linux/block/blk-throttle.c
<<
>>
Prefs
   1/*
   2 * Interface for controlling IO bandwidth on a request queue
   3 *
   4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   5 */
   6
   7#include <linux/module.h>
   8#include <linux/slab.h>
   9#include <linux/blkdev.h>
  10#include <linux/bio.h>
  11#include <linux/blktrace_api.h>
  12#include "blk-cgroup.h"
  13
  14/* Max dispatch from a group in 1 round */
  15static int throtl_grp_quantum = 8;
  16
  17/* Total max dispatch from all groups in one round */
  18static int throtl_quantum = 32;
  19
  20/* Throttling is performed over 100ms slice and after that slice is renewed */
  21static unsigned long throtl_slice = HZ/10;      /* 100 ms */
  22
  23/* A workqueue to queue throttle related work */
  24static struct workqueue_struct *kthrotld_workqueue;
  25static void throtl_schedule_delayed_work(struct throtl_data *td,
  26                                unsigned long delay);
  27
  28struct throtl_rb_root {
  29        struct rb_root rb;
  30        struct rb_node *left;
  31        unsigned int count;
  32        unsigned long min_disptime;
  33};
  34
  35#define THROTL_RB_ROOT  (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
  36                        .count = 0, .min_disptime = 0}
  37
  38#define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  39
  40struct throtl_grp {
  41        /* List of throtl groups on the request queue*/
  42        struct hlist_node tg_node;
  43
  44        /* active throtl group service_tree member */
  45        struct rb_node rb_node;
  46
  47        /*
  48         * Dispatch time in jiffies. This is the estimated time when group
  49         * will unthrottle and is ready to dispatch more bio. It is used as
  50         * key to sort active groups in service tree.
  51         */
  52        unsigned long disptime;
  53
  54        struct blkio_group blkg;
  55        atomic_t ref;
  56        unsigned int flags;
  57
  58        /* Two lists for READ and WRITE */
  59        struct bio_list bio_lists[2];
  60
  61        /* Number of queued bios on READ and WRITE lists */
  62        unsigned int nr_queued[2];
  63
  64        /* bytes per second rate limits */
  65        uint64_t bps[2];
  66
  67        /* IOPS limits */
  68        unsigned int iops[2];
  69
  70        /* Number of bytes disptached in current slice */
  71        uint64_t bytes_disp[2];
  72        /* Number of bio's dispatched in current slice */
  73        unsigned int io_disp[2];
  74
  75        /* When did we start a new slice */
  76        unsigned long slice_start[2];
  77        unsigned long slice_end[2];
  78
  79        /* Some throttle limits got updated for the group */
  80        int limits_changed;
  81};
  82
  83struct throtl_data
  84{
  85        /* List of throtl groups */
  86        struct hlist_head tg_list;
  87
  88        /* service tree for active throtl groups */
  89        struct throtl_rb_root tg_service_tree;
  90
  91        struct throtl_grp root_tg;
  92        struct request_queue *queue;
  93
  94        /* Total Number of queued bios on READ and WRITE lists */
  95        unsigned int nr_queued[2];
  96
  97        /*
  98         * number of total undestroyed groups
  99         */
 100        unsigned int nr_undestroyed_grps;
 101
 102        /* Work for dispatching throttled bios */
 103        struct delayed_work throtl_work;
 104
 105        int limits_changed;
 106};
 107
 108enum tg_state_flags {
 109        THROTL_TG_FLAG_on_rr = 0,       /* on round-robin busy list */
 110};
 111
 112#define THROTL_TG_FNS(name)                                             \
 113static inline void throtl_mark_tg_##name(struct throtl_grp *tg)         \
 114{                                                                       \
 115        (tg)->flags |= (1 << THROTL_TG_FLAG_##name);                    \
 116}                                                                       \
 117static inline void throtl_clear_tg_##name(struct throtl_grp *tg)        \
 118{                                                                       \
 119        (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name);                   \
 120}                                                                       \
 121static inline int throtl_tg_##name(const struct throtl_grp *tg)         \
 122{                                                                       \
 123        return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0;       \
 124}
 125
 126THROTL_TG_FNS(on_rr);
 127
 128#define throtl_log_tg(td, tg, fmt, args...)                             \
 129        blk_add_trace_msg((td)->queue, "throtl %s " fmt,                \
 130                                blkg_path(&(tg)->blkg), ##args);        \
 131
 132#define throtl_log(td, fmt, args...)    \
 133        blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
 134
 135static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
 136{
 137        if (blkg)
 138                return container_of(blkg, struct throtl_grp, blkg);
 139
 140        return NULL;
 141}
 142
 143static inline int total_nr_queued(struct throtl_data *td)
 144{
 145        return (td->nr_queued[0] + td->nr_queued[1]);
 146}
 147
 148static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
 149{
 150        atomic_inc(&tg->ref);
 151        return tg;
 152}
 153
 154static void throtl_put_tg(struct throtl_grp *tg)
 155{
 156        BUG_ON(atomic_read(&tg->ref) <= 0);
 157        if (!atomic_dec_and_test(&tg->ref))
 158                return;
 159        kfree(tg);
 160}
 161
 162static struct throtl_grp * throtl_find_alloc_tg(struct throtl_data *td,
 163                        struct blkio_cgroup *blkcg)
 164{
 165        struct throtl_grp *tg = NULL;
 166        void *key = td;
 167        struct backing_dev_info *bdi = &td->queue->backing_dev_info;
 168        unsigned int major, minor;
 169
 170        /*
 171         * TODO: Speed up blkiocg_lookup_group() by maintaining a radix
 172         * tree of blkg (instead of traversing through hash list all
 173         * the time.
