 key = cfqg_key(st, cfqg);
 799         int left = 1;
 800
 801         while (*node != NULL) {
 802                 parent = *node;
 803                 __cfqg = rb_entry_cfqg(parent);
 804
 805                 if (key < cfqg_key(st, __cfqg))
 806                         node = &parent->rb_left;
 807                 else {
 808                         node = &parent->rb_right;
 809                         left = 0;
 810                 }
 811         }
 812
 813         if (left)
 814                 st->left = &cfqg->rb_node;
 815
 816         rb_link_node(&cfqg->rb_node, parent, node);
 817         rb_insert_color(&cfqg->rb_node, &st->rb);
 818 }
 819
 820 static void
 821 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
 822 {
 823         struct cfq_rb_root *st = &cfqd->grp_service_tree;
 824         struct cfq_group *__cfqg;
 825         struct rb_node *n;
 826
 827         cfqg->nr_cfqq++;
 828         if (cfqg->on_st)
 829                 return;
 830
 831         /*
 832          * Currently put the group at the end. Later implement something
 833          * so that groups get lesser vtime based on their weights, so that
 834          * if group does not loose all if it was not continously backlogged.
 835          */
 836         n = rb_last(&st->rb);
 837         if (n) {
 838                 __cfqg = rb_entry_cfqg(n);
 839                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
 840         } else
 841                 cfqg->vdisktime = st->min_vdisktime;
 842
 843         __cfq_group_service_tree_add(st, cfqg);
 844         cfqg->on_st = true;
 845         cfqd->nr_groups++;
 846         st->total_weight += cfqg->weight;
 847 }
 848
 849 static void
 850 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
 851 {
 852         struct cfq_rb_root *st = &cfqd->grp_service_tree;
 853
 854         if (st->active == &cfqg->rb_node)
 855                 st->active = NULL;
 856
 857         BUG_ON(cfqg->nr_cfqq < 1);
 858         cfqg->nr_cfqq--;
 859
 860         /* If there are other cfq queues under this group, don't delete it */
 861         if (cfqg->nr_cfqq)
 862                 return;
 863
 864         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
 865         cfqg->on_st = false;
 866         cfqd->nr_groups--;
 867         st->total_weight -= cfqg->weight;
 868         if (!RB_EMPTY_NODE(&cfqg->rb_node))
 869                 cfq_rb_erase(&cfqg->rb_node, st);
 870         cfqg->saved_workload_slice = 0;
 871         blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
 872 }
 873
 874 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
 875 {
 876         unsigned int slice_used;
 877
 878         /*
 879          * Queue got expired before even a single request completed or
 880          * got expired immediately after first request completion.
 881          */
 882         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
 883                 /*
 884                  * Also charge the seek time incurred to the group, otherwise
 885                  * if there are mutiple queues in the group, each can dispatch
 886                  * a single request on seeky media and cause lots of seek time
 887                  * and group will never know it.
 888                  */
 889                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
 890                                         1);
 891         } else {
 892                 slice_used = jiffies - cfqq->slice_start;
 893                 if (slice_used > cfqq->allocated_slice)
 894                         slice_used = cfqq->allocated_slice;
 895         }
 896
 897         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
 898                                 cfqq->nr_sectors);
 899         return slice_used;
 900 }
 901
 902 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
 903                                 struct cfq_queue *cfqq)
 904 {
 905         struct cfq_rb_root *st = &cfqd->grp_service_tree;
 906         unsigned int used_sl, charge_sl;
 907         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
 908                         - cfqg->service_tree_idle.count;
 909
 910         BUG_ON(nr_sync < 0);
 911         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
 912
 913         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
 914                 charge_sl = cfqq->allocated_slice;
 915
 916         /* Can't update vdisktime while group is on service tree */
 917         cfq_rb_erase(&cfqg->rb_node, st);
 918         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
 919         __cfq_group_service_tree_add(st, cfqg);
 920
 921         /* This group is being expired. Save the context */
 922         if (time_after(cfqd->workload_expires, jiffies)) {
 923                 cfqg->saved_workload_slice = cfqd->workload_expires
 924                                                 - jiffies;
 925                 cfqg->saved_workload = cfqd->serving_type;
 926                 cfqg->saved_serving_prio = cfqd->serving_prio;
 927         } else
 928                 cfqg->saved_workload_slice = 0;
 929
 930         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
 931                                         st->min_vdisktime);
 932         blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
 933                                                 cfqq->nr_sectors);
 934 }
 935
 936 #ifdef CONFIG_CFQ_GROUP_IOSCHED
 937 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
 938 {
 939         if (blkg)
 940                 return container_of(blkg, struct cfq_group, blkg);
 941         return NULL;
 942 }
 943
 944 void
 945 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
 946 {
 947         cfqg_of_blkg(blkg)->weight = weight;
 948 }
 949
 950 static struct cfq_group *
 951 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
 952 {
 953         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
 954         struct cfq_group *cfqg = NULL;
 955         void *key = cfqd;
 956         int i, j;
 957         struct cfq_rb_root *st;
 958         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
 959         unsigned int major, minor;
 960
 961         /* Do we need to take this reference */
 962         if (!blkiocg_css_tryget(blkcg))
 963                 return NULL;;
 964
 965         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
 966         if (cfqg || !create)
 967                 goto done;
 968
 969         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
 970         if (!cfqg)
 971                 goto done;
 972
 973         cfqg->weight = blkcg->weight;
 974         for_each_cfqg_st(cfqg, i, j, st)
 975                 *st = CFQ_RB_ROOT;
 976         RB_CLEAR_NODE(&cfqg->rb_node);
 977
 978         /*
 979          * Take the initial reference that will be released on destroy
 980          * This can be thought of a joint reference by cgroup and
 981          * elevator which will be dropped by either elevator exit
 982          * or cgroup deletion path depending on who is exiting first.
 983          */
 984         atomic_set(&cfqg->ref, 1);
 985
 986         /* Add group onto cgroup list */
 987         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
 988         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
 989                                         MKDEV(major, minor));
 990
 991         /* Add group on cfqd list */
 992         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
 993
 994 done:
 995         blkiocg_css_put(blkcg);
 996         return cfqg;
 997 }
 998
 999 /*
1000  * Search for the cfq group current task belongs to. If create = 1, then also
1001  * create the cfq group if it does not exist. request_queue lock must be held.
1002  */
1003 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1004 {
1005         struct cgroup *cgroup;
1006         struct cfq_group *cfqg = NULL;
1007
1008         rcu_read_lock();
1009         cgroup = task_cgroup(current, blkio_subsys_id);
1010         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1011         if (!cfqg && create)
1012                 cfqg = &cfqd->root_group;
1013         rcu_read_unlock();
1014         return cfqg;
1015 }
1016
1017 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1018 {
1019         /* Currently, all async queues are mapped to root group */
1020         if (!cfq_cfqq_sync(cfqq))
1021                 cfqg = &cfqq->cfqd->root_group;
1022
1023         cfqq->cfqg = cfqg;
1024         /* cfqq reference on cfqg */
1025         atomic_inc(&cfqq->cfqg->ref);
1026 }
1027
1028 static void cfq_put_cfqg(struct cfq_group *cfqg)
1029 {
1030         struct cfq_rb_root *st;
1031         int i, j;
1032
1033         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1034         if (!atomic_dec_and_test(&cfqg->ref))
1035                 return;
1036         for_each_cfqg_st(cfqg, i, j, st)
1037                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1038         kfree(cfqg);
1039 }
1040
1041 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1042 {
1043         /* Something wrong if we are trying to remove same group twice */
1044         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1045
1046         hlist_del_init(&cfqg->cfqd_node);
1047
1048         /*
1049          * Put the reference taken at the time of creation so that when all
1050          * queues are gone, group can be destroyed.
1051          */
1052         cfq_put_cfqg(cfqg);
1053 }
1054
1055 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1056 {
1057         struct hlist_node *pos, *n;
1058         struct cfq_group *cfqg;
1059
1060         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1061                 /*
1062                  * If cgroup removal path got to blk_group first and removed
1063                  * it from cgroup list, then it will take care of destroying
1064                  * cfqg also.
1065                  */
1066                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1067                         cfq_destroy_cfqg(cfqd, cfqg);
1068         }
1069 }
1070
1071 /*
1072  * Blk cgroup controller notification saying that blkio_group object is being
1073  * delinked as associated cgroup object is going away. That also means that
1074  * no new IO will come in this group. So get rid of this group as soon as
1075  * any pending IO in the group is finished.
1076  *
1077  * This function is called under rcu_read_lock(). key is the rcu protected
1078  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1079  * read lock.
1080  *
1081  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1082  * it should not be NULL as even if elevator was exiting, cgroup deltion
1083  * path got to it first.
1084  */
1085 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1086 {
1087         unsigned long  flags;
1088         struct cfq_data *cfqd = key;
1089
1090         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1091         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1092         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1093 }
1094
1095 #else /* GROUP_IOSCHED */
1096 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1097 {
1098         return &cfqd->root_group;
1099 }
1100 static inline void
1101 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1102         cfqq->cfqg = cfqg;
1103 }
1104
1105 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1106 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1107
1108 #endif /* GROUP_IOSCHED */
1109
1110 /*
1111  * The cfqd->service_trees holds all pending cfq_queue's that have
1112  * requests waiting to be processed. It is sorted in the order that
1113  * we will service the queues.
1114  */
1115 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1116                                  bool add_front)
1117 {
1118         struct rb_node **p, *parent;
1119         struct cfq_queue *__cfqq;
1120         unsigned long rb_key;
1121         struct cfq_rb_root *service_tree;
1122         int left;
1123         int new_cfqq = 1;
1124         int group_changed = 0;
1125
1126 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1127         if (!cfqd->cfq_group_isolation
1128             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1129             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1130                 /* Move this cfq to root group */
1131                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1132                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1133                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1134                 cfqq->orig_cfqg = cfqq->cfqg;
1135                 cfqq->cfqg = &cfqd->root_group;
1136                 atomic_inc(&cfqd->root_group.ref);
1137                 group_changed = 1;
1138         } else if (!cfqd->cfq_group_isolation
1139                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1140                 /* cfqq is sequential now needs to go to its original group */
1141                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1142                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1143                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1144                 cfq_put_cfqg(cfqq->cfqg);
1145                 cfqq->cfqg = cfqq->orig_cfqg;
1146                 cfqq->orig_cfqg = NULL;
1147                 group_changed = 1;
1148                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1149         }
1150 #endif
1151
1152         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1153                                                 cfqq_type(cfqq), cfqd);
1154         if (cfq_class_idle(cfqq)) {
1155                 rb_key = CFQ_IDLE_DELAY;
1156                 parent = rb_last(&service_tree->rb);
1157                 if (parent && parent != &cfqq->rb_node) {
1158                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1159                         rb_key += __cfqq->rb_key;
1160                 } else
1161                         rb_key += jiffies;
1162         } else if (!add_front) {
1163                 /*
1164                  * Get our rb key offset. Subtract any residual slice
1165                  * value carried from last service. A negative resid
1166                  * count indicates slice overrun, and this should position
1167                  * the next service time further away in the tree.
1168                  */
1169                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1170                 rb_key -= cfqq->slice_resid;
1171                 cfqq->slice_resid = 0;
1172         } else {
1173                 rb_key = -HZ;
1174                 __cfqq = cfq_rb_first(service_tree);
1175                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1176         }
1177
1178         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1179                 new_cfqq = 0;
1180                 /*
1181                  * same position, nothing more to do
1182                  */
1183                 if (rb_key == cfqq->rb_key &&
1184                     cfqq->service_tree == service_tree)
1185                         return;
1186
1187                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1188                 cfqq->service_tree = NULL;
1189         }
1190
1191         left = 1;
1192         parent = NULL;
1193         cfqq->service_tree = service_tree;
1194         p = &service_tree->rb.rb_node;
1195         while (*p) {
1196                 struct rb_node **n;
1197
1198                 parent = *p;
1199                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1200
1201                 /*
1202                  * sort by key, that represents service time.
