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#include "blk.h"
  14
  15/* Max dispatch from a group in 1 round */
  16static int throtl_grp_quantum = 8;
  17
  18/* Total max dispatch from all groups in one round */
  19static int throtl_quantum = 32;
  20
  21/* Throttling is performed over 100ms slice and after that slice is renewed */
  22static unsigned long throtl_slice = HZ/10;      /* 100 ms */
  23
  24static struct blkcg_policy blkcg_policy_throtl;
  25
  26/* A workqueue to queue throttle related work */
  27static struct workqueue_struct *kthrotld_workqueue;
  28static void throtl_schedule_delayed_work(struct throtl_data *td,
  29                                unsigned long delay);
  30
  31struct throtl_rb_root {
  32        struct rb_root rb;
  33        struct rb_node *left;
  34        unsigned int count;
  35        unsigned long min_disptime;
  36};
  37
  38#define THROTL_RB_ROOT  (struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
  39                        .count = 0, .min_disptime = 0}
  40
  41#define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  42
  43/* Per-cpu group stats */
  44struct tg_stats_cpu {
  45        /* total bytes transferred */
  46        struct blkg_rwstat              service_bytes;
  47        /* total IOs serviced, post merge */
  48        struct blkg_rwstat              serviced;
  49};
  50
  51struct throtl_grp {
  52        /* must be the first member */
  53        struct blkg_policy_data pd;
  54
  55        /* active throtl group service_tree member */
  56        struct rb_node rb_node;
  57
  58        /*
  59         * Dispatch time in jiffies. This is the estimated time when group
  60         * will unthrottle and is ready to dispatch more bio. It is used as
  61         * key to sort active groups in service tree.
  62         */
  63        unsigned long disptime;
  64
  65        unsigned int flags;
  66
  67        /* Two lists for READ and WRITE */
  68        struct bio_list bio_lists[2];
  69
  70        /* Number of queued bios on READ and WRITE lists */
  71        unsigned int nr_queued[2];
  72
  73        /* bytes per second rate limits */
  74        uint64_t bps[2];
  75
  76        /* IOPS limits */
  77        unsigned int iops[2];
  78
  79        /* Number of bytes disptached in current slice */
  80        uint64_t bytes_disp[2];
  81        /* Number of bio's dispatched in current slice */
  82        unsigned int io_disp[2];
  83
  84        /* When did we start a new slice */
  85        unsigned long slice_start[2];
  86        unsigned long slice_end[2];
  87
  88        /* Some throttle limits got updated for the group */
  89        int limits_changed;
  90
  91        /* Per cpu stats pointer */
  92        struct tg_stats_cpu __percpu *stats_cpu;
  93
  94        /* List of tgs waiting for per cpu stats memory to be allocated */
  95        struct list_head stats_alloc_node;
  96};
  97
  98struct throtl_data
  99{
 100        /* service tree for active throtl groups */
 101        struct throtl_rb_root tg_service_tree;
 102
 103        struct request_queue *queue;
 104
 105        /* Total Number of queued bios on READ and WRITE lists */
 106        unsigned int nr_queued[2];
 107
 108        /*
 109         * number of total undestroyed groups
 110         */
 111        unsigned int nr_undestroyed_grps;
 112
 113        /* Work for dispatching throttled bios */
 114        struct delayed_work throtl_work;
 115
 116        int limits_changed;
 117};
 118
 119/* list and work item to allocate percpu group stats */
 120static DEFINE_SPINLOCK(tg_stats_alloc_lock);
 121static LIST_HEAD(tg_stats_alloc_list);
 122
 123static void tg_stats_alloc_fn(struct work_struct *);
 124static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
 125
 126static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 127{
 128        return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 129}
 130
 131static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 132{
 133        return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 134}
 135
 136static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 137{