 174         */
 175
 176        /*
 177         * This is the common case when there are no blkio cgroups.
 178         * Avoid lookup in this case
 179         */
 180        if (blkcg == &blkio_root_cgroup)
 181                tg = &td->root_tg;
 182        else
 183                tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
 184
 185        /* Fill in device details for root group */
 186        if (tg && !tg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
 187                sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
 188                tg->blkg.dev = MKDEV(major, minor);
 189                goto done;
 190        }
 191
 192        if (tg)
 193                goto done;
 194
 195        tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
 196        if (!tg)
 197                goto done;
 198
 199        INIT_HLIST_NODE(&tg->tg_node);
 200        RB_CLEAR_NODE(&tg->rb_node);
 201        bio_list_init(&tg->bio_lists[0]);
 202        bio_list_init(&tg->bio_lists[1]);
 203        td->limits_changed = false;
 204
 205        /*
 206         * Take the initial reference that will be released on destroy
 207         * This can be thought of a joint reference by cgroup and
 208         * request queue which will be dropped by either request queue
 209         * exit or cgroup deletion path depending on who is exiting first.
 210         */
 211        atomic_set(&tg->ref, 1);
 212
 213        /* Add group onto cgroup list */
 214        sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
 215        blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
 216                                MKDEV(major, minor), BLKIO_POLICY_THROTL);
 217
 218        tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
 219        tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
 220        tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
 221        tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
 222
 223        hlist_add_head(&tg->tg_node, &td->tg_list);
 224        td->nr_undestroyed_grps++;
 225done:
 226        return tg;
 227}
 228
 229static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
 230{
 231        struct throtl_grp *tg = NULL;
 232        struct blkio_cgroup *blkcg;
 233
 234        rcu_read_lock();
 235        blkcg = task_blkio_cgroup(current);
 236        tg = throtl_find_alloc_tg(td, blkcg);
 237        if (!tg)
 238                tg = &td->root_tg;
 239        rcu_read_unlock();
 240        return tg;
 241}
 242
 243static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
 244{
 245        /* Service tree is empty */
 246        if (!root->count)
 247                return NULL;
 248
 249        if (!root->left)
 250                root->left = rb_first(&root->rb);
 251
 252        if (root->left)
 253                return rb_entry_tg(root->left);
 254
 255        return NULL;
 256}
 257
 258static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 259{
 260        rb_erase(n, root);
 261        RB_CLEAR_NODE(n);
 262}
 263
 264static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
 265{
 266        if (root->left == n)
 267                root->left = NULL;
 268        rb_erase_init(n, &root->rb);
 269        --root->count;
 270}
 271
 272static void update_min_dispatch_time(struct throtl_rb_root *st)
 273{
 274        struct throtl_grp *tg;
 275
 276        tg = throtl_rb_first(st);
 277        if (!tg)
 278                return;
 279
 280        st->min_disptime = tg->disptime;
 281}
 282
 283static void
 284tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
 285{
 286        struct rb_node **node = &st->rb.rb_node;
 287        struct rb_node *parent = NULL;
 288        struct throtl_grp *__tg;
 289        unsigned long key = tg->disptime;
 290        int left = 1;
 291
 292        while (*node != NULL) {
 293                parent = *node;
 294                __tg = rb_entry_tg(parent);
 295
 296                if (time_before(key, __tg->disptime))
 297                        node = &parent->rb_left;
 298                else {
 299                        node = &parent->rb_right;
 300                        left = 0;
 301                }
 302        }
 303
 304        if (left)
 305                st->left = &tg->rb_node;
 306
 307        rb_link_node(&tg->rb_node, parent, node);
 308        rb_insert_color(&tg->rb_node, &st->rb);
 309}
 310
 311static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
 312{
 313        struct throtl_rb_root *st = &td->tg_service_tree;
 314
 315        tg_service_tree_add(st, tg);
 316        throtl_mark_tg_on_rr(tg);
 317        st->count++;
 318}
 319
 320static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
 321{
 322        if (!throtl_tg_on_rr(tg))
 323                __throtl_enqueue_tg(td, tg);
 324}
 325
 326static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
 327{
 328        throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
 329        throtl_clear_tg_on_rr(tg);
 330}
 331
 332static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
 333{
 334        if (throtl_tg_on_rr(tg))
 335                __throtl_dequeue_tg(td, tg);
 336}
 337
 338static void throtl_schedule_next_dispatch(struct throtl_data *td)
 339{
 340        struct throtl_rb_root *st = &td->tg_service_tree;
 341
 342        /*
 343         * If there are more bios pending, schedule more work.