1203                  */
1204                 if (time_before(rb_key, __cfqq->rb_key))
1205                         n = &(*p)->rb_left;
1206                 else {
1207                         n = &(*p)->rb_right;
1208                         left = 0;
1209                 }
1210
1211                 p = n;
1212         }
1213
1214         if (left)
1215                 service_tree->left = &cfqq->rb_node;
1216
1217         cfqq->rb_key = rb_key;
1218         rb_link_node(&cfqq->rb_node, parent, p);
1219         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1220         service_tree->count++;
1221         if ((add_front || !new_cfqq) && !group_changed)
1222                 return;
1223         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1224 }
1225
1226 static struct cfq_queue *
1227 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1228                      sector_t sector, struct rb_node **ret_parent,
1229                      struct rb_node ***rb_link)
1230 {
1231         struct rb_node **p, *parent;
1232         struct cfq_queue *cfqq = NULL;
1233
1234         parent = NULL;
1235         p = &root->rb_node;
1236         while (*p) {
1237                 struct rb_node **n;
1238
1239                 parent = *p;
1240                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1241
1242                 /*
1243                  * Sort strictly based on sector.  Smallest to the left,
1244                  * largest to the right.
1245                  */
1246                 if (sector > blk_rq_pos(cfqq->next_rq))
1247                         n = &(*p)->rb_right;
1248                 else if (sector < blk_rq_pos(cfqq->next_rq))
1249                         n = &(*p)->rb_left;
1250                 else
1251                         break;
1252                 p = n;
1253                 cfqq = NULL;
1254         }
1255
1256         *ret_parent = parent;
1257         if (rb_link)
1258                 *rb_link = p;
1259         return cfqq;
1260 }
1261
1262 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1263 {
1264         struct rb_node **p, *parent;
1265         struct cfq_queue *__cfqq;
1266
1267         if (cfqq->p_root) {
1268                 rb_erase(&cfqq->p_node, cfqq->p_root);
1269                 cfqq->p_root = NULL;
1270         }
1271
1272         if (cfq_class_idle(cfqq))
1273                 return;
1274         if (!cfqq->next_rq)
1275                 return;
1276
1277         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1278         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1279                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1280         if (!__cfqq) {
1281                 rb_link_node(&cfqq->p_node, parent, p);
1282                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1283         } else
1284                 cfqq->p_root = NULL;
1285 }
1286
1287 /*
1288  * Update cfqq's position in the service tree.
1289  */
1290 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1291 {
1292         /*
1293          * Resorting requires the cfqq to be on the RR list already.
1294          */
1295         if (cfq_cfqq_on_rr(cfqq)) {
1296                 cfq_service_tree_add(cfqd, cfqq, 0);
1297                 cfq_prio_tree_add(cfqd, cfqq);
1298         }
1299 }
1300
1301 /*
1302  * add to busy list of queues for service, trying to be fair in ordering
1303  * the pending list according to last request service
1304  */
1305 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1306 {
1307         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1308         BUG_ON(cfq_cfqq_on_rr(cfqq));
1309         cfq_mark_cfqq_on_rr(cfqq);
1310         cfqd->busy_queues++;
1311
1312         cfq_resort_rr_list(cfqd, cfqq);
1313 }
1314
1315 /*
1316  * Called when the cfqq no longer has requests pending, remove it from
1317  * the service tree.
1318  */
1319 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1320 {
1321         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1322         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1323         cfq_clear_cfqq_on_rr(cfqq);
1324
1325         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1326                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1327                 cfqq->service_tree = NULL;
1328         }
1329         if (cfqq->p_root) {
1330                 rb_erase(&cfqq->p_node, cfqq->p_root);
1331                 cfqq->p_root = NULL;
1332         }
1333
1334         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1335         BUG_ON(!cfqd->busy_queues);
1336         cfqd->busy_queues--;
1337 }
1338
1339 /*
1340  * rb tree support functions
1341  */
1342 static void cfq_del_rq_rb(struct request *rq)
1343 {
1344         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1345         const int sync = rq_is_sync(rq);
1346
1347         BUG_ON(!cfqq->queued[sync]);
1348         cfqq->queued[sync]--;
1349
1350         elv_rb_del(&cfqq->sort_list, rq);
1351
1352         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1353                 /*
1354                  * Queue will be deleted from service tree when we actually
1355                  * expire it later. Right now just remove it from prio tree
1356                  * as it is empty.
1357                  */
1358                 if (cfqq->p_root) {
1359                         rb_erase(&cfqq->p_node, cfqq->p_root);
1360                         cfqq->p_root = NULL;
1361                 }
1362         }
1363 }
1364
1365 static void cfq_add_rq_rb(struct request *rq)
1366 {
1367         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1368         struct cfq_data *cfqd = cfqq->cfqd;
1369         struct request *__alias, *prev;
1370
1371         cfqq->queued[rq_is_sync(rq)]++;
1372
1373         /*
1374          * looks a little odd, but the first insert might return an alias.
1375          * if that happens, put the alias on the dispatch list
1376          */
1377         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1378                 cfq_dispatch_insert(cfqd->queue, __alias);
1379
1380         if (!cfq_cfqq_on_rr(cfqq))
1381                 cfq_add_cfqq_rr(cfqd, cfqq);
1382
1383         /*
1384          * check if this request is a better next-serve candidate
1385          */
1386         prev = cfqq->next_rq;
1387         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1388
1389         /*
1390          * adjust priority tree position, if ->next_rq changes
1391          */
1392         if (prev != cfqq->next_rq)
1393                 cfq_prio_tree_add(cfqd, cfqq);
1394
1395         BUG_ON(!cfqq->next_rq);
1396 }
1397
1398 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1399 {
1400         elv_rb_del(&cfqq->sort_list, rq);
1401         cfqq->queued[rq_is_sync(rq)]--;
1402         cfq_add_rq_rb(rq);
1403 }
1404
1405 static struct request *
1406 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1407 {
1408         struct task_struct *tsk = current;
1409         struct cfq_io_context *cic;
1410         struct cfq_queue *cfqq;
1411
1412         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1413         if (!cic)
1414                 return NULL;
1415
1416         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1417         if (cfqq) {
1418                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1419
1420                 return elv_rb_find(&cfqq->sort_list, sector);
1421         }
1422
1423         return NULL;
1424 }
1425
1426 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1427 {
1428         struct cfq_data *cfqd = q->elevator->elevator_data;
1429
1430         cfqd->rq_in_driver[rq_is_sync(rq)]++;
1431         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1432                                                 rq_in_driver(cfqd));
1433
1434         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1435 }
1436
1437 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1438 {
1439         struct cfq_data *cfqd = q->elevator->elevator_data;
1440         const int sync = rq_is_sync(rq);
1441
1442         WARN_ON(!cfqd->rq_in_driver[sync]);
1443         cfqd->rq_in_driver[sync]--;
1444         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1445                                                 rq_in_driver(cfqd));
1446 }
1447
1448 static void cfq_remove_request(struct request *rq)
1449 {
1450         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1451
1452         if (cfqq->next_rq == rq)
1453                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1454
1455         list_del_init(&rq->queuelist);
1456         cfq_del_rq_rb(rq);
1457
1458         cfqq->cfqd->rq_queued--;
1459         if (rq_is_meta(rq)) {
1460                 WARN_ON(!cfqq->meta_pending);
1461                 cfqq->meta_pending--;
1462         }
1463 }
1464
1465 static int cfq_merge(struct request_queue *q, struct request **req,
1466                      struct bio *bio)
1467 {
1468         struct cfq_data *cfqd = q->elevator->elevator_data;
1469         struct request *__rq;
1470
1471         __rq = cfq_find_rq_fmerge(cfqd, bio);
1472         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1473                 *req = __rq;
1474                 return ELEVATOR_FRONT_MERGE;
1475         }
1476
1477         return ELEVATOR_NO_MERGE;
1478 }
1479
1480 static void cfq_merged_request(struct request_queue *q, struct request *req,
1481                                int type)
1482 {
1483         if (type == ELEVATOR_FRONT_MERGE) {
1484                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1485
1486                 cfq_reposition_rq_rb(cfqq, req);
1487         }
1488 }
1489
1490 static void
1491 cfq_merged_requests(struct request_queue *q, struct request *rq,
1492                     struct request *next)
1493 {
1494         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1495         /*
1496          * reposition in fifo if next is older than rq
1497          */
1498         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1499             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1500                 list_move(&rq->queuelist, &next->queuelist);
1501                 rq_set_fifo_time(rq, rq_fifo_time(next));
1502         }
1503
1504         if (cfqq->next_rq == next)
1505                 cfqq->next_rq = rq;
1506         cfq_remove_request(next);
1507 }
1508
1509 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1510                            struct bio *bio)
1511 {
1512         struct cfq_data *cfqd = q->elevator->elevator_data;
1513         struct cfq_io_context *cic;
1514         struct cfq_queue *cfqq;
1515
1516         /* Deny merge if bio and rq don't belong to same cfq group */
1517         if ((RQ_CFQQ(rq))->cfqg != cfq_get_cfqg(cfqd, 0))
1518                 return false;
1519         /*
1520          * Disallow merge of a sync bio into an async request.
1521          */
1522         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1523                 return false;
1524
1525         /*
1526          * Lookup the cfqq that this bio will be queued with. Allow
1527          * merge only if rq is queued there.
1528          */
1529         cic = cfq_cic_lookup(cfqd, current->io_context);
1530         if (!cic)
1531                 return false;
1532
1533         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1534         return cfqq == RQ_CFQQ(rq);
1535 }
1536
1537 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1538                                    struct cfq_queue *cfqq)
1539 {
1540         if (cfqq) {
1541                 cfq_log_cfqq(cfqd, cfqq, "set_active");
1542                 cfqq->slice_start = 0;
1543                 cfqq->dispatch_start = jiffies;
1544                 cfqq->allocated_slice = 0;
1545                 cfqq->slice_end = 0;
1546                 cfqq->slice_dispatch = 0;
1547                 cfqq->nr_sectors = 0;
1548
1549                 cfq_clear_cfqq_wait_request(cfqq);
1550                 cfq_clear_cfqq_must_dispatch(cfqq);
1551                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1552                 cfq_clear_cfqq_fifo_expire(cfqq);
1553                 cfq_mark_cfqq_slice_new(cfqq);
1554
1555                 del_timer(&cfqd->idle_slice_timer);
1556         }
1557
1558         cfqd->active_queue = cfqq;
1559 }
1560
1561 /*
1562  * current cfqq expired its slice (or was too idle), select new one
1563  */
1564 static void
1565 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1566                     bool timed_out)
1567 {
1568         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1569
1570         if (cfq_cfqq_wait_request(cfqq))
1571                 del_timer(&cfqd->idle_slice_timer);
1572
1573         cfq_clear_cfqq_wait_request(cfqq);
1574         cfq_clear_cfqq_wait_busy(cfqq);
1575
1576         /*
1577          * store what was left of this slice, if the queue idled/timed out
1578          */
1579         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1580                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1581                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1582         }
1583
1584         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1585
1586         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1587                 cfq_del_cfqq_rr(cfqd, cfqq);
1588
1589         cfq_resort_rr_list(cfqd, cfqq);
1590
1591         if (cfqq == cfqd->active_queue)
1592                 cfqd->active_queue = NULL;
1593
1594         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1595                 cfqd->grp_service_tree.active = NULL;
1596
1597         if (cfqd->active_cic) {
1598                 put_io_context(cfqd->active_cic->ioc);
1599                 cfqd->active_cic = NULL;
1600         }
1601 }
1602
1603 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1604 {
1605         struct cfq_queue *cfqq = cfqd->active_queue;
1606
1607         if (cfqq)
1608                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1609 }
1610
1611 /*
1612  * Get next queue for service. Unless we have a queue preemption,
1613  * we'll simply select the first cfqq in the service tree.