 138        return pd_to_blkg(&tg->pd);
 139}
 140
 141static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
 142{
 143        return blkg_to_tg(td->queue->root_blkg);
 144}
 145
 146enum tg_state_flags {
 147        THROTL_TG_FLAG_on_rr = 0,       /* on round-robin busy list */
 148};
 149
 150#define THROTL_TG_FNS(name)                                             \
 151static inline void throtl_mark_tg_##name(struct throtl_grp *tg)         \
 152{                                                                       \
 153        (tg)->flags |= (1 << THROTL_TG_FLAG_##name);                    \
 154}                                                                       \
 155static inline void throtl_clear_tg_##name(struct throtl_grp *tg)        \
 156{                                                                       \
 157        (tg)->flags &= ~(1 << THROTL_TG_FLAG_##name);                   \
 158}                                                                       \
 159static inline int throtl_tg_##name(const struct throtl_grp *tg)         \
 160{                                                                       \
 161        return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0;       \
 162}
 163
 164THROTL_TG_FNS(on_rr);
 165
 166#define throtl_log_tg(td, tg, fmt, args...)     do {                    \
 167        char __pbuf[128];                                               \
 168                                                                        \
 169        blkg_path(tg_to_blkg(tg), __pbuf, sizeof(__pbuf));              \
 170        blk_add_trace_msg((td)->queue, "throtl %s " fmt, __pbuf, ##args); \
 171} while (0)
 172
 173#define throtl_log(td, fmt, args...)    \
 174        blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
 175
 176static inline unsigned int total_nr_queued(struct throtl_data *td)
 177{
 178        return td->nr_queued[0] + td->nr_queued[1];
 179}
 180
 181/*
 182 * Worker for allocating per cpu stat for tgs. This is scheduled on the
 183 * system_nrt_wq once there are some groups on the alloc_list waiting for
 184 * allocation.
 185 */
 186static void tg_stats_alloc_fn(struct work_struct *work)
 187{
 188        static struct tg_stats_cpu *stats_cpu;  /* this fn is non-reentrant */
 189        struct delayed_work *dwork = to_delayed_work(work);
 190        bool empty = false;
 191
 192alloc_stats:
 193        if (!stats_cpu) {
 194                stats_cpu = alloc_percpu(struct tg_stats_cpu);
 195                if (!stats_cpu) {
 196                        /* allocation failed, try again after some time */
 197                        queue_delayed_work(system_nrt_wq, dwork,
 198                                           msecs_to_jiffies(10));
 199                        return;
 200                }
 201        }
 202
 203        spin_lock_irq(&tg_stats_alloc_lock);
 204
 205        if (!list_empty(&tg_stats_alloc_list)) {
 206                struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
 207                                                         struct throtl_grp,
 208                                                         stats_alloc_node);
 209                swap(tg->stats_cpu, stats_cpu);
 210                list_del_init(&tg->stats_alloc_node);
 211        }
 212
 213        empty = list_empty(&tg_stats_alloc_list);
 214        spin_unlock_irq(&tg_stats_alloc_lock);
 215        if (!empty)
 216                goto alloc_stats;
 217}
 218
 219static void throtl_pd_init(struct blkcg_gq *blkg)
 220{
 221        struct throtl_grp *tg = blkg_to_tg(blkg);
 222        unsigned long flags;
 223
 224        RB_CLEAR_NODE(&tg->rb_node);
 225        bio_list_init(&tg->bio_lists[0]);
 226        bio_list_init(&tg->bio_lists[1]);
 227        tg->limits_changed = false;
 228
 229        tg->bps[READ] = -1;
 230        tg->bps[WRITE] = -1;
 231        tg->iops[READ] = -1;
 232        tg->iops[WRITE] = -1;
 233
 234        /*
 235         * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
 236         * but percpu allocator can't be called from IO path.  Queue tg on
 237         * tg_stats_alloc_list and allocate from work item.