 344         */
 345        if (!total_nr_queued(td))
 346                return;
 347
 348        BUG_ON(!st->count);
 349
 350        update_min_dispatch_time(st);
 351
 352        if (time_before_eq(st->min_disptime, jiffies))
 353                throtl_schedule_delayed_work(td, 0);
 354        else
 355                throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
 356}
 357
 358static inline void
 359throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 360{
 361        tg->bytes_disp[rw] = 0;
 362        tg->io_disp[rw] = 0;
 363        tg->slice_start[rw] = jiffies;
 364        tg->slice_end[rw] = jiffies + throtl_slice;
 365        throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
 366                        rw == READ ? 'R' : 'W', tg->slice_start[rw],
 367                        tg->slice_end[rw], jiffies);
 368}
 369
 370static inline void throtl_set_slice_end(struct throtl_data *td,
 371                struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
 372{
 373        tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 374}
 375
 376static inline void throtl_extend_slice(struct throtl_data *td,
 377                struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
 378{
 379        tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 380        throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 381                        rw == READ ? 'R' : 'W', tg->slice_start[rw],
 382                        tg->slice_end[rw], jiffies);
 383}
 384
 385/* Determine if previously allocated or extended slice is complete or not */
 386static bool
 387throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 388{
 389        if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 390                return 0;
 391
 392        return 1;
 393}
 394
 395/* Trim the used slices and adjust slice start accordingly */
 396static inline void
 397throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 398{
 399        unsigned long nr_slices, time_elapsed, io_trim;
 400        u64 bytes_trim, tmp;
 401
 402        BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 403
 404        /*
 405         * If bps are unlimited (-1), then time slice don't get
 406         * renewed. Don't try to trim the slice if slice is used. A new
 407         * slice will start when appropriate.
 408         */
 409        if (throtl_slice_used(td, tg, rw))
 410                return;
 411
 412        /*
 413         * A bio has been dispatched. Also adjust slice_end. It might happen
 414         * that initially cgroup limit was very low resulting in high
 415         * slice_end, but later limit was bumped up and bio was dispached
 416         * sooner, then we need to reduce slice_end. A high bogus slice_end
 417         * is bad because it does not allow new slice to start.
 418         */
 419
 420        throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
 421
 422        time_elapsed = jiffies - tg->slice_start[rw];
 423
 424        nr_slices = time_elapsed / throtl_slice;
 425
 426        if (!nr_slices)
 427                return;
 428        tmp = tg->bps[rw] * throtl_slice * nr_slices;
 429        do_div(tmp, HZ);
 430        bytes_trim = tmp;
 431
 432        io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
 433
 434        if (!bytes_trim && !io_trim)
 435                return;
 436
 437        if (tg->bytes_disp[rw] >= bytes_trim)
 438                tg->bytes_disp[rw] -= bytes_trim;
 439        else
 440                tg->bytes_disp[rw] = 0;
 441
 442        if (tg->io_disp[rw] >= io_trim)
 443                tg->io_disp[rw] -= io_trim;
 444        else
 445                tg->io_disp[rw] = 0;
 446
 447        tg->slice_start[rw] += nr_slices * throtl_slice;
 448
 449        throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
 450                        " start=%lu end=%lu jiffies=%lu",
 451                        rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 452                        tg->slice_start[rw], tg->slice_end[rw], jiffies);
 453}
 454
 455static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
 456                struct bio *bio, unsigned long *wait)
 457{
 458        bool rw = bio_data_dir(bio);
 459        unsigned int io_allowed;
 460        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 461        u64 tmp;
 462
 463        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 464
 465        /* Slice has just started. Consider one slice interval */
 466        if (!jiffy_elapsed)
 467                jiffy_elapsed_rnd = throtl_slice;
 468
 469        jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 470
 471        /*
 472         * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 473         * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 474         * will allow dispatch after 1 second and after that slice should
 475         * have been trimmed.