1614  */
1615 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1616 {
1617         struct cfq_rb_root *service_tree =
1618                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1619                                         cfqd->serving_type, cfqd);
1620
1621         if (!cfqd->rq_queued)
1622                 return NULL;
1623
1624         /* There is nothing to dispatch */
1625         if (!service_tree)
1626                 return NULL;
1627         if (RB_EMPTY_ROOT(&service_tree->rb))
1628                 return NULL;
1629         return cfq_rb_first(service_tree);
1630 }
1631
1632 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1633 {
1634         struct cfq_group *cfqg;
1635         struct cfq_queue *cfqq;
1636         int i, j;
1637         struct cfq_rb_root *st;
1638
1639         if (!cfqd->rq_queued)
1640                 return NULL;
1641
1642         cfqg = cfq_get_next_cfqg(cfqd);
1643         if (!cfqg)
1644                 return NULL;
1645
1646         for_each_cfqg_st(cfqg, i, j, st)
1647                 if ((cfqq = cfq_rb_first(st)) != NULL)
1648                         return cfqq;
1649         return NULL;
1650 }
1651
1652 /*
1653  * Get and set a new active queue for service.
1654  */
1655 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1656                                               struct cfq_queue *cfqq)
1657 {
1658         if (!cfqq)
1659                 cfqq = cfq_get_next_queue(cfqd);
1660
1661         __cfq_set_active_queue(cfqd, cfqq);
1662         return cfqq;
1663 }
1664
1665 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1666                                           struct request *rq)
1667 {
1668         if (blk_rq_pos(rq) >= cfqd->last_position)
1669                 return blk_rq_pos(rq) - cfqd->last_position;
1670         else
1671                 return cfqd->last_position - blk_rq_pos(rq);
1672 }
1673
1674 #define CFQQ_SEEK_THR           8 * 1024
1675 #define CFQQ_SEEKY(cfqq)        ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1676
1677 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1678                                struct request *rq)
1679 {
1680         sector_t sdist = cfqq->seek_mean;
1681
1682         if (!sample_valid(cfqq->seek_samples))
1683                 sdist = CFQQ_SEEK_THR;
1684
1685         return cfq_dist_from_last(cfqd, rq) <= sdist;
1686 }
1687
1688 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1689                                     struct cfq_queue *cur_cfqq)
1690 {
1691         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1692         struct rb_node *parent, *node;
1693         struct cfq_queue *__cfqq;
1694         sector_t sector = cfqd->last_position;
1695
1696         if (RB_EMPTY_ROOT(root))
1697                 return NULL;
1698
1699         /*
1700          * First, if we find a request starting at the end of the last
1701          * request, choose it.
1702          */
1703         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1704         if (__cfqq)
1705                 return __cfqq;
1706
1707         /*
1708          * If the exact sector wasn't found, the parent of the NULL leaf
1709          * will contain the closest sector.
1710          */
1711         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1712         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1713                 return __cfqq;
1714
1715         if (blk_rq_pos(__cfqq->next_rq) < sector)
1716                 node = rb_next(&__cfqq->p_node);
1717         else
1718                 node = rb_prev(&__cfqq->p_node);
1719         if (!node)
1720                 return NULL;
1721
1722         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1723         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1724                 return __cfqq;
1725
1726         return NULL;
1727 }
1728
1729 /*
1730  * cfqd - obvious
1731  * cur_cfqq - passed in so that we don't decide that the current queue is
1732  *            closely cooperating with itself.
1733  *
1734  * So, basically we're assuming that that cur_cfqq has dispatched at least
1735  * one request, and that cfqd->last_position reflects a position on the disk
1736  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1737  * assumption.
1738  */
1739 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1740                                               struct cfq_queue *cur_cfqq)
1741 {
1742         struct cfq_queue *cfqq;
1743
1744         if (!cfq_cfqq_sync(cur_cfqq))
1745                 return NULL;
1746         if (CFQQ_SEEKY(cur_cfqq))
1747                 return NULL;
1748
1749         /*
1750          * Don't search priority tree if it's the only queue in the group.
1751          */
1752         if (cur_cfqq->cfqg->nr_cfqq == 1)
1753                 return NULL;
1754
1755         /*
1756          * We should notice if some of the queues are cooperating, eg
1757          * working closely on the same area of the disk. In that case,
1758          * we can group them together and don't waste time idling.
1759          */
1760         cfqq = cfqq_close(cfqd, cur_cfqq);
1761         if (!cfqq)
1762                 return NULL;
1763
1764         /* If new queue belongs to different cfq_group, don't choose it */
1765         if (cur_cfqq->cfqg != cfqq->cfqg)
1766                 return NULL;
1767
1768         /*
1769          * It only makes sense to merge sync queues.
1770          */
1771         if (!cfq_cfqq_sync(cfqq))
1772                 return NULL;
1773         if (CFQQ_SEEKY(cfqq))
1774                 return NULL;
1775
1776         /*
1777          * Do not merge queues of different priority classes
1778          */
1779         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1780                 return NULL;
1781
1782         return cfqq;
1783 }
1784
1785 /*
1786  * Determine whether we should enforce idle window for this queue.
1787  */
1788
1789 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1790 {
1791         enum wl_prio_t prio = cfqq_prio(cfqq);
1792         struct cfq_rb_root *service_tree = cfqq->service_tree;
1793
1794         BUG_ON(!service_tree);
1795         BUG_ON(!service_tree->count);
1796
1797         /* We never do for idle class queues. */
1798         if (prio == IDLE_WORKLOAD)
1799                 return false;
1800
1801         /* We do for queues that were marked with idle window flag. */
1802         if (cfq_cfqq_idle_window(cfqq) &&
1803            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1804                 return true;
1805
1806         /*
1807          * Otherwise, we do only if they are the last ones
1808          * in their service tree.
1809          */
1810         return service_tree->count == 1;
1811 }
1812
1813 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1814 {
1815         struct cfq_queue *cfqq = cfqd->active_queue;
1816         struct cfq_io_context *cic;
1817         unsigned long sl;
1818
1819         /*
1820          * SSD device without seek penalty, disable idling. But only do so
1821          * for devices that support queuing, otherwise we still have a problem
1822          * with sync vs async workloads.
1823          */
1824         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1825                 return;
1826
1827         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1828         WARN_ON(cfq_cfqq_slice_new(cfqq));
1829
1830         /*
1831          * idle is disabled, either manually or by past process history
1832          */
1833         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1834                 return;
1835
1836         /*
1837          * still active requests from this queue, don't idle
1838          */
1839         if (cfqq->dispatched)
1840                 return;
1841
1842         /*
1843          * task has exited, don't wait
1844          */
1845         cic = cfqd->active_cic;
1846         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1847                 return;
1848
1849         /*
1850          * If our average think time is larger than the remaining time
1851          * slice, then don't idle. This avoids overrunning the allotted
1852          * time slice.
1853          */
1854         if (sample_valid(cic->ttime_samples) &&
1855             (cfqq->slice_end - jiffies < cic->ttime_mean))
1856                 return;
1857
1858         cfq_mark_cfqq_wait_request(cfqq);
1859
1860         sl = cfqd->cfq_slice_idle;
1861
1862         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1863         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1864 }
1865
1866 /*
1867  * Move request from internal lists to the request queue dispatch list.
1868  */
1869 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1870 {
1871         struct cfq_data *cfqd = q->elevator->elevator_data;
1872         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1873
1874         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1875
1876         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1877         cfq_remove_request(rq);
1878         cfqq->dispatched++;
1879         elv_dispatch_sort(q, rq);
1880
1881         if (cfq_cfqq_sync(cfqq))
1882                 cfqd->sync_flight++;
1883         cfqq->nr_sectors += blk_rq_sectors(rq);
1884 }
1885
1886 /*
1887  * return expired entry, or NULL to just start from scratch in rbtree
1888  */
1889 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1890 {
1891         struct request *rq = NULL;
1892
1893         if (cfq_cfqq_fifo_expire(cfqq))
1894                 return NULL;
1895
1896         cfq_mark_cfqq_fifo_expire(cfqq);
1897
1898         if (list_empty(&cfqq->fifo))
1899                 return NULL;
1900
1901         rq = rq_entry_fifo(cfqq->fifo.next);
1902         if (time_before(jiffies, rq_fifo_time(rq)))
1903                 rq = NULL;
1904
1905         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1906         return rq;
1907 }
1908
1909 static inline int
1910 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1911 {
1912         const int base_rq = cfqd->cfq_slice_async_rq;
1913
1914         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1915
1916         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1917 }
1918
1919 /*
1920  * Must be called with the queue_lock held.
1921  */
1922 static int cfqq_process_refs(struct cfq_queue *cfqq)
1923 {
1924         int process_refs, io_refs;
1925
1926         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1927         process_refs = atomic_read(&cfqq->ref) - io_refs;
1928         BUG_ON(process_refs < 0);
1929         return process_refs;
1930 }
1931
1932 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1933 {
1934         int process_refs, new_process_refs;
1935         struct cfq_queue *__cfqq;
1936
1937         /* Avoid a circular list and skip interim queue merges */
1938         while ((__cfqq = new_cfqq->new_cfqq)) {
1939                 if (__cfqq == cfqq)
1940                         return;
1941                 new_cfqq = __cfqq;
1942         }
1943
1944         process_refs = cfqq_process_refs(cfqq);
1945         /*
1946          * If the process for the cfqq has gone away, there is no
1947          * sense in merging the queues.
1948          */
1949         if (process_refs == 0)
1950                 return;
1951
1952         /*
1953          * Merge in the direction of the lesser amount of work.