 238         */
 239        spin_lock_irqsave(&tg_stats_alloc_lock, flags);
 240        list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
 241        queue_delayed_work(system_nrt_wq, &tg_stats_alloc_work, 0);
 242        spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
 243}
 244
 245static void throtl_pd_exit(struct blkcg_gq *blkg)
 246{
 247        struct throtl_grp *tg = blkg_to_tg(blkg);
 248        unsigned long flags;
 249
 250        spin_lock_irqsave(&tg_stats_alloc_lock, flags);
 251        list_del_init(&tg->stats_alloc_node);
 252        spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
 253
 254        free_percpu(tg->stats_cpu);
 255}
 256
 257static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
 258{
 259        struct throtl_grp *tg = blkg_to_tg(blkg);
 260        int cpu;
 261
 262        if (tg->stats_cpu == NULL)
 263                return;
 264
 265        for_each_possible_cpu(cpu) {
 266                struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
 267
 268                blkg_rwstat_reset(&sc->service_bytes);
 269                blkg_rwstat_reset(&sc->serviced);
 270        }
 271}
 272
 273static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
 274                                           struct blkcg *blkcg)
 275{
 276        /*
 277         * This is the common case when there are no blkcgs.  Avoid lookup
 278         * in this case
 279         */
 280        if (blkcg == &blkcg_root)
 281                return td_root_tg(td);
 282
 283        return blkg_to_tg(blkg_lookup(blkcg, td->queue));
 284}
 285
 286static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
 287                                                  struct blkcg *blkcg)
 288{
 289        struct request_queue *q = td->queue;
 290        struct throtl_grp *tg = NULL;
 291
 292        /*
 293         * This is the common case when there are no blkcgs.  Avoid lookup
 294         * in this case
 295         */
 296        if (blkcg == &blkcg_root) {
 297                tg = td_root_tg(td);
 298        } else {
 299                struct blkcg_gq *blkg;
 300
 301                blkg = blkg_lookup_create(blkcg, q);
 302
 303                /* if %NULL and @q is alive, fall back to root_tg */
 304                if (!IS_ERR(blkg))
 305                        tg = blkg_to_tg(blkg);
 306                else if (!blk_queue_dead(q))
 307                        tg = td_root_tg(td);
 308        }
 309
 310        return tg;
 311}
 312
 313static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
 314{
 315        /* Service tree is empty */
 316        if (!root->count)
 317                return NULL;
 318
 319        if (!root->left)
 320                root->left = rb_first(&root->rb);
 321
 322        if (root->left)
 323                return rb_entry_tg(root->left);
 324
 325        return NULL;
 326}
 327
 328static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 329{
 330        rb_erase(n, root);
 331        RB_CLEAR_NODE(n);
 332}
 333
 334static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
 335{
 336        if (root->left == n)
 337                root->left = NULL;
 338        rb_erase_init(n, &root->rb);
 339        --root->count;
 340}
 341
 342static void update_min_dispatch_time(struct throtl_rb_root *st)
 343{
 344        struct throtl_grp *tg;
 345
 346        tg = throtl_rb_first(st);
 347        if (!tg)
 348                return;
 349
 350        st->min_disptime = tg->disptime;
 351}
 352
 353static void
 354tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
 355{
 356        struct rb_node **node = &st->rb.rb_node;
 357        struct rb_node *parent = NULL;
 358        struct throtl_grp *__tg;
 359        unsigned long key = tg->disptime;
 360        int left = 1;
 361
 362        while (*node != NULL) {
 363                parent = *node;
 364                __tg = rb_entry_tg(parent);
 365
 366                if (time_before(key, __tg->disptime))
 367                        node = &parent->rb_left;
 368                else {
 369                        node = &parent->rb_right;
 370                        left = 0;
 371                }
 372        }
 373
 374        if (left)
 375                st->left = &tg->rb_node;
 376
 377        rb_link_node(&tg->rb_node, parent, node);
 378        rb_insert_color(&tg->rb_node, &st->rb);
 379}
 380
 381static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
 382{
 383        struct throtl_rb_root *st = &td->tg_service_tree;
 384
 385        tg_service_tree_add(st, tg);
 386        throtl_mark_tg_on_rr(tg);
 387        st->count++;
 388}
 389
 390static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
 391{
 392        if (!throtl_tg_on_rr(tg))
 393                __throtl_enqueue_tg(td, tg);
 394}
 395
 396static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
 397{
 398        throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
 399        throtl_clear_tg_on_rr(tg);
 400}
 401
 402static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
 403{
 404        if (throtl_tg_on_rr(tg))
 405                __throtl_dequeue_tg(td, tg);
 406}
 407
 408static void throtl_schedule_next_dispatch(struct throtl_data *td)
 409{
 410        struct throtl_rb_root *st = &td->tg_service_tree;
 411
 412        /*
 413         * If there are more bios pending, schedule more work.