 476         */
 477
 478        tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
 479        do_div(tmp, HZ);
 480
 481        if (tmp > UINT_MAX)
 482                io_allowed = UINT_MAX;
 483        else
 484                io_allowed = tmp;
 485
 486        if (tg->io_disp[rw] + 1 <= io_allowed) {
 487                if (wait)
 488                        *wait = 0;
 489                return 1;
 490        }
 491
 492        /* Calc approx time to dispatch */
 493        jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
 494
 495        if (jiffy_wait > jiffy_elapsed)
 496                jiffy_wait = jiffy_wait - jiffy_elapsed;
 497        else
 498                jiffy_wait = 1;
 499
 500        if (wait)
 501                *wait = jiffy_wait;
 502        return 0;
 503}
 504
 505static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
 506                struct bio *bio, unsigned long *wait)
 507{
 508        bool rw = bio_data_dir(bio);
 509        u64 bytes_allowed, extra_bytes, tmp;
 510        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 511
 512        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 513
 514        /* Slice has just started. Consider one slice interval */
 515        if (!jiffy_elapsed)
 516                jiffy_elapsed_rnd = throtl_slice;
 517
 518        jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 519
 520        tmp = tg->bps[rw] * jiffy_elapsed_rnd;
 521        do_div(tmp, HZ);
 522        bytes_allowed = tmp;
 523
 524        if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
 525                if (wait)
 526                        *wait = 0;
 527                return 1;
 528        }
 529
 530        /* Calc approx time to dispatch */
 531        extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
 532        jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
 533
 534        if (!jiffy_wait)
 535                jiffy_wait = 1;
 536
 537        /*
 538         * This wait time is without taking into consideration the rounding
 539         * up we did. Add that time also.
 540         */
 541        jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 542        if (wait)
 543                *wait = jiffy_wait;
 544        return 0;
 545}
 546
 547/*
 548 * Returns whether one can dispatch a bio or not. Also returns approx number
 549 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 550 */
 551static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
 552                                struct bio *bio, unsigned long *wait)
 553{
 554        bool rw = bio_data_dir(bio);
 555        unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 556
 557        /*
 558         * Currently whole state machine of group depends on first bio
 559         * queued in the group bio list. So one should not be calling
 560         * this function with a different bio if there are other bios
 561         * queued.
 562         */
 563        BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
 564
 565        /* If tg->bps = -1, then BW is unlimited */
 566        if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
 567                if (wait)
 568                        *wait = 0;
 569                return 1;
 570        }
 571
 572        /*
 573         * If previous slice expired, start a new one otherwise renew/extend
 574         * existing slice to make sure it is at least throtl_slice interval
 575         * long since now.
 576         */
 577        if (throtl_slice_used(td, tg, rw))
 578                throtl_start_new_slice(td, tg, rw);
 579        else {
 580                if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
 581                        throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
 582        }
 583
 584        if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
 585            && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
 586                if (wait)
 587                        *wait = 0;
 588                return 1;
 589        }
 590
 591        max_wait = max(bps_wait, iops_wait);
 592
 593        if (wait)
 594                *wait = max_wait;
 595
 596        if (time_before(tg->slice_end[rw], jiffies + max_wait))
 597                throtl_extend_slice(td, tg, rw, jiffies + max_wait);
 598
 599        return 0;
 600}
 601
 602static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 603{
 604        bool rw = bio_data_dir(bio);
 605        bool sync = bio->bi_rw & REQ_SYNC;
 606
 607        /* Charge the bio to the group */
 608        tg->bytes_disp[rw] += bio->bi_size;
 609        tg->io_disp[rw]++;
 610
 611        /*
 612         * TODO: This will take blkg->stats_lock. Figure out a way
 613         * to avoid this cost.