1954          */
1955         new_process_refs = cfqq_process_refs(new_cfqq);
1956         if (new_process_refs >= process_refs) {
1957                 cfqq->new_cfqq = new_cfqq;
1958                 atomic_add(process_refs, &new_cfqq->ref);
1959         } else {
1960                 new_cfqq->new_cfqq = cfqq;
1961                 atomic_add(new_process_refs, &cfqq->ref);
1962         }
1963 }
1964
1965 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1966                                 struct cfq_group *cfqg, enum wl_prio_t prio,
1967                                 bool prio_changed)
1968 {
1969         struct cfq_queue *queue;
1970         int i;
1971         bool key_valid = false;
1972         unsigned long lowest_key = 0;
1973         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1974
1975         if (prio_changed) {
1976                 /*
1977                  * When priorities switched, we prefer starting
1978                  * from SYNC_NOIDLE (first choice), or just SYNC
1979                  * over ASYNC
1980                  */
1981                 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1982                         return cur_best;
1983                 cur_best = SYNC_WORKLOAD;
1984                 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1985                         return cur_best;
1986
1987                 return ASYNC_WORKLOAD;
1988         }
1989
1990         for (i = 0; i < 3; ++i) {
1991                 /* otherwise, select the one with lowest rb_key */
1992                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1993                 if (queue &&
1994                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1995                         lowest_key = queue->rb_key;
1996                         cur_best = i;
1997                         key_valid = true;
1998                 }
1999         }
2000
2001         return cur_best;
2002 }
2003
2004 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2005 {
2006         enum wl_prio_t previous_prio = cfqd->serving_prio;
2007         bool prio_changed;
2008         unsigned slice;
2009         unsigned count;
2010         struct cfq_rb_root *st;
2011         unsigned group_slice;
2012
2013         if (!cfqg) {
2014                 cfqd->serving_prio = IDLE_WORKLOAD;
2015                 cfqd->workload_expires = jiffies + 1;
2016                 return;
2017         }
2018
2019         /* Choose next priority. RT > BE > IDLE */
2020         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2021                 cfqd->serving_prio = RT_WORKLOAD;
2022         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2023                 cfqd->serving_prio = BE_WORKLOAD;
2024         else {
2025                 cfqd->serving_prio = IDLE_WORKLOAD;
2026                 cfqd->workload_expires = jiffies + 1;
2027                 return;
2028         }
2029
2030         /*
2031          * For RT and BE, we have to choose also the type
2032          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2033          * expiration time
2034          */
2035         prio_changed = (cfqd->serving_prio != previous_prio);
2036         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2037                                 cfqd);
2038         count = st->count;
2039
2040         /*
2041          * If priority didn't change, check workload expiration,
2042          * and that we still have other queues ready
2043          */
2044         if (!prio_changed && count &&
2045             !time_after(jiffies, cfqd->workload_expires))
2046                 return;
2047
2048         /* otherwise select new workload type */
2049         cfqd->serving_type =
2050                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
2051         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
2052                                 cfqd);
2053         count = st->count;
2054
2055         /*
2056          * the workload slice is computed as a fraction of target latency
2057          * proportional to the number of queues in that workload, over
2058          * all the queues in the same priority class
2059          */
2060         group_slice = cfq_group_slice(cfqd, cfqg);
2061
2062         slice = group_slice * count /
2063                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2064                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2065
2066         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2067                 unsigned int tmp;
2068
2069                 /*
2070                  * Async queues are currently system wide. Just taking
2071                  * proportion of queues with-in same group will lead to higher
2072                  * async ratio system wide as generally root group is going
2073                  * to have higher weight. A more accurate thing would be to
2074                  * calculate system wide asnc/sync ratio.
2075                  */
2076                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2077                 tmp = tmp/cfqd->busy_queues;
2078                 slice = min_t(unsigned, slice, tmp);
2079
2080                 /* async workload slice is scaled down according to
2081                  * the sync/async slice ratio. */
2082                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2083         } else
2084                 /* sync workload slice is at least 2 * cfq_slice_idle */
2085                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2086
2087         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2088         cfqd->workload_expires = jiffies + slice;
2089         cfqd->noidle_tree_requires_idle = false;
2090 }
2091
2092 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2093 {
2094         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2095         struct cfq_group *cfqg;
2096
2097         if (RB_EMPTY_ROOT(&st->rb))
2098                 return NULL;
2099         cfqg = cfq_rb_first_group(st);
2100         st->active = &cfqg->rb_node;
2101         update_min_vdisktime(st);
2102         return cfqg;
2103 }
2104
2105 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2106 {
2107         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2108
2109         cfqd->serving_group = cfqg;
2110
2111         /* Restore the workload type data */
2112         if (cfqg->saved_workload_slice) {
2113                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2114                 cfqd->serving_type = cfqg->saved_workload;
2115                 cfqd->serving_prio = cfqg->saved_serving_prio;
2116         } else
2117                 cfqd->workload_expires = jiffies - 1;
2118
2119         choose_service_tree(cfqd, cfqg);
2120 }
2121
2122 /*
2123  * Select a queue for service. If we have a current active queue,
2124  * check whether to continue servicing it, or retrieve and set a new one.
2125  */
2126 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2127 {
2128         struct cfq_queue *cfqq, *new_cfqq = NULL;
2129
2130         cfqq = cfqd->active_queue;
2131         if (!cfqq)
2132                 goto new_queue;
2133
2134         if (!cfqd->rq_queued)
2135                 return NULL;
2136
2137         /*
2138          * We were waiting for group to get backlogged. Expire the queue
2139          */
2140         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2141                 goto expire;
2142
2143         /*
2144          * The active queue has run out of time, expire it and select new.
2145          */
2146         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2147                 /*
2148                  * If slice had not expired at the completion of last request
2149                  * we might not have turned on wait_busy flag. Don't expire
2150                  * the queue yet. Allow the group to get backlogged.
2151                  *
2152                  * The very fact that we have used the slice, that means we
2153                  * have been idling all along on this queue and it should be
2154                  * ok to wait for this request to complete.
2155                  */
2156                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2157                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2158                         cfqq = NULL;
2159                         goto keep_queue;
2160                 } else
2161                         goto expire;
2162         }
2163
2164         /*
2165          * The active queue has requests and isn't expired, allow it to
2166          * dispatch.
2167          */
2168         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2169                 goto keep_queue;
2170
2171         /*
2172          * If another queue has a request waiting within our mean seek
2173          * distance, let it run.  The expire code will check for close
2174          * cooperators and put the close queue at the front of the service
2175          * tree.  If possible, merge the expiring queue with the new cfqq.
2176          */
2177         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2178         if (new_cfqq) {
2179                 if (!cfqq->new_cfqq)
2180                         cfq_setup_merge(cfqq, new_cfqq);
2181                 goto expire;
2182         }
2183
2184         /*
2185          * No requests pending. If the active queue still has requests in
2186          * flight or is idling for a new request, allow either of these
2187          * conditions to happen (or time out) before selecting a new queue.
2188          */
2189         if (timer_pending(&cfqd->idle_slice_timer) ||
2190             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2191                 cfqq = NULL;
2192                 goto keep_queue;
2193         }
2194
2195 expire:
2196         cfq_slice_expired(cfqd, 0);
2197 new_queue:
2198         /*
2199          * Current queue expired. Check if we have to switch to a new
2200          * service tree
2201          */
2202         if (!new_cfqq)
2203                 cfq_choose_cfqg(cfqd);
2204
2205         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2206 keep_queue:
2207         return cfqq;
2208 }
2209
2210 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2211 {
2212         int dispatched = 0;
2213
2214         while (cfqq->next_rq) {
2215                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2216                 dispatched++;
2217         }
2218
2219         BUG_ON(!list_empty(&cfqq->fifo));
2220
2221         /* By default cfqq is not expired if it is empty. Do it explicitly */
2222         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2223         return dispatched;
2224 }
2225
2226 /*
2227  * Drain our current requests. Used for barriers and when switching
2228  * io schedulers on-the-fly.
2229  */
2230 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2231 {
2232         struct cfq_queue *cfqq;
2233         int dispatched = 0;
2234
2235         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2236                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2237
2238         cfq_slice_expired(cfqd, 0);
2239         BUG_ON(cfqd->busy_queues);
2240
2241         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2242         return dispatched;
2243 }
2244
2245 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2246 {
2247         unsigned int max_dispatch;
2248
2249         /*
2250          * Drain async requests before we start sync IO
2251          */
2252         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2253                 return false;
2254
2255         /*
2256          * If this is an async queue and we have sync IO in flight, let it wait
2257          */
2258         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2259                 return false;
2260
2261         max_dispatch = cfqd->cfq_quantum;
2262         if (cfq_class_idle(cfqq))
2263                 max_dispatch = 1;
2264
2265         /*
2266          * Does this cfqq already have too much IO in flight?
2267          */
2268         if (cfqq->dispatched >= max_dispatch) {
2269                 /*
2270                  * idle queue must always only have a single IO in flight
2271                  */
2272                 if (cfq_class_idle(cfqq))
2273                         return false;
2274
2275                 /*
2276                  * We have other queues, don't allow more IO from this one
2277                  */
2278                 if (cfqd->busy_queues > 1)
2279                         return false;
2280
2281                 /*
2282                  * Sole queue user, no limit
2283                  */
2284                 max_dispatch = -1;
2285         }
2286
2287         /*
2288          * Async queues must wait a bit before being allowed dispatch.
2289          * We also ramp up the dispatch depth gradually for async IO,
2290          * based on the last sync IO we serviced
2291          */
2292         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2293                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2294                 unsigned int depth;
2295
2296                 depth = last_sync / cfqd->cfq_slice[1];
2297                 if (!depth && !cfqq->dispatched)
2298                         depth = 1;
2299                 if (depth < max_dispatch)
2300                         max_dispatch = depth;
2301         }
2302
2303         /*
2304          * If we're below the current max, allow a dispatch
2305          */
2306         return cfqq->dispatched < max_dispatch;
2307 }
2308
2309 /*
2310  * Dispatch a request from cfqq, moving them to the request queue
2311  * dispatch list.
2312  */
2313 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2314 {
2315         struct request *rq;
2316
2317         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2318
2319         if (!cfq_may_dispatch(cfqd, cfqq))
2320                 return false;
2321
2322         /*
2323          * follow expired path, else get first next available
2324          */
2325         rq = cfq_check_fifo(cfqq);
2326         if (!rq)
2327                 rq = cfqq->next_rq;
2328
2329         /*
2330          * insert request into driver dispatch list
2331          */
2332         cfq_dispatch_insert(cfqd->queue, rq);
2333
2334         if (!cfqd->active_cic) {
2335                 struct cfq_io_context *cic = RQ_CIC(rq);
2336
2337                 atomic_long_inc(&cic->ioc->refcount);
2338                 cfqd->active_cic = cic;
2339         }
2340
2341         return true;
2342 }
2343
2344 /*
2345  * Find the cfqq that we need to service and move a request from that to the
2346  * dispatch list
2347  */
2348 static int cfq_dispatch_requests(struct request_queue *q, int force)
2349 {
2350         struct cfq_data *cfqd = q->elevator->elevator_data;
2351         struct cfq_queue *cfqq;
2352
2353         if (!cfqd->busy_queues)
2354                 return 0;
2355
2356         if (unlikely(force))
2357                 return cfq_forced_dispatch(cfqd);
2358
2359         cfqq = cfq_select_queue(cfqd);
2360         if (!cfqq)
2361                 return 0;
2362
2363         /*
2364          * Dispatch a request from this cfqq, if it is allowed
2365          */
2366         if (!cfq_dispatch_request(cfqd, cfqq))
2367                 return 0;
2368
2369         cfqq->slice_dispatch++;
2370         cfq_clear_cfqq_must_dispatch(cfqq);
2371
2372         /*
2373          * expire an async queue immediately if it has used up its slice. idle
2374          * queue always expire after 1 dispatch round.
2375          */
2376         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2377             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2378             cfq_class_idle(cfqq))) {
2379                 cfqq->slice_end = jiffies + 1;
2380                 cfq_slice_expired(cfqd, 0);
2381         }
2382
2383         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2384         return 1;
2385 }
2386
2387 /*
2388  * task holds one reference to the queue, dropped when task exits. each rq
2389  * in-flight on this queue also holds a reference, dropped when rq is freed.
2390  *
2391  * Each cfq queue took a reference on the parent group. Drop it now.
2392  * queue lock must be held here.
2393  */
2394 static void cfq_put_queue(struct cfq_queue *cfqq)
2395 {
2396         struct cfq_data *cfqd = cfqq->cfqd;
2397         struct cfq_group *cfqg, *orig_cfqg;
2398
2399         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2400
2401         if (!atomic_dec_and_test(&cfqq->ref))
2402                 return;
2403
2404         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2405         BUG_ON(rb_first(&cfqq->sort_list));
2406         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2407         cfqg = cfqq->cfqg;
2408         orig_cfqg = cfqq->orig_cfqg;
2409
2410         if (unlikely(cfqd->active_queue == cfqq)) {
2411                 __cfq_slice_expired(cfqd, cfqq, 0);
2412                 cfq_schedule_dispatch(cfqd);
2413         }
2414
2415         BUG_ON(cfq_cfqq_on_rr(cfqq));
2416         kmem_cache_free(cfq_pool, cfqq);
2417         cfq_put_cfqg(cfqg);
2418         if (orig_cfqg)
2419                 cfq_put_cfqg(orig_cfqg);
2420 }
2421
2422 /*
2423  * Must always be called with the rcu_read_lock() held
2424  */
2425 static void
2426 __call_for_each_cic(struct io_context *ioc,
2427                     void (*func)(struct io_context *, struct cfq_io_context *))
2428 {
2429         struct cfq_io_context *cic;
2430         struct hlist_node *n;
2431
2432         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2433                 func(ioc, cic);
2434 }
2435
2436 /*
2437  * Call func for each cic attached to this ioc.