 414         */
 415        if (!total_nr_queued(td))
 416                return;
 417
 418        BUG_ON(!st->count);
 419
 420        update_min_dispatch_time(st);
 421
 422        if (time_before_eq(st->min_disptime, jiffies))
 423                throtl_schedule_delayed_work(td, 0);
 424        else
 425                throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
 426}
 427
 428static inline void
 429throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 430{
 431        tg->bytes_disp[rw] = 0;
 432        tg->io_disp[rw] = 0;
 433        tg->slice_start[rw] = jiffies;
 434        tg->slice_end[rw] = jiffies + throtl_slice;
 435        throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
 436                        rw == READ ? 'R' : 'W', tg->slice_start[rw],
 437                        tg->slice_end[rw], jiffies);
 438}
 439
 440static inline void throtl_set_slice_end(struct throtl_data *td,
 441                struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
 442{
 443        tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 444}
 445
 446static inline void throtl_extend_slice(struct throtl_data *td,
 447                struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
 448{
 449        tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 450        throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 451                        rw == READ ? 'R' : 'W', tg->slice_start[rw],
 452                        tg->slice_end[rw], jiffies);
 453}
 454
 455/* Determine if previously allocated or extended slice is complete or not */
 456static bool
 457throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 458{
 459        if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 460                return 0;
 461
 462        return 1;
 463}
 464
 465/* Trim the used slices and adjust slice start accordingly */
 466static inline void
 467throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 468{
 469        unsigned long nr_slices, time_elapsed, io_trim;
 470        u64 bytes_trim, tmp;
 471
 472        BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 473
 474        /*
 475         * If bps are unlimited (-1), then time slice don't get
 476         * renewed. Don't try to trim the slice if slice is used. A new
 477         * slice will start when appropriate.
 478         */
 479        if (throtl_slice_used(td, tg, rw))
 480                return;
 481
 482        /*
 483         * A bio has been dispatched. Also adjust slice_end. It might happen
 484         * that initially cgroup limit was very low resulting in high
 485         * slice_end, but later limit was bumped up and bio was dispached
 486         * sooner, then we need to reduce slice_end. A high bogus slice_end
 487         * is bad because it does not allow new slice to start.
 488         */
 489
 490        throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
 491
 492        time_elapsed = jiffies - tg->slice_start[rw];
 493
 494        nr_slices = time_elapsed / throtl_slice;
 495
 496        if (!nr_slices)
 497                return;
 498        tmp = tg->bps[rw] * throtl_slice * nr_slices;
 499        do_div(tmp, HZ);
 500        bytes_trim = tmp;
 501
 502        io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
 503
 504        if (!bytes_trim && !io_trim)
 505                return;
 506
 507        if (tg->bytes_disp[rw] >= bytes_trim)
 508                tg->bytes_disp[rw] -= bytes_trim;
 509        else
 510                tg->bytes_disp[rw] = 0;
 511
 512        if (tg->io_disp[rw] >= io_trim)
 513                tg->io_disp[rw] -= io_trim;
 514        else
 515                tg->io_disp[rw] = 0;
 516
 517        tg->slice_start[rw] += nr_slices * throtl_slice;
 518
 519        throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
 520                        " start=%lu end=%lu jiffies=%lu",
 521                        rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 522                        tg->slice_start[rw], tg->slice_end[rw], jiffies);
 523}
 524
 525static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
 526                struct bio *bio, unsigned long *wait)
 527{
 528        bool rw = bio_data_dir(bio);
 529        unsigned int io_allowed;
 530        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 531        u64 tmp;
 532
 533        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 534
 535        /* Slice has just started. Consider one slice interval */
 536        if (!jiffy_elapsed)
 537                jiffy_elapsed_rnd = throtl_slice;
 538
 539        jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 540
 541        /*
 542         * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 543         * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 544         * will allow dispatch after 1 second and after that slice should
 545         * have been trimmed.