 614         */
 615        blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
 616}
 617
 618static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
 619                        struct bio *bio)
 620{
 621        bool rw = bio_data_dir(bio);
 622
 623        bio_list_add(&tg->bio_lists[rw], bio);
 624        /* Take a bio reference on tg */
 625        throtl_ref_get_tg(tg);
 626        tg->nr_queued[rw]++;
 627        td->nr_queued[rw]++;
 628        throtl_enqueue_tg(td, tg);
 629}
 630
 631static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
 632{
 633        unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
 634        struct bio *bio;
 635
 636        if ((bio = bio_list_peek(&tg->bio_lists[READ])))
 637                tg_may_dispatch(td, tg, bio, &read_wait);
 638
 639        if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
 640                tg_may_dispatch(td, tg, bio, &write_wait);
 641
 642        min_wait = min(read_wait, write_wait);
 643        disptime = jiffies + min_wait;
 644
 645        /* Update dispatch time */
 646        throtl_dequeue_tg(td, tg);
 647        tg->disptime = disptime;
 648        throtl_enqueue_tg(td, tg);
 649}
 650
 651static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
 652                                bool rw, struct bio_list *bl)
 653{
 654        struct bio *bio;
 655
 656        bio = bio_list_pop(&tg->bio_lists[rw]);
 657        tg->nr_queued[rw]--;
 658        /* Drop bio reference on tg */
 659        throtl_put_tg(tg);
 660
 661        BUG_ON(td->nr_queued[rw] <= 0);
 662        td->nr_queued[rw]--;
 663
 664        throtl_charge_bio(tg, bio);
 665        bio_list_add(bl, bio);
 666        bio->bi_rw |= REQ_THROTTLED;
 667
 668        throtl_trim_slice(td, tg, rw);
 669}
 670
 671static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
 672                                struct bio_list *bl)
 673{
 674        unsigned int nr_reads = 0, nr_writes = 0;
 675        unsigned int max_nr_reads = throtl_grp_quantum*3/4;
 676        unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
 677        struct bio *bio;
 678
 679        /* Try to dispatch 75% READS and 25% WRITES */
 680
 681        while ((bio = bio_list_peek(&tg->bio_lists[READ]))
 682                && tg_may_dispatch(td, tg, bio, NULL)) {
 683
 684                tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
 685                nr_reads++;
 686
 687                if (nr_reads >= max_nr_reads)
 688                        break;
 689        }
 690
 691        while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
 692                && tg_may_dispatch(td, tg, bio, NULL)) {
 693
 694                tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
 695                nr_writes++;
 696
 697                if (nr_writes >= max_nr_writes)
 698                        break;
 699        }
 700
 701        return nr_reads + nr_writes;
 702}
 703
 704static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
 705{
 706        unsigned int nr_disp = 0;
 707        struct throtl_grp *tg;
 708        struct throtl_rb_root *st = &td->tg_service_tree;
 709
 710        while (1) {
 711                tg = throtl_rb_first(st);
 712
 713                if (!tg)
 714                        break;
 715
 716                if (time_before(jiffies, tg->disptime))
 717                        break;
 718
 719                throtl_dequeue_tg(td, tg);
 720
 721                nr_disp += throtl_dispatch_tg(td, tg, bl);
 722
 723                if (tg->nr_queued[0] || tg->nr_queued[1]) {
 724                        tg_update_disptime(td, tg);
 725                        throtl_enqueue_tg(td, tg);
 726                }
 727
 728                if (nr_disp >= throtl_quantum)
 729                        break;
 730        }
 731
 732        return nr_disp;
 733}
 734
 735static void throtl_process_limit_change(struct throtl_data *td)
 736{
 737        struct throtl_grp *tg;
 738        struct hlist_node *pos, *n;
 739
 740        if (!td->limits_changed)
 741                return;
 742
 743        xchg(&td->limits_changed, false);
 744
 745        throtl_log(td, "limits changed");
 746
 747        hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
 748                if (!tg->limits_changed)
 749                        continue;
 750
 751                if (!xchg(&tg->limits_changed, false))
 752                        continue;
 753
 754                throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
 755                        " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
 756                        tg->iops[READ], tg->iops[WRITE]);
 757
 758                /*
 759                 * Restart the slices for both READ and WRITES. It
 760                 * might happen that a group's limit are dropped