2438  */
2439 static void
2440 call_for_each_cic(struct io_context *ioc,
2441                   void (*func)(struct io_context *, struct cfq_io_context *))
2442 {
2443         rcu_read_lock();
2444         __call_for_each_cic(ioc, func);
2445         rcu_read_unlock();
2446 }
2447
2448 static void cfq_cic_free_rcu(struct rcu_head *head)
2449 {
2450         struct cfq_io_context *cic;
2451
2452         cic = container_of(head, struct cfq_io_context, rcu_head);
2453
2454         kmem_cache_free(cfq_ioc_pool, cic);
2455         elv_ioc_count_dec(cfq_ioc_count);
2456
2457         if (ioc_gone) {
2458                 /*
2459                  * CFQ scheduler is exiting, grab exit lock and check
2460                  * the pending io context count. If it hits zero,
2461                  * complete ioc_gone and set it back to NULL
2462                  */
2463                 spin_lock(&ioc_gone_lock);
2464                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2465                         complete(ioc_gone);
2466                         ioc_gone = NULL;
2467                 }
2468                 spin_unlock(&ioc_gone_lock);
2469         }
2470 }
2471
2472 static void cfq_cic_free(struct cfq_io_context *cic)
2473 {
2474         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2475 }
2476
2477 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2478 {
2479         unsigned long flags;
2480
2481         BUG_ON(!cic->dead_key);
2482
2483         spin_lock_irqsave(&ioc->lock, flags);
2484         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2485         hlist_del_rcu(&cic->cic_list);
2486         spin_unlock_irqrestore(&ioc->lock, flags);
2487
2488         cfq_cic_free(cic);
2489 }
2490
2491 /*
2492  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2493  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2494  * and ->trim() which is called with the task lock held
2495  */
2496 static void cfq_free_io_context(struct io_context *ioc)
2497 {
2498         /*
2499          * ioc->refcount is zero here, or we are called from elv_unregister(),
2500          * so no more cic's are allowed to be linked into this ioc.  So it
2501          * should be ok to iterate over the known list, we will see all cic's
2502          * since no new ones are added.
2503          */
2504         __call_for_each_cic(ioc, cic_free_func);
2505 }
2506
2507 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2508 {
2509         struct cfq_queue *__cfqq, *next;
2510
2511         if (unlikely(cfqq == cfqd->active_queue)) {
2512                 __cfq_slice_expired(cfqd, cfqq, 0);
2513                 cfq_schedule_dispatch(cfqd);
2514         }
2515
2516         /*
2517          * If this queue was scheduled to merge with another queue, be
2518          * sure to drop the reference taken on that queue (and others in
2519          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2520          */
2521         __cfqq = cfqq->new_cfqq;
2522         while (__cfqq) {
2523                 if (__cfqq == cfqq) {
2524                         WARN(1, "cfqq->new_cfqq loop detected\n");
2525                         break;
2526                 }
2527                 next = __cfqq->new_cfqq;
2528                 cfq_put_queue(__cfqq);
2529                 __cfqq = next;
2530         }
2531
2532         cfq_put_queue(cfqq);
2533 }
2534
2535 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2536                                          struct cfq_io_context *cic)
2537 {
2538         struct io_context *ioc = cic->ioc;
2539
2540         list_del_init(&cic->queue_list);
2541
2542         /*
2543          * Make sure key == NULL is seen for dead queues
2544          */
2545         smp_wmb();
2546         cic->dead_key = (unsigned long) cic->key;
2547         cic->key = NULL;
2548
2549         if (ioc->ioc_data == cic)
2550                 rcu_assign_pointer(ioc->ioc_data, NULL);
2551
2552         if (cic->cfqq[BLK_RW_ASYNC]) {
2553                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2554                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2555         }
2556
2557         if (cic->cfqq[BLK_RW_SYNC]) {
2558                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2559                 cic->cfqq[BLK_RW_SYNC] = NULL;
2560         }
2561 }
2562
2563 static void cfq_exit_single_io_context(struct io_context *ioc,
2564                                        struct cfq_io_context *cic)
2565 {
2566         struct cfq_data *cfqd = cic->key;
2567
2568         if (cfqd) {
2569                 struct request_queue *q = cfqd->queue;
2570                 unsigned long flags;
2571
2572                 spin_lock_irqsave(q->queue_lock, flags);
2573
2574                 /*
2575                  * Ensure we get a fresh copy of the ->key to prevent
2576                  * race between exiting task and queue
2577                  */
2578                 smp_read_barrier_depends();
2579                 if (cic->key)
2580                         __cfq_exit_single_io_context(cfqd, cic);
2581
2582                 spin_unlock_irqrestore(q->queue_lock, flags);
2583         }
2584 }
2585
2586 /*
2587  * The process that ioc belongs to has exited, we need to clean up
2588  * and put the internal structures we have that belongs to that process.
2589  */
2590 static void cfq_exit_io_context(struct io_context *ioc)
2591 {
2592         call_for_each_cic(ioc, cfq_exit_single_io_context);
2593 }
2594
2595 static struct cfq_io_context *
2596 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2597 {
2598         struct cfq_io_context *cic;
2599
2600         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2601                                                         cfqd->queue->node);
2602         if (cic) {
2603                 cic->last_end_request = jiffies;
2604                 INIT_LIST_HEAD(&cic->queue_list);
2605                 INIT_HLIST_NODE(&cic->cic_list);
2606                 cic->dtor = cfq_free_io_context;
2607                 cic->exit = cfq_exit_io_context;
2608                 elv_ioc_count_inc(cfq_ioc_count);
2609         }
2610
2611         return cic;
2612 }
2613
2614 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2615 {
2616         struct task_struct *tsk = current;
2617         int ioprio_class;
2618
2619         if (!cfq_cfqq_prio_changed(cfqq))
2620                 return;
2621
2622         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2623         switch (ioprio_class) {
2624         default:
2625                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2626         case IOPRIO_CLASS_NONE:
2627                 /*
2628                  * no prio set, inherit CPU scheduling settings
2629                  */
2630                 cfqq->ioprio = task_nice_ioprio(tsk);
2631                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2632                 break;
2633         case IOPRIO_CLASS_RT:
2634                 cfqq->ioprio = task_ioprio(ioc);
2635                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2636                 break;
2637         case IOPRIO_CLASS_BE:
2638                 cfqq->ioprio = task_ioprio(ioc);
2639                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2640                 break;
2641         case IOPRIO_CLASS_IDLE:
2642                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2643                 cfqq->ioprio = 7;
2644                 cfq_clear_cfqq_idle_window(cfqq);
2645                 break;
2646         }
2647
2648         /*
2649          * keep track of original prio settings in case we have to temporarily
2650          * elevate the priority of this queue
2651          */
2652         cfqq->org_ioprio = cfqq->ioprio;
2653         cfqq->org_ioprio_class = cfqq->ioprio_class;
2654         cfq_clear_cfqq_prio_changed(cfqq);
2655 }
2656
2657 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2658 {
2659         struct cfq_data *cfqd = cic->key;
2660         struct cfq_queue *cfqq;
2661         unsigned long flags;
2662
2663         if (unlikely(!cfqd))
2664                 return;
2665
2666         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2667
2668         cfqq = cic->cfqq[BLK_RW_ASYNC];
2669         if (cfqq) {
2670                 struct cfq_queue *new_cfqq;
2671                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2672                                                 GFP_ATOMIC);
2673                 if (new_cfqq) {
2674                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2675                         cfq_put_queue(cfqq);
2676                 }
2677         }
2678
2679         cfqq = cic->cfqq[BLK_RW_SYNC];
2680         if (cfqq)
2681                 cfq_mark_cfqq_prio_changed(cfqq);
2682
2683         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2684 }
2685
2686 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2687 {
2688         call_for_each_cic(ioc, changed_ioprio);
2689         ioc->ioprio_changed = 0;
2690 }
2691
2692 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2693                           pid_t pid, bool is_sync)
2694 {
2695         RB_CLEAR_NODE(&cfqq->rb_node);
2696         RB_CLEAR_NODE(&cfqq->p_node);
2697         INIT_LIST_HEAD(&cfqq->fifo);
2698
2699         atomic_set(&cfqq->ref, 0);
2700         cfqq->cfqd = cfqd;
2701
2702         cfq_mark_cfqq_prio_changed(cfqq);
2703
2704         if (is_sync) {
2705                 if (!cfq_class_idle(cfqq))
2706                         cfq_mark_cfqq_idle_window(cfqq);
2707                 cfq_mark_cfqq_sync(cfqq);
2708         }
2709         cfqq->pid = pid;
2710 }
2711
2712 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2713 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2714 {
2715         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2716         struct cfq_data *cfqd = cic->key;
2717         unsigned long flags;
2718         struct request_queue *q;
2719
2720         if (unlikely(!cfqd))
2721                 return;
2722
2723         q = cfqd->queue;
2724
2725         spin_lock_irqsave(q->queue_lock, flags);
2726
2727         if (sync_cfqq) {
2728                 /*
2729                  * Drop reference to sync queue. A new sync queue will be
2730                  * assigned in new group upon arrival of a fresh request.
2731                  */
2732                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2733                 cic_set_cfqq(cic, NULL, 1);
2734                 cfq_put_queue(sync_cfqq);
2735         }
2736
2737         spin_unlock_irqrestore(q->queue_lock, flags);
2738 }
2739
2740 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2741 {
2742         call_for_each_cic(ioc, changed_cgroup);
2743         ioc->cgroup_changed = 0;
2744 }
2745 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2746
2747 static struct cfq_queue *
2748 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2749                      struct io_context *ioc, gfp_t gfp_mask)
2750 {
2751         struct cfq_queue *cfqq, *new_cfqq = NULL;
2752         struct cfq_io_context *cic;
2753         struct cfq_group *cfqg;
2754
2755 retry:
2756         cfqg = cfq_get_cfqg(cfqd, 1);
2757         cic = cfq_cic_lookup(cfqd, ioc);
2758         /* cic always exists here */
2759         cfqq = cic_to_cfqq(cic, is_sync);
2760
2761         /*
2762          * Always try a new alloc if we fell back to the OOM cfqq
2763          * originally, since it should just be a temporary situation.