 546         */
 547
 548        tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
 549        do_div(tmp, HZ);
 550
 551        if (tmp > UINT_MAX)
 552                io_allowed = UINT_MAX;
 553        else
 554                io_allowed = tmp;
 555
 556        if (tg->io_disp[rw] + 1 <= io_allowed) {
 557                if (wait)
 558                        *wait = 0;
 559                return 1;
 560        }
 561
 562        /* Calc approx time to dispatch */
 563        jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
 564
 565        if (jiffy_wait > jiffy_elapsed)
 566                jiffy_wait = jiffy_wait - jiffy_elapsed;
 567        else
 568                jiffy_wait = 1;
 569
 570        if (wait)
 571                *wait = jiffy_wait;
 572        return 0;
 573}
 574
 575static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
 576                struct bio *bio, unsigned long *wait)
 577{
 578        bool rw = bio_data_dir(bio);
 579        u64 bytes_allowed, extra_bytes, tmp;
 580        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 581
 582        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 583
 584        /* Slice has just started. Consider one slice interval */
 585        if (!jiffy_elapsed)
 586                jiffy_elapsed_rnd = throtl_slice;
 587
 588        jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 589
 590        tmp = tg->bps[rw] * jiffy_elapsed_rnd;
 591        do_div(tmp, HZ);
 592        bytes_allowed = tmp;
 593
 594        if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
 595                if (wait)
 596                        *wait = 0;
 597                return 1;
 598        }
 599
 600        /* Calc approx time to dispatch */
 601        extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
 602        jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
 603
 604        if (!jiffy_wait)
 605                jiffy_wait = 1;
 606
 607        /*
 608         * This wait time is without taking into consideration the rounding
 609         * up we did. Add that time also.
 610         */
 611        jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 612        if (wait)
 613                *wait = jiffy_wait;
 614        return 0;
 615}
 616
 617static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
 618        if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
 619                return 1;
 620        return 0;
 621}
 622
 623/*
 624 * Returns whether one can dispatch a bio or not. Also returns approx number
 625 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 626 */
 627static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
 628                                struct bio *bio, unsigned long *wait)
 629{
 630        bool rw = bio_data_dir(bio);
 631        unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 632
 633        /*
 634         * Currently whole state machine of group depends on first bio
 635         * queued in the group bio list. So one should not be calling
 636         * this function with a different bio if there are other bios
 637         * queued.
 638         */
 639        BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
 640
 641        /* If tg->bps = -1, then BW is unlimited */
 642        if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
 643                if (wait)
 644                        *wait = 0;
 645                return 1;
 646        }
 647
 648        /*
 649         * If previous slice expired, start a new one otherwise renew/extend
 650         * existing slice to make sure it is at least throtl_slice interval
 651         * long since now.
 652         */
 653        if (throtl_slice_used(td, tg, rw))
 654                throtl_start_new_slice(td, tg, rw);
 655        else {
 656                if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
 657                        throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
 658        }
 659
 660        if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
 661            && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
 662                if (wait)
 663                        *wait = 0;
 664                return 1;
 665        }
 666
 667        max_wait = max(bps_wait, iops_wait);
 668
 669        if (wait)
 670                *wait = max_wait;
 671
 672        if (time_before(tg->slice_end[rw], jiffies + max_wait))
 673                throtl_extend_slice(td, tg, rw, jiffies + max_wait);
 674
 675        return 0;
 676}
 677
 678static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
 679                                         int rw)
 680{
 681        struct throtl_grp *tg = blkg_to_tg(blkg);
 682        struct tg_stats_cpu *stats_cpu;
 683        unsigned long flags;
 684
 685        /* If per cpu stats are not allocated yet, don't do any accounting. */
 686        if (tg->stats_cpu == NULL)
 687                return;
 688
 689        /*
 690         * Disabling interrupts to provide mutual exclusion between two
 691         * writes on same cpu. It probably is not needed for 64bit. Not
 692         * optimizing that case yet.