 761                 * suddenly and we don't want to account recently
 762                 * dispatched IO with new low rate
 763                 */
 764                throtl_start_new_slice(td, tg, 0);
 765                throtl_start_new_slice(td, tg, 1);
 766
 767                if (throtl_tg_on_rr(tg))
 768                        tg_update_disptime(td, tg);
 769        }
 770}
 771
 772/* Dispatch throttled bios. Should be called without queue lock held. */
 773static int throtl_dispatch(struct request_queue *q)
 774{
 775        struct throtl_data *td = q->td;
 776        unsigned int nr_disp = 0;
 777        struct bio_list bio_list_on_stack;
 778        struct bio *bio;
 779        struct blk_plug plug;
 780
 781        spin_lock_irq(q->queue_lock);
 782
 783        throtl_process_limit_change(td);
 784
 785        if (!total_nr_queued(td))
 786                goto out;
 787
 788        bio_list_init(&bio_list_on_stack);
 789
 790        throtl_log(td, "dispatch nr_queued=%lu read=%u write=%u",
 791                        total_nr_queued(td), td->nr_queued[READ],
 792                        td->nr_queued[WRITE]);
 793
 794        nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
 795
 796        if (nr_disp)
 797                throtl_log(td, "bios disp=%u", nr_disp);
 798
 799        throtl_schedule_next_dispatch(td);
 800out:
 801        spin_unlock_irq(q->queue_lock);
 802
 803        /*
 804         * If we dispatched some requests, unplug the queue to make sure
 805         * immediate dispatch
 806         */
 807        if (nr_disp) {
 808                blk_start_plug(&plug);
 809                while((bio = bio_list_pop(&bio_list_on_stack)))
 810                        generic_make_request(bio);
 811                blk_finish_plug(&plug);
 812        }
 813        return nr_disp;
 814}
 815
 816void blk_throtl_work(struct work_struct *work)
 817{
 818        struct throtl_data *td = container_of(work, struct throtl_data,
 819                                        throtl_work.work);
 820        struct request_queue *q = td->queue;
 821
 822        throtl_dispatch(q);
 823}
 824
 825/* Call with queue lock held */
 826static void
 827throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
 828{
 829
 830        struct delayed_work *dwork = &td->throtl_work;
 831
 832        /* schedule work if limits changed even if no bio is queued */
 833        if (total_nr_queued(td) > 0 || td->limits_changed) {
 834                /*
 835                 * We might have a work scheduled to be executed in future.
 836                 * Cancel that and schedule a new one.
 837                 */
 838                __cancel_delayed_work(dwork);
 839                queue_delayed_work(kthrotld_workqueue, dwork, delay);
 840                throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
 841                                delay, jiffies);
 842        }
 843}
 844
 845static void
 846throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
 847{
 848        /* Something wrong if we are trying to remove same group twice */
 849        BUG_ON(hlist_unhashed(&tg->tg_node));
 850
 851        hlist_del_init(&tg->tg_node);
 852
 853        /*
 854         * Put the reference taken at the time of creation so that when all
 855         * queues are gone, group can be destroyed.
 856         */
 857        throtl_put_tg(tg);
 858        td->nr_undestroyed_grps--;
 859}
 860
 861static void throtl_release_tgs(struct throtl_data *td)
 862{
 863        struct hlist_node *pos, *n;
 864        struct throtl_grp *tg;
 865
 866        hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
 867                /*
 868                 * If cgroup removal path got to blk_group first and removed
 869                 * it from cgroup list, then it will take care of destroying
 870                 * cfqg also.
 871                 */
 872                if (!blkiocg_del_blkio_group(&tg->blkg))
 873                        throtl_destroy_tg(td, tg);
 874        }
 875}
 876
 877static void throtl_td_free(struct throtl_data *td)
 878{
 879        kfree(td);
 880}
 881
 882/*
 883 * Blk cgroup controller notification saying that blkio_group object is being
 884 * delinked as associated cgroup object is going away. That also means that
 885 * no new IO will come in this group. So get rid of this group as soon as
 886 * any pending IO in the group is finished.
 887 *
 888 * This function is called under rcu_read_lock(). key is the rcu protected
 889 * pointer. That means "key" is a valid throtl_data pointer as long as we are
 890 * rcu read lock.
 891 *
 892 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
 893 * it should not be NULL as even if queue was going away, cgroup deltion
 894 * path got to it first.