2764          */
2765         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2766                 cfqq = NULL;
2767                 if (new_cfqq) {
2768                         cfqq = new_cfqq;
2769                         new_cfqq = NULL;
2770                 } else if (gfp_mask & __GFP_WAIT) {
2771                         spin_unlock_irq(cfqd->queue->queue_lock);
2772                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2773                                         gfp_mask | __GFP_ZERO,
2774                                         cfqd->queue->node);
2775                         spin_lock_irq(cfqd->queue->queue_lock);
2776                         if (new_cfqq)
2777                                 goto retry;
2778                 } else {
2779                         cfqq = kmem_cache_alloc_node(cfq_pool,
2780                                         gfp_mask | __GFP_ZERO,
2781                                         cfqd->queue->node);
2782                 }
2783
2784                 if (cfqq) {
2785                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2786                         cfq_init_prio_data(cfqq, ioc);
2787                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2788                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2789                 } else
2790                         cfqq = &cfqd->oom_cfqq;
2791         }
2792
2793         if (new_cfqq)
2794                 kmem_cache_free(cfq_pool, new_cfqq);
2795
2796         return cfqq;
2797 }
2798
2799 static struct cfq_queue **
2800 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2801 {
2802         switch (ioprio_class) {
2803         case IOPRIO_CLASS_RT:
2804                 return &cfqd->async_cfqq[0][ioprio];
2805         case IOPRIO_CLASS_BE:
2806                 return &cfqd->async_cfqq[1][ioprio];
2807         case IOPRIO_CLASS_IDLE:
2808                 return &cfqd->async_idle_cfqq;
2809         default:
2810                 BUG();
2811         }
2812 }
2813
2814 static struct cfq_queue *
2815 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2816               gfp_t gfp_mask)
2817 {
2818         const int ioprio = task_ioprio(ioc);
2819         const int ioprio_class = task_ioprio_class(ioc);
2820         struct cfq_queue **async_cfqq = NULL;
2821         struct cfq_queue *cfqq = NULL;
2822
2823         if (!is_sync) {
2824                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2825                 cfqq = *async_cfqq;
2826         }
2827
2828         if (!cfqq)
2829                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2830
2831         /*
2832          * pin the queue now that it's allocated, scheduler exit will prune it
2833          */
2834         if (!is_sync && !(*async_cfqq)) {
2835                 atomic_inc(&cfqq->ref);
2836                 *async_cfqq = cfqq;
2837         }
2838
2839         atomic_inc(&cfqq->ref);
2840         return cfqq;
2841 }
2842
2843 /*
2844  * We drop cfq io contexts lazily, so we may find a dead one.
2845  */
2846 static void
2847 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2848                   struct cfq_io_context *cic)
2849 {
2850         unsigned long flags;
2851
2852         WARN_ON(!list_empty(&cic->queue_list));
2853
2854         spin_lock_irqsave(&ioc->lock, flags);
2855
2856         BUG_ON(ioc->ioc_data == cic);
2857
2858         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2859         hlist_del_rcu(&cic->cic_list);
2860         spin_unlock_irqrestore(&ioc->lock, flags);
2861
2862         cfq_cic_free(cic);
2863 }
2864
2865 static struct cfq_io_context *
2866 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2867 {
2868         struct cfq_io_context *cic;
2869         unsigned long flags;
2870         void *k;
2871
2872         if (unlikely(!ioc))
2873                 return NULL;
2874
2875         rcu_read_lock();
2876
2877         /*
2878          * we maintain a last-hit cache, to avoid browsing over the tree
2879          */
2880         cic = rcu_dereference(ioc->ioc_data);
2881         if (cic && cic->key == cfqd) {
2882                 rcu_read_unlock();
2883                 return cic;
2884         }
2885
2886         do {
2887                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2888                 rcu_read_unlock();
2889                 if (!cic)
2890                         break;
2891                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2892                 k = cic->key;
2893                 if (unlikely(!k)) {
2894                         cfq_drop_dead_cic(cfqd, ioc, cic);
2895                         rcu_read_lock();
2896                         continue;
2897                 }
2898
2899                 spin_lock_irqsave(&ioc->lock, flags);
2900                 rcu_assign_pointer(ioc->ioc_data, cic);
2901                 spin_unlock_irqrestore(&ioc->lock, flags);
2902                 break;
2903         } while (1);
2904
2905         return cic;
2906 }
2907
2908 /*
2909  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2910  * the process specific cfq io context when entered from the block layer.
2911  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2912  */
2913 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2914                         struct cfq_io_context *cic, gfp_t gfp_mask)
2915 {
2916         unsigned long flags;
2917         int ret;
2918
2919         ret = radix_tree_preload(gfp_mask);
2920         if (!ret) {
2921                 cic->ioc = ioc;
2922                 cic->key = cfqd;
2923
2924                 spin_lock_irqsave(&ioc->lock, flags);
2925                 ret = radix_tree_insert(&ioc->radix_root,
2926                                                 (unsigned long) cfqd, cic);
2927                 if (!ret)
2928                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2929                 spin_unlock_irqrestore(&ioc->lock, flags);
2930
2931                 radix_tree_preload_end();
2932
2933                 if (!ret) {
2934                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2935                         list_add(&cic->queue_list, &cfqd->cic_list);
2936                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2937                 }
2938         }
2939
2940         if (ret)
2941                 printk(KERN_ERR "cfq: cic link failed!\n");
2942
2943         return ret;
2944 }
2945
2946 /*
2947  * Setup general io context and cfq io context. There can be several cfq
2948  * io contexts per general io context, if this process is doing io to more
2949  * than one device managed by cfq.
2950  */
2951 static struct cfq_io_context *
2952 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2953 {
2954         struct io_context *ioc = NULL;
2955         struct cfq_io_context *cic;
2956
2957         might_sleep_if(gfp_mask & __GFP_WAIT);
2958
2959         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2960         if (!ioc)
2961                 return NULL;
2962
2963         cic = cfq_cic_lookup(cfqd, ioc);
2964         if (cic)
2965                 goto out;
2966
2967         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2968         if (cic == NULL)
2969                 goto err;
2970
2971         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2972                 goto err_free;
2973
2974 out:
2975         smp_read_barrier_depends();
2976         if (unlikely(ioc->ioprio_changed))
2977                 cfq_ioc_set_ioprio(ioc);
2978
2979 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2980         if (unlikely(ioc->cgroup_changed))
2981                 cfq_ioc_set_cgroup(ioc);
2982 #endif
2983         return cic;
2984 err_free:
2985         cfq_cic_free(cic);
2986 err:
2987         put_io_context(ioc);
2988         return NULL;
2989 }
2990
2991 static void
2992 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2993 {
2994         unsigned long elapsed = jiffies - cic->last_end_request;
2995         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2996
2997         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2998         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2999         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3000 }
3001
3002 static void
3003 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3004                        struct request *rq)
3005 {
3006         sector_t sdist;
3007         u64 total;
3008
3009         if (!cfqq->last_request_pos)
3010                 sdist = 0;
3011         else if (cfqq->last_request_pos < blk_rq_pos(rq))
3012                 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3013         else
3014                 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3015
3016         /*
3017          * Don't allow the seek distance to get too large from the
3018          * odd fragment, pagein, etc
3019          */
3020         if (cfqq->seek_samples <= 60) /* second&third seek */
3021                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
3022         else
3023                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
3024
3025         cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
3026         cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
3027         total = cfqq->seek_total + (cfqq->seek_samples/2);
3028         do_div(total, cfqq->seek_samples);
3029         cfqq->seek_mean = (sector_t)total;
3030
3031         /*
3032          * If this cfqq is shared between multiple processes, check to
3033          * make sure that those processes are still issuing I/Os within
3034          * the mean seek distance.  If not, it may be time to break the
3035          * queues apart again.
3036          */
3037         if (cfq_cfqq_coop(cfqq)) {
3038                 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
3039                         cfqq->seeky_start = jiffies;
3040                 else if (!CFQQ_SEEKY(cfqq))
3041                         cfqq->seeky_start = 0;
3042         }
3043 }
3044
3045 /*
3046  * Disable idle window if the process thinks too long or seeks so much that
3047  * it doesn't matter
3048  */
3049 static void
3050 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3051                        struct cfq_io_context *cic)
3052 {
3053         int old_idle, enable_idle;
3054
3055         /*
3056          * Don't idle for async or idle io prio class
3057          */
3058         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3059                 return;
3060
3061         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3062
3063         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3064                 cfq_mark_cfqq_deep(cfqq);
3065
3066         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3067             (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3068              && CFQQ_SEEKY(cfqq)))
3069                 enable_idle = 0;
3070         else if (sample_valid(cic->ttime_samples)) {
3071                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3072                         enable_idle = 0;
3073                 else
3074                         enable_idle = 1;
3075         }
3076
3077         if (old_idle != enable_idle) {
3078                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3079                 if (enable_idle)
3080                         cfq_mark_cfqq_idle_window(cfqq);
3081                 else
3082                         cfq_clear_cfqq_idle_window(cfqq);
3083         }
3084 }
3085
3086 /*
3087  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3088  * no or if we aren't sure, a 1 will cause a preempt.
3089  */
3090 static bool
3091 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3092                    struct request *rq)
3093 {
3094         struct cfq_queue *cfqq;
3095
3096         cfqq = cfqd->active_queue;
3097         if (!cfqq)
3098                 return false;
3099
3100         if (cfq_class_idle(new_cfqq))
3101                 return false;
3102
3103         if (cfq_class_idle(cfqq))
3104                 return true;
3105
3106         /*
3107          * if the new request is sync, but the currently running queue is
3108          * not, let the sync request have priority.
3109          */
3110         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3111                 return true;
3112
3113         if (new_cfqq->cfqg != cfqq->cfqg)
3114                 return false;
3115
3116         if (cfq_slice_used(cfqq))
3117                 return true;
3118
3119         /* Allow preemption only if we are idling on sync-noidle tree */
3120         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3121             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3122             new_cfqq->service_tree->count == 2 &&
3123             RB_EMPTY_ROOT(&cfqq->sort_list))
3124                 return true;
3125
3126         /*
3127          * So both queues are sync. Let the new request get disk time if
3128          * it's a metadata request and the current queue is doing regular IO.
3129          */
3130         if (rq_is_meta(rq) && !cfqq->meta_pending)
3131                 return true;
3132
3133         /*
3134          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3135          */
3136         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3137                 return true;
3138
3139         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3140                 return false;
3141
3142         /*
3143          * if this request is as-good as one we would expect from the
3144          * current cfqq, let it preempt
3145          */
3146         if (cfq_rq_close(cfqd, cfqq, rq))
3147                 return true;
3148
3149         return false;
3150 }
3151
3152 /*
3153  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3154  * let it have half of its nominal slice.
3155  */
3156 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3157 {
3158         cfq_log_cfqq(cfqd, cfqq, "preempt");
3159         cfq_slice_expired(cfqd, 1);
3160
3161         /*
3162          * Put the new queue at the front of the of the current list,
3163          * so we know that it will be selected next.
3164          */
3165         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3166
3167         cfq_service_tree_add(cfqd, cfqq, 1);
3168
3169         cfqq->slice_end = 0;
3170         cfq_mark_cfqq_slice_new(cfqq);
3171 }
3172
3173 /*
3174  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3175  * something we should do about it
3176  */
3177 static void
3178 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3179                 struct request *rq)
3180 {
3181         struct cfq_io_context *cic = RQ_CIC(rq);
3182
3183         cfqd->rq_queued++;
3184         if (rq_is_meta(rq))
3185                 cfqq->meta_pending++;
3186
3187         cfq_update_io_thinktime(cfqd, cic);
3188         cfq_update_io_seektime(cfqd, cfqq, rq);
3189         cfq_update_idle_window(cfqd, cfqq, cic);
3190
3191         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3192
3193         if (cfqq == cfqd->active_queue) {
3194                 /*
3195                  * Remember that we saw a request from this process, but
3196                  * don't start queuing just yet. Otherwise we risk seeing lots
3197                  * of tiny requests, because we disrupt the normal plugging
3198                  * and merging. If the request is already larger than a single
3199                  * page, let it rip immediately. For that case we assume that
3200                  * merging is already done. Ditto for a busy system that
3201                  * has other work pending, don't risk delaying until the
3202                  * idle timer unplug to continue working.