 693         */
 694        local_irq_save(flags);
 695
 696        stats_cpu = this_cpu_ptr(tg->stats_cpu);
 697
 698        blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
 699        blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
 700
 701        local_irq_restore(flags);
 702}
 703
 704static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 705{
 706        bool rw = bio_data_dir(bio);
 707
 708        /* Charge the bio to the group */
 709        tg->bytes_disp[rw] += bio->bi_size;
 710        tg->io_disp[rw]++;
 711
 712        throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size, bio->bi_rw);
 713}
 714
 715static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
 716                        struct bio *bio)
 717{
 718        bool rw = bio_data_dir(bio);
 719
 720        bio_list_add(&tg->bio_lists[rw], bio);
 721        /* Take a bio reference on tg */
 722        blkg_get(tg_to_blkg(tg));
 723        tg->nr_queued[rw]++;
 724        td->nr_queued[rw]++;
 725        throtl_enqueue_tg(td, tg);
 726}
 727
 728static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
 729{
 730        unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
 731        struct bio *bio;
 732
 733        if ((bio = bio_list_peek(&tg->bio_lists[READ])))
 734                tg_may_dispatch(td, tg, bio, &read_wait);
 735
 736        if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
 737                tg_may_dispatch(td, tg, bio, &write_wait);
 738
 739        min_wait = min(read_wait, write_wait);
 740        disptime = jiffies + min_wait;
 741
 742        /* Update dispatch time */
 743        throtl_dequeue_tg(td, tg);
 744        tg->disptime = disptime;
 745        throtl_enqueue_tg(td, tg);
 746}
 747
 748static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
 749                                bool rw, struct bio_list *bl)
 750{
 751        struct bio *bio;
 752
 753        bio = bio_list_pop(&tg->bio_lists[rw]);
 754        tg->nr_queued[rw]--;
 755        /* Drop bio reference on blkg */
 756        blkg_put(tg_to_blkg(tg));
 757
 758        BUG_ON(td->nr_queued[rw] <= 0);
 759        td->nr_queued[rw]--;
 760
 761        throtl_charge_bio(tg, bio);
 762        bio_list_add(bl, bio);
 763        bio->bi_rw |= REQ_THROTTLED;
 764
 765        throtl_trim_slice(td, tg, rw);
 766}
 767
 768static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
 769                                struct bio_list *bl)
 770{
 771        unsigned int nr_reads = 0, nr_writes = 0;
 772        unsigned int max_nr_reads = throtl_grp_quantum*3/4;
 773        unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
 774        struct bio *bio;
 775
 776        /* Try to dispatch 75% READS and 25% WRITES */
 777
 778        while ((bio = bio_list_peek(&tg->bio_lists[READ]))
 779                && tg_may_dispatch(td, tg, bio, NULL)) {
 780
 781                tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
 782                nr_reads++;
 783
 784                if (nr_reads >= max_nr_reads)
 785                        break;
 786        }
 787
 788        while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
 789                && tg_may_dispatch(td, tg, bio, NULL)) {
 790
 791                tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
 792                nr_writes++;
 793
 794                if (nr_writes >= max_nr_writes)
 795                        break;
 796        }
 797
 798        return nr_reads + nr_writes;
 799}
 800
 801static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
 802{
 803        unsigned int nr_disp = 0;
 804        struct throtl_grp *tg;
 805        struct throtl_rb_root *st = &td->tg_service_tree;
 806
 807        while (1) {
 808                tg = throtl_rb_first(st);
 809
 810                if (!tg)
 811                        break;
 812
 813                if (time_before(jiffies, tg->disptime))
 814                        break;
 815
 816                throtl_dequeue_tg(td, tg);
 817
 818                nr_disp += throtl_dispatch_tg(td, tg, bl);
 819
 820                if (tg->nr_queued[0] || tg->nr_queued[1]) {
 821                        tg_update_disptime(td, tg);
 822                        throtl_enqueue_tg(td, tg);
 823                }
 824
 825                if (nr_disp >= throtl_quantum)
 826                        break;
 827        }
 828
 829        return nr_disp;
 830}
 831
 832static void throtl_process_limit_change(struct throtl_data *td)
 833{
 834        struct request_queue *q = td->queue;
 835        struct blkcg_gq *blkg, *n;
 836
 837        if (!td->limits_changed)
 838                return;
 839
 840        xchg(&td->limits_changed, false);
 841
 842        throtl_log(td, "limits changed");
 843