 895 */
 896void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
 897{
 898        unsigned long flags;
 899        struct throtl_data *td = key;
 900
 901        spin_lock_irqsave(td->queue->queue_lock, flags);
 902        throtl_destroy_tg(td, tg_of_blkg(blkg));
 903        spin_unlock_irqrestore(td->queue->queue_lock, flags);
 904}
 905
 906static void throtl_update_blkio_group_common(struct throtl_data *td,
 907                                struct throtl_grp *tg)
 908{
 909        xchg(&tg->limits_changed, true);
 910        xchg(&td->limits_changed, true);
 911        /* Schedule a work now to process the limit change */
 912        throtl_schedule_delayed_work(td, 0);
 913}
 914
 915/*
 916 * For all update functions, key should be a valid pointer because these
 917 * update functions are called under blkcg_lock, that means, blkg is
 918 * valid and in turn key is valid. queue exit path can not race because
 919 * of blkcg_lock
 920 *
 921 * Can not take queue lock in update functions as queue lock under blkcg_lock
 922 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
 923 */
 924static void throtl_update_blkio_group_read_bps(void *key,
 925                                struct blkio_group *blkg, u64 read_bps)
 926{
 927        struct throtl_data *td = key;
 928        struct throtl_grp *tg = tg_of_blkg(blkg);
 929
 930        tg->bps[READ] = read_bps;
 931        throtl_update_blkio_group_common(td, tg);
 932}
 933
 934static void throtl_update_blkio_group_write_bps(void *key,
 935                                struct blkio_group *blkg, u64 write_bps)
 936{
 937        struct throtl_data *td = key;
 938        struct throtl_grp *tg = tg_of_blkg(blkg);
 939
 940        tg->bps[WRITE] = write_bps;
 941        throtl_update_blkio_group_common(td, tg);
 942}
 943
 944static void throtl_update_blkio_group_read_iops(void *key,
 945                        struct blkio_group *blkg, unsigned int read_iops)
 946{
 947        struct throtl_data *td = key;
 948        struct throtl_grp *tg = tg_of_blkg(blkg);
 949
 950        tg->iops[READ] = read_iops;
 951        throtl_update_blkio_group_common(td, tg);
 952}
 953
 954static void throtl_update_blkio_group_write_iops(void *key,
 955                        struct blkio_group *blkg, unsigned int write_iops)
 956{
 957        struct throtl_data *td = key;
 958        struct throtl_grp *tg = tg_of_blkg(blkg);
 959
 960        tg->iops[WRITE] = write_iops;
 961        throtl_update_blkio_group_common(td, tg);
 962}
 963
 964static void throtl_shutdown_wq(struct request_queue *q)
 965{
 966        struct throtl_data *td = q->td;
 967
 968        cancel_delayed_work_sync(&td->throtl_work);
 969}
 970
 971static struct blkio_policy_type blkio_policy_throtl = {
 972        .ops = {
 973                .blkio_unlink_group_fn = throtl_unlink_blkio_group,
 974                .blkio_update_group_read_bps_fn =
 975                                        throtl_update_blkio_group_read_bps,
 976                .blkio_update_group_write_bps_fn =
 977                                        throtl_update_blkio_group_write_bps,
 978                .blkio_update_group_read_iops_fn =
 979                                        throtl_update_blkio_group_read_iops,
 980                .blkio_update_group_write_iops_fn =
 981                                        throtl_update_blkio_group_write_iops,
 982        },
 983        .plid = BLKIO_POLICY_THROTL,
 984};
 985
 986int blk_throtl_bio(struct request_queue *q, struct bio **biop)
 987{
 988        struct throtl_data *td = q->td;
 989        struct throtl_grp *tg;
 990        struct bio *bio = *biop;
 991        bool rw = bio_data_dir(bio), update_disptime = true;
 992
 993        if (bio->bi_rw & REQ_THROTTLED) {
 994                bio->bi_rw &= ~REQ_THROTTLED;
 995                return 0;
 996        }
 997
 998        spin_lock_irq(q->queue_lock);
 999        tg = throtl_get_tg(td);
1000
1001        if (tg->nr_queued[rw]) {
1002                /*
1003                 * There is already another bio queued in same dir. No
1004                 * need to update dispatch time.
1005                 */
1006                update_disptime = false;
1007                goto queue_bio;
1008
1009        }
1010
1011        /* Bio is with-in rate limit of group */
1012        if (tg_may_dispatch(td, tg, bio, NULL)) {
1013                throtl_charge_bio(tg, bio);
1014
1015                /*
1016                 * We need to trim slice even when bios are not being queued
1017                 * otherwise it might happen that a bio is not queued for
1018                 * a long time and slice keeps on extending and trim is not
1019                 * called for a long time. Now if limits are reduced suddenly
1020                 * we take into account all the IO dispatched so far at new
1021                 * low rate and * newly queued IO gets a really long dispatch
1022                 * time.