3203                  */
3204                 if (cfq_cfqq_wait_request(cfqq)) {
3205                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3206                             cfqd->busy_queues > 1) {
3207                                 del_timer(&cfqd->idle_slice_timer);
3208                                 cfq_clear_cfqq_wait_request(cfqq);
3209                                 __blk_run_queue(cfqd->queue);
3210                         } else
3211                                 cfq_mark_cfqq_must_dispatch(cfqq);
3212                 }
3213         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3214                 /*
3215                  * not the active queue - expire current slice if it is
3216                  * idle and has expired it's mean thinktime or this new queue
3217                  * has some old slice time left and is of higher priority or
3218                  * this new queue is RT and the current one is BE
3219                  */
3220                 cfq_preempt_queue(cfqd, cfqq);
3221                 __blk_run_queue(cfqd->queue);
3222         }
3223 }
3224
3225 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3226 {
3227         struct cfq_data *cfqd = q->elevator->elevator_data;
3228         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3229
3230         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3231         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3232
3233         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3234         list_add_tail(&rq->queuelist, &cfqq->fifo);
3235         cfq_add_rq_rb(rq);
3236
3237         cfq_rq_enqueued(cfqd, cfqq, rq);
3238 }
3239
3240 /*
3241  * Update hw_tag based on peak queue depth over 50 samples under
3242  * sufficient load.
3243  */
3244 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3245 {
3246         struct cfq_queue *cfqq = cfqd->active_queue;
3247
3248         if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3249                 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3250
3251         if (cfqd->hw_tag == 1)
3252                 return;
3253
3254         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3255             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3256                 return;
3257
3258         /*
3259          * If active queue hasn't enough requests and can idle, cfq might not
3260          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3261          * case
3262          */
3263         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3264             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3265             CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3266                 return;
3267
3268         if (cfqd->hw_tag_samples++ < 50)
3269                 return;
3270
3271         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3272                 cfqd->hw_tag = 1;
3273         else
3274                 cfqd->hw_tag = 0;
3275 }
3276
3277 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3278 {
3279         struct cfq_io_context *cic = cfqd->active_cic;
3280
3281         /* If there are other queues in the group, don't wait */
3282         if (cfqq->cfqg->nr_cfqq > 1)
3283                 return false;
3284
3285         if (cfq_slice_used(cfqq))
3286                 return true;
3287
3288         /* if slice left is less than think time, wait busy */
3289         if (cic && sample_valid(cic->ttime_samples)
3290             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3291                 return true;
3292
3293         /*
3294          * If think times is less than a jiffy than ttime_mean=0 and above
3295          * will not be true. It might happen that slice has not expired yet
3296          * but will expire soon (4-5 ns) during select_queue(). To cover the
3297          * case where think time is less than a jiffy, mark the queue wait
3298          * busy if only 1 jiffy is left in the slice.
3299          */
3300         if (cfqq->slice_end - jiffies == 1)
3301                 return true;
3302
3303         return false;
3304 }
3305
3306 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3307 {
3308         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3309         struct cfq_data *cfqd = cfqq->cfqd;
3310         const int sync = rq_is_sync(rq);
3311         unsigned long now;
3312
3313         now = jiffies;
3314         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3315
3316         cfq_update_hw_tag(cfqd);
3317
3318         WARN_ON(!cfqd->rq_in_driver[sync]);
3319         WARN_ON(!cfqq->dispatched);
3320         cfqd->rq_in_driver[sync]--;
3321         cfqq->dispatched--;
3322
3323         if (cfq_cfqq_sync(cfqq))
3324                 cfqd->sync_flight--;
3325
3326         if (sync) {
3327                 RQ_CIC(rq)->last_end_request = now;
3328                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3329                         cfqd->last_delayed_sync = now;
3330         }
3331
3332         /*
3333          * If this is the active queue, check if it needs to be expired,
3334          * or if we want to idle in case it has no pending requests.
3335          */
3336         if (cfqd->active_queue == cfqq) {
3337                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3338
3339                 if (cfq_cfqq_slice_new(cfqq)) {
3340                         cfq_set_prio_slice(cfqd, cfqq);
3341                         cfq_clear_cfqq_slice_new(cfqq);
3342                 }
3343
3344                 /*
3345                  * Should we wait for next request to come in before we expire
3346                  * the queue.
3347                  */
3348                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3349                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3350                         cfq_mark_cfqq_wait_busy(cfqq);
3351                 }
3352
3353                 /*
3354                  * Idling is not enabled on:
3355                  * - expired queues
3356                  * - idle-priority queues
3357                  * - async queues
3358                  * - queues with still some requests queued
3359                  * - when there is a close cooperator
3360                  */
3361                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3362                         cfq_slice_expired(cfqd, 1);
3363                 else if (sync && cfqq_empty &&
3364                          !cfq_close_cooperator(cfqd, cfqq)) {
3365                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3366                         /*
3367                          * Idling is enabled for SYNC_WORKLOAD.
3368                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3369                          * only if we processed at least one !rq_noidle request
3370                          */
3371                         if (cfqd->serving_type == SYNC_WORKLOAD
3372                             || cfqd->noidle_tree_requires_idle
3373                             || cfqq->cfqg->nr_cfqq == 1)
3374                                 cfq_arm_slice_timer(cfqd);
3375                 }
3376         }
3377
3378         if (!rq_in_driver(cfqd))
3379                 cfq_schedule_dispatch(cfqd);
3380 }
3381
3382 /*
3383  * we temporarily boost lower priority queues if they are holding fs exclusive
3384  * resources. they are boosted to normal prio (CLASS_BE/4)
3385  */
3386 static void cfq_prio_boost(struct cfq_queue *cfqq)
3387 {
3388         if (has_fs_excl()) {
3389                 /*
3390                  * boost idle prio on transactions that would lock out other
3391                  * users of the filesystem
3392                  */
3393                 if (cfq_class_idle(cfqq))
3394                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3395                 if (cfqq->ioprio > IOPRIO_NORM)
3396                         cfqq->ioprio = IOPRIO_NORM;
3397         } else {
3398                 /*
3399                  * unboost the queue (if needed)
3400                  */
3401                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3402                 cfqq->ioprio = cfqq->org_ioprio;
3403         }
3404 }
3405
3406 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3407 {
3408         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3409                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3410                 return ELV_MQUEUE_MUST;
3411         }
3412
3413         return ELV_MQUEUE_MAY;
3414 }
3415
3416 static int cfq_may_queue(struct request_queue *q, int rw)
3417 {
3418         struct cfq_data *cfqd = q->elevator->elevator_data;
3419         struct task_struct *tsk = current;
3420         struct cfq_io_context *cic;
3421         struct cfq_queue *cfqq;
3422
3423         /*
3424          * don't force setup of a queue from here, as a call to may_queue
3425          * does not necessarily imply that a request actually will be queued.
3426          * so just lookup a possibly existing queue, or return 'may queue'
3427          * if that fails
3428          */
3429         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3430         if (!cic)
3431                 return ELV_MQUEUE_MAY;
3432
3433         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3434         if (cfqq) {
3435                 cfq_init_prio_data(cfqq, cic->ioc);
3436                 cfq_prio_boost(cfqq);
3437
3438                 return __cfq_may_queue(cfqq);
3439         }
3440
3441         return ELV_MQUEUE_MAY;
3442 }
3443
3444 /*
3445  * queue lock held here
3446  */
3447 static void cfq_put_request(struct request *rq)
3448 {
3449         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3450
3451         if (cfqq) {
3452                 const int rw = rq_data_dir(rq);
3453
3454                 BUG_ON(!cfqq->allocated[rw]);
3455                 cfqq->allocated[rw]--;
3456
3457                 put_io_context(RQ_CIC(rq)->ioc);
3458
3459                 rq->elevator_private = NULL;
3460                 rq->elevator_private2 = NULL;
3461
3462                 cfq_put_queue(cfqq);
3463         }
3464 }
3465
3466 static struct cfq_queue *
3467 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3468                 struct cfq_queue *cfqq)
3469 {
3470         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3471         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3472         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3473         cfq_put_queue(cfqq);
3474         return cic_to_cfqq(cic, 1);
3475 }
3476
3477 static int should_split_cfqq(struct cfq_queue *cfqq)
3478 {
3479         if (cfqq->seeky_start &&
3480             time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3481                 return 1;
3482         return 0;
3483 }
3484
3485 /*
3486  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3487  * was the last process referring to said cfqq.
3488  */
3489 static struct cfq_queue *
3490 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3491 {
3492         if (cfqq_process_refs(cfqq) == 1) {
3493                 cfqq->seeky_start = 0;
3494                 cfqq->pid = current->pid;
3495                 cfq_clear_cfqq_coop(cfqq);
3496                 return cfqq;
3497         }
3498
3499         cic_set_cfqq(cic, NULL, 1);
3500         cfq_put_queue(cfqq);
3501         return NULL;
3502 }
3503 /*
3504  * Allocate cfq data structures associated with this request.
3505  */
3506 static int
3507 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3508 {
3509         struct cfq_data *cfqd = q->elevator->elevator_data;
3510         struct cfq_io_context *cic;
3511         const int rw = rq_data_dir(rq);
3512         const bool is_sync = rq_is_sync(rq);
3513         struct cfq_queue *cfqq;
3514         unsigned long flags;
3515
3516         might_sleep_if(gfp_mask & __GFP_WAIT);
3517
3518         cic = cfq_get_io_context(cfqd, gfp_mask);
3519
3520         spin_lock_irqsave(q->queue_lock, flags);
3521
3522         if (!cic)
3523                 goto queue_fail;
3524
3525 new_queue:
3526         cfqq = cic_to_cfqq(cic, is_sync);
3527         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3528                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3529                 cic_set_cfqq(cic, cfqq, is_sync);
3530         } else {
3531                 /*
3532                  * If the queue was seeky for too long, break it apart.
3533                  */
3534                 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3535                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3536                         cfqq = split_cfqq(cic, cfqq);
3537                         if (!cfqq)
3538                                 goto new_queue;
3539                 }
3540
3541                 /*
3542                  * Check to see if this queue is scheduled to merge with
3543                  * another, closely cooperating queue.  The merging of
3544                  * queues happens here as it must be done in process context.
3545                  * The reference on new_cfqq was taken in merge_cfqqs.