 844        list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
 845                struct throtl_grp *tg = blkg_to_tg(blkg);
 846
 847                if (!tg->limits_changed)
 848                        continue;
 849
 850                if (!xchg(&tg->limits_changed, false))
 851                        continue;
 852
 853                throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
 854                        " riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
 855                        tg->iops[READ], tg->iops[WRITE]);
 856
 857                /*
 858                 * Restart the slices for both READ and WRITES. It
 859                 * might happen that a group's limit are dropped
 860                 * suddenly and we don't want to account recently
 861                 * dispatched IO with new low rate
 862                 */
 863                throtl_start_new_slice(td, tg, 0);
 864                throtl_start_new_slice(td, tg, 1);
 865
 866                if (throtl_tg_on_rr(tg))
 867                        tg_update_disptime(td, tg);
 868        }
 869}
 870
 871/* Dispatch throttled bios. Should be called without queue lock held. */
 872static int throtl_dispatch(struct request_queue *q)
 873{
 874        struct throtl_data *td = q->td;
 875        unsigned int nr_disp = 0;
 876        struct bio_list bio_list_on_stack;
 877        struct bio *bio;
 878        struct blk_plug plug;
 879
 880        spin_lock_irq(q->queue_lock);
 881
 882        throtl_process_limit_change(td);
 883
 884        if (!total_nr_queued(td))
 885                goto out;
 886
 887        bio_list_init(&bio_list_on_stack);
 888
 889        throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
 890                        total_nr_queued(td), td->nr_queued[READ],
 891                        td->nr_queued[WRITE]);
 892
 893        nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
 894
 895        if (nr_disp)
 896                throtl_log(td, "bios disp=%u", nr_disp);
 897
 898        throtl_schedule_next_dispatch(td);
 899out:
 900        spin_unlock_irq(q->queue_lock);
 901
 902        /*
 903         * If we dispatched some requests, unplug the queue to make sure
 904         * immediate dispatch
 905         */
 906        if (nr_disp) {
 907                blk_start_plug(&plug);
 908                while((bio = bio_list_pop(&bio_list_on_stack)))
 909                        generic_make_request(bio);
 910                blk_finish_plug(&plug);
 911        }
 912        return nr_disp;
 913}
 914
 915void blk_throtl_work(struct work_struct *work)
 916{
 917        struct throtl_data *td = container_of(work, struct throtl_data,
 918                                        throtl_work.work);
 919        struct request_queue *q = td->queue;
 920
 921        throtl_dispatch(q);
 922}
 923
 924/* Call with queue lock held */
 925static void
 926throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
 927{
 928
 929        struct delayed_work *dwork = &td->throtl_work;
 930
 931        /* schedule work if limits changed even if no bio is queued */
 932        if (total_nr_queued(td) || td->limits_changed) {
 933                /*
 934                 * We might have a work scheduled to be executed in future.
 935                 * Cancel that and schedule a new one.
 936                 */
 937                __cancel_delayed_work(dwork);
 938                queue_delayed_work(kthrotld_workqueue, dwork, delay);
 939                throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
 940                                delay, jiffies);
 941        }
 942}
 943
 944static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
 945                                struct blkg_policy_data *pd, int off)
 946{
 947        struct throtl_grp *tg = pd_to_tg(pd);
 948        struct blkg_rwstat rwstat = { }, tmp;
 949        int i, cpu;
 950
 951        for_each_possible_cpu(cpu) {
 952                struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
 953
 954                tmp = blkg_rwstat_read((void *)sc + off);
 955                for (i = 0; i < BLKG_RWSTAT_NR; i++)
 956                        rwstat.cnt[i] += tmp.cnt[i];
 957        }
 958
 959        return __blkg_prfill_rwstat(sf, pd, &rwstat);
 960}
 961
 962static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
 963                               struct seq_file *sf)
 964{
 965        struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
 966
 967        blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
 968                          cft->private, true);
 969        return 0;
 970}
 971
 972static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
 973                              int off)
 974{
 975        struct throtl_grp *tg = pd_to_tg(pd);