1023                 *
1024                 * So keep on trimming slice even if bio is not queued.
1025                 */
1026                throtl_trim_slice(td, tg, rw);
1027                goto out;
1028        }
1029
1030queue_bio:
1031        throtl_log_tg(td, tg, "[%c] bio. bdisp=%u sz=%u bps=%llu"
1032                        " iodisp=%u iops=%u queued=%d/%d",
1033                        rw == READ ? 'R' : 'W',
1034                        tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1035                        tg->io_disp[rw], tg->iops[rw],
1036                        tg->nr_queued[READ], tg->nr_queued[WRITE]);
1037
1038        throtl_add_bio_tg(q->td, tg, bio);
1039        *biop = NULL;
1040
1041        if (update_disptime) {
1042                tg_update_disptime(td, tg);
1043                throtl_schedule_next_dispatch(td);
1044        }
1045
1046out:
1047        spin_unlock_irq(q->queue_lock);
1048        return 0;
1049}
1050
1051int blk_throtl_init(struct request_queue *q)
1052{
1053        struct throtl_data *td;
1054        struct throtl_grp *tg;
1055
1056        td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1057        if (!td)
1058                return -ENOMEM;
1059
1060        INIT_HLIST_HEAD(&td->tg_list);
1061        td->tg_service_tree = THROTL_RB_ROOT;
1062        td->limits_changed = false;
1063
1064        /* Init root group */
1065        tg = &td->root_tg;
1066        INIT_HLIST_NODE(&tg->tg_node);
1067        RB_CLEAR_NODE(&tg->rb_node);
1068        bio_list_init(&tg->bio_lists[0]);
1069        bio_list_init(&tg->bio_lists[1]);
1070
1071        /* Practically unlimited BW */
1072        tg->bps[0] = tg->bps[1] = -1;
1073        tg->iops[0] = tg->iops[1] = -1;
1074        td->limits_changed = false;
1075
1076        /*
1077         * Set root group reference to 2. One reference will be dropped when
1078         * all groups on tg_list are being deleted during queue exit. Other
1079         * reference will remain there as we don't want to delete this group
1080         * as it is statically allocated and gets destroyed when throtl_data
1081         * goes away.
1082         */
1083        atomic_set(&tg->ref, 2);
1084        hlist_add_head(&tg->tg_node, &td->tg_list);
1085        td->nr_undestroyed_grps++;
1086
1087        INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1088
1089        rcu_read_lock();
1090        blkiocg_add_blkio_group(&blkio_root_cgroup, &tg->blkg, (void *)td,
1091                                        0, BLKIO_POLICY_THROTL);
1092        rcu_read_unlock();
1093
1094        /* Attach throtl data to request queue */
1095        td->queue = q;
1096        q->td = td;
1097        return 0;
1098}
1099
1100void blk_throtl_exit(struct request_queue *q)
1101{
1102        struct throtl_data *td = q->td;
1103        bool wait = false;
1104
1105        BUG_ON(!td);
1106
1107        throtl_shutdown_wq(q);
1108
1109        spin_lock_irq(q->queue_lock);
1110        throtl_release_tgs(td);
1111
1112        /* If there are other groups */
1113        if (td->nr_undestroyed_grps > 0)
1114                wait = true;
1115
1116        spin_unlock_irq(q->queue_lock);
1117
1118        /*
1119         * Wait for tg->blkg->key accessors to exit their grace periods.
1120         * Do this wait only if there are other undestroyed groups out
1121         * there (other than root group). This can happen if cgroup deletion
1122         * path claimed the responsibility of cleaning up a group before
1123         * queue cleanup code get to the group.
1124         *
1125         * Do not call synchronize_rcu() unconditionally as there are drivers
1126         * which create/delete request queue hundreds of times during scan/boot
1127         * and synchronize_rcu() can take significant time and slow down boot.
1128         */
1129        if (wait)
1130                synchronize_rcu();
1131
1132        /*
1133         * Just being safe to make sure after previous flush if some body did
1134         * update limits through cgroup and another work got queued, cancel
1135         * it.
1136         */
1137        throtl_shutdown_wq(q);
1138        throtl_td_free(td);
1139}
1140
1141static int __init throtl_init(void)
1142{
1143        kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1144        if (!kthrotld_workqueue)
1145                panic("Failed to create kthrotld\n");
1146
1147        blkio_policy_register(&blkio_policy_throtl);
1148        return 0;
1149}
1150
1151module_init(throtl_init);
1152