3546                  */
3547                 if (cfqq->new_cfqq)
3548                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3549         }
3550
3551         cfqq->allocated[rw]++;
3552         atomic_inc(&cfqq->ref);
3553
3554         spin_unlock_irqrestore(q->queue_lock, flags);
3555
3556         rq->elevator_private = cic;
3557         rq->elevator_private2 = cfqq;
3558         return 0;
3559
3560 queue_fail:
3561         if (cic)
3562                 put_io_context(cic->ioc);
3563
3564         cfq_schedule_dispatch(cfqd);
3565         spin_unlock_irqrestore(q->queue_lock, flags);
3566         cfq_log(cfqd, "set_request fail");
3567         return 1;
3568 }
3569
3570 static void cfq_kick_queue(struct work_struct *work)
3571 {
3572         struct cfq_data *cfqd =
3573                 container_of(work, struct cfq_data, unplug_work);
3574         struct request_queue *q = cfqd->queue;
3575
3576         spin_lock_irq(q->queue_lock);
3577         __blk_run_queue(cfqd->queue);
3578         spin_unlock_irq(q->queue_lock);
3579 }
3580
3581 /*
3582  * Timer running if the active_queue is currently idling inside its time slice
3583  */
3584 static void cfq_idle_slice_timer(unsigned long data)
3585 {
3586         struct cfq_data *cfqd = (struct cfq_data *) data;
3587         struct cfq_queue *cfqq;
3588         unsigned long flags;
3589         int timed_out = 1;
3590
3591         cfq_log(cfqd, "idle timer fired");
3592
3593         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3594
3595         cfqq = cfqd->active_queue;
3596         if (cfqq) {
3597                 timed_out = 0;
3598
3599                 /*
3600                  * We saw a request before the queue expired, let it through
3601                  */
3602                 if (cfq_cfqq_must_dispatch(cfqq))
3603                         goto out_kick;
3604
3605                 /*
3606                  * expired
3607                  */
3608                 if (cfq_slice_used(cfqq))
3609                         goto expire;
3610
3611                 /*
3612                  * only expire and reinvoke request handler, if there are
3613                  * other queues with pending requests
3614                  */
3615                 if (!cfqd->busy_queues)
3616                         goto out_cont;
3617
3618                 /*
3619                  * not expired and it has a request pending, let it dispatch
3620                  */
3621                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3622                         goto out_kick;
3623
3624                 /*
3625                  * Queue depth flag is reset only when the idle didn't succeed
3626                  */
3627                 cfq_clear_cfqq_deep(cfqq);
3628         }
3629 expire:
3630         cfq_slice_expired(cfqd, timed_out);
3631 out_kick:
3632         cfq_schedule_dispatch(cfqd);
3633 out_cont:
3634         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3635 }
3636
3637 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3638 {
3639         del_timer_sync(&cfqd->idle_slice_timer);
3640         cancel_work_sync(&cfqd->unplug_work);
3641 }
3642
3643 static void cfq_put_async_queues(struct cfq_data *cfqd)
3644 {
3645         int i;
3646
3647         for (i = 0; i < IOPRIO_BE_NR; i++) {
3648                 if (cfqd->async_cfqq[0][i])
3649                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3650                 if (cfqd->async_cfqq[1][i])
3651                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3652         }
3653
3654         if (cfqd->async_idle_cfqq)
3655                 cfq_put_queue(cfqd->async_idle_cfqq);
3656 }
3657
3658 static void cfq_cfqd_free(struct rcu_head *head)
3659 {
3660         kfree(container_of(head, struct cfq_data, rcu));
3661 }
3662
3663 static void cfq_exit_queue(struct elevator_queue *e)
3664 {
3665         struct cfq_data *cfqd = e->elevator_data;
3666         struct request_queue *q = cfqd->queue;
3667
3668         cfq_shutdown_timer_wq(cfqd);
3669
3670         spin_lock_irq(q->queue_lock);
3671
3672         if (cfqd->active_queue)
3673                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3674
3675         while (!list_empty(&cfqd->cic_list)) {
3676                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3677                                                         struct cfq_io_context,
3678                                                         queue_list);
3679
3680                 __cfq_exit_single_io_context(cfqd, cic);
3681         }
3682
3683         cfq_put_async_queues(cfqd);
3684         cfq_release_cfq_groups(cfqd);
3685         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3686
3687         spin_unlock_irq(q->queue_lock);
3688
3689         cfq_shutdown_timer_wq(cfqd);
3690
3691         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3692         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3693 }
3694
3695 static void *cfq_init_queue(struct request_queue *q)
3696 {
3697         struct cfq_data *cfqd;
3698         int i, j;
3699         struct cfq_group *cfqg;
3700         struct cfq_rb_root *st;
3701
3702         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3703         if (!cfqd)
3704                 return NULL;
3705
3706         /* Init root service tree */
3707         cfqd->grp_service_tree = CFQ_RB_ROOT;
3708
3709         /* Init root group */
3710         cfqg = &cfqd->root_group;
3711         for_each_cfqg_st(cfqg, i, j, st)
3712                 *st = CFQ_RB_ROOT;
3713         RB_CLEAR_NODE(&cfqg->rb_node);
3714
3715         /* Give preference to root group over other groups */
3716         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3717
3718 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3719         /*
3720          * Take a reference to root group which we never drop. This is just
3721          * to make sure that cfq_put_cfqg() does not try to kfree root group
3722          */
3723         atomic_set(&cfqg->ref, 1);
3724         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3725                                         0);
3726 #endif
3727         /*
3728          * Not strictly needed (since RB_ROOT just clears the node and we
3729          * zeroed cfqd on alloc), but better be safe in case someone decides
3730          * to add magic to the rb code
3731          */
3732         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3733                 cfqd->prio_trees[i] = RB_ROOT;
3734
3735         /*
3736          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3737          * Grab a permanent reference to it, so that the normal code flow
3738          * will not attempt to free it.
3739          */
3740         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3741         atomic_inc(&cfqd->oom_cfqq.ref);
3742         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3743
3744         INIT_LIST_HEAD(&cfqd->cic_list);
3745
3746         cfqd->queue = q;
3747
3748         init_timer(&cfqd->idle_slice_timer);
3749         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3750         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3751
3752         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3753
3754         cfqd->cfq_quantum = cfq_quantum;
3755         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3756         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3757         cfqd->cfq_back_max = cfq_back_max;
3758         cfqd->cfq_back_penalty = cfq_back_penalty;
3759         cfqd->cfq_slice[0] = cfq_slice_async;
3760         cfqd->cfq_slice[1] = cfq_slice_sync;
3761         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3762         cfqd->cfq_slice_idle = cfq_slice_idle;
3763         cfqd->cfq_latency = 1;
3764         cfqd->cfq_group_isolation = 0;
3765         cfqd->hw_tag = -1;
3766         /*
3767          * we optimistically start assuming sync ops weren't delayed in last
3768          * second, in order to have larger depth for async operations.
3769          */
3770         cfqd->last_delayed_sync = jiffies - HZ;
3771         INIT_RCU_HEAD(&cfqd->rcu);
3772         return cfqd;
3773 }
3774
3775 static void cfq_slab_kill(void)
3776 {
3777         /*
3778          * Caller already ensured that pending RCU callbacks are completed,
3779          * so we should have no busy allocations at this point.
3780          */
3781         if (cfq_pool)
3782                 kmem_cache_destroy(cfq_pool);
3783         if (cfq_ioc_pool)
3784                 kmem_cache_destroy(cfq_ioc_pool);
3785 }
3786
3787 static int __init cfq_slab_setup(void)
3788 {
3789         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3790         if (!cfq_pool)
3791                 goto fail;
3792
3793         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3794         if (!cfq_ioc_pool)
3795                 goto fail;
3796
3797         return 0;
3798 fail:
3799         cfq_slab_kill();
3800         return -ENOMEM;
3801 }
3802
3803 /*
3804  * sysfs parts below -->
3805  */
3806 static ssize_t
3807 cfq_var_show(unsigned int var, char *page)
3808 {
3809         return sprintf(page, "%d\n", var);
3810 }
3811
3812 static ssize_t
3813 cfq_var_store(unsigned int *var, const char *page, size_t count)
3814 {
3815         char *p = (char *) page;
3816
3817         *var = simple_strtoul(p, &p, 10);
3818         return count;
3819 }
3820
3821 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3822 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3823 {                                                                       \
3824         struct cfq_data *cfqd = e->elevator_data;                       \
3825         unsigned int __data = __VAR;                                    \
3826         if (__CONV)                                                     \
3827                 __data = jiffies_to_msecs(__data);                      \
3828         return cfq_var_show(__data, (page));                            \
3829 }
3830 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3831 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3832 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3833 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3834 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3835 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3836 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3837 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3838 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3839 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3840 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3841 #undef SHOW_FUNCTION
3842
3843 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3844 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3845 {                                                                       \
3846         struct cfq_data *cfqd = e->elevator_data;                       \
3847         unsigned int __data;                                            \
3848         int ret = cfq_var_store(&__data, (page), count);                \
3849         if (__data < (MIN))                                             \
3850                 __data = (MIN);                                         \
3851         else if (__data > (MAX))                                        \
3852                 __data = (MAX);                                         \
3853         if (__CONV)                                                     \
3854                 *(__PTR) = msecs_to_jiffies(__data);                    \
3855         else                                                            \
3856                 *(__PTR) = __data;                                      \
3857         return ret;                                                     \
3858 }
3859 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3860 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3861                 UINT_MAX, 1);
3862 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3863                 UINT_MAX, 1);
3864 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3865 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3866                 UINT_MAX, 0);
3867 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3868 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3869 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3870 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3871                 UINT_MAX, 0);
3872 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3873 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3874 #undef STORE_FUNCTION
3875
3876 #define CFQ_ATTR(name) \
3877         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3878
3879 static struct elv_fs_entry cfq_attrs[] = {
3880         CFQ_ATTR(quantum),
3881         CFQ_ATTR(fifo_expire_sync),
3882         CFQ_ATTR(fifo_expire_async),
3883         CFQ_ATTR(back_seek_max),
3884         CFQ_ATTR(back_seek_penalty),
3885         CFQ_ATTR(slice_sync),
3886         CFQ_ATTR(slice_async),
3887         CFQ_ATTR(slice_async_rq),
3888         CFQ_ATTR(slice_idle),
3889         CFQ_ATTR(low_latency),
3890         CFQ_ATTR(group_isolation),
3891         __ATTR_NULL
3892 };
3893
3894 static struct elevator_type iosched_cfq = {
3895         .ops = {
3896                 .elevator_merge_fn =            cfq_merge,
3897                 .elevator_merged_fn =           cfq_merged_request,
3898                 .elevator_merge_req_fn =        cfq_merged_requests,
3899                 .elevator_allow_merge_fn =      cfq_allow_merge,
3900                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3901                 .elevator_add_req_fn =          cfq_insert_request,
3902                 .elevator_activate_req_fn =     cfq_activate_request,
3903                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3904                 .elevator_queue_empty_fn =      cfq_queue_empty,
3905                 .elevator_completed_req_fn =    cfq_completed_request,
3906                 .elevator_former_req_fn =       elv_rb_former_request,
3907                 .elevator_latter_req_fn =       elv_rb_latter_request,
3908                 .elevator_set_req_fn =          cfq_set_request,
3909                 .elevator_put_req_fn =          cfq_put_request,
3910                 .elevator_may_queue_fn =        cfq_may_queue,
3911                 .elevator_init_fn =             cfq_init_queue,
3912                 .elevator_exit_fn =             cfq_exit_queue,
3913                 .trim =                         cfq_free_io_context,
3914         },
3915         .elevator_attrs =       cfq_attrs,
3916         .elevator_name =        "cfq",
3917         .elevator_owner =       THIS_MODULE,
3918 };
3919
3920 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3921 static struct blkio_policy_type blkio_policy_cfq = {
3922         .ops = {
3923                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3924                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3925         },
3926 };
3927 #else
3928 static struct blkio_policy_type blkio_policy_cfq;
3929 #endif
3930
3931 static int __init cfq_init(void)
3932 {
3933         /*
3934          * could be 0 on HZ < 1000 setups
3935          */
3936         if (!cfq_slice_async)
3937                 cfq_slice_async = 1;
3938         if (!cfq_slice_idle)
3939                 cfq_slice_idle = 1;
3940
3941         if (cfq_slab_setup())
3942                 return -ENOMEM;
3943
3944         elv_register(&iosched_cfq);
3945         blkio_policy_register(&blkio_policy_cfq);
3946
3947         return 0;
3948 }
3949
3950 static void __exit cfq_exit(void)
3951 {
3952         DECLARE_COMPLETION_ONSTACK(all_gone);
3953         blkio_policy_unregister(&blkio_policy_cfq);
3954         elv_unregister(&iosched_cfq);
3955         ioc_gone = &all_gone;
3956         /* ioc_gone's update must be visible before reading ioc_count */
3957         smp_wmb();
3958
3959         /*
3960          * this also protects us from entering cfq_slab_kill() with
3961          * pending RCU callbacks
3962          */
3963         if (elv_ioc_count_read(cfq_ioc_count))
3964                 wait_for_completion(&all_gone);
3965         cfq_slab_kill();
3966 }
3967
3968 module_init(cfq_init);
3969 module_exit(cfq_exit);
3970
3971 MODULE_AUTHOR("Jens Axboe");
3972 MODULE_LICENSE("GPL");
3973 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");