 976        u64 v = *(u64 *)((void *)tg + off);
 977
 978        if (v == -1)
 979                return 0;
 980        return __blkg_prfill_u64(sf, pd, v);
 981}
 982
 983static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
 984                               int off)
 985{
 986        struct throtl_grp *tg = pd_to_tg(pd);
 987        unsigned int v = *(unsigned int *)((void *)tg + off);
 988
 989        if (v == -1)
 990                return 0;
 991        return __blkg_prfill_u64(sf, pd, v);
 992}
 993
 994static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
 995                             struct seq_file *sf)
 996{
 997        blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
 998                          &blkcg_policy_throtl, cft->private, false);
 999        return 0;
1000}
1001
1002static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1003                              struct seq_file *sf)
1004{
1005        blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1006                          &blkcg_policy_throtl, cft->private, false);
1007        return 0;
1008}
1009
1010static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1011                       bool is_u64)
1012{
1013        struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1014        struct blkg_conf_ctx ctx;
1015        struct throtl_grp *tg;
1016        struct throtl_data *td;
1017        int ret;
1018
1019        ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1020        if (ret)
1021                return ret;
1022
1023        tg = blkg_to_tg(ctx.blkg);
1024        td = ctx.blkg->q->td;
1025
1026        if (!ctx.v)
1027                ctx.v = -1;
1028
1029        if (is_u64)
1030                *(u64 *)((void *)tg + cft->private) = ctx.v;
1031        else
1032                *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1033
1034        /* XXX: we don't need the following deferred processing */
1035        xchg(&tg->limits_changed, true);
1036        xchg(&td->limits_changed, true);
1037        throtl_schedule_delayed_work(td, 0);
1038
1039        blkg_conf_finish(&ctx);
1040        return 0;
1041}
1042
1043static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1044                           const char *buf)
1045{
1046        return tg_set_conf(cgrp, cft, buf, true);
1047}
1048
1049static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1050                            const char *buf)
1051{
1052        return tg_set_conf(cgrp, cft, buf, false);
1053}
1054
1055static struct cftype throtl_files[] = {
1056        {
1057                .name = "throttle.read_bps_device",
1058                .private = offsetof(struct throtl_grp, bps[READ]),
1059                .read_seq_string = tg_print_conf_u64,
1060                .write_string = tg_set_conf_u64,
1061                .max_write_len = 256,
1062        },
1063        {
1064                .name = "throttle.write_bps_device",
1065                .private = offsetof(struct throtl_grp, bps[WRITE]),
1066                .read_seq_string = tg_print_conf_u64,
1067                .write_string = tg_set_conf_u64,
1068                .max_write_len = 256,
1069        },
1070        {
1071                .name = "throttle.read_iops_device",
1072                .private = offsetof(struct throtl_grp, iops[READ]),
1073                .read_seq_string = tg_print_conf_uint,
1074                .write_string = tg_set_conf_uint,
1075                .max_write_len = 256,
1076        },
1077        {
1078                .name = "throttle.write_iops_device",
1079                .private = offsetof(struct throtl_grp, iops[WRITE]),
1080                .read_seq_string = tg_print_conf_uint,
1081                .write_string = tg_set_conf_uint,
1082                .max_write_len = 256,
1083        },
1084        {
1085                .name = "throttle.io_service_bytes",
1086                .private = offsetof(struct tg_stats_cpu, service_bytes),
1087                .read_seq_string = tg_print_cpu_rwstat,
1088        },
1089        {
1090                .name = "throttle.io_serviced",
1091                .private = offsetof(struct tg_stats_cpu, serviced),
1092                .read_seq_string = tg_print_cpu_rwstat,
1093        },
1094        { }     /* terminate */
1095};
1096
1097static void throtl_shutdown_wq(struct request_queue *q)
1098{
1099        struct throtl_data *td = q->td;
1100
1101        cancel_delayed_work_sync(&td->throtl_work);
1102}
1103
1104static struct blkcg_policy blkcg_policy_throtl = {
1105        .pd_size                = sizeof(struct throtl_grp),
1106        .cftypes                = throtl_files,
1107
1108        .pd_init_fn             = throtl_pd_init,
1109        .pd_exit_fn             = throtl_pd_exit,
1110        .pd_reset_stats_fn      = throtl_pd_reset_stats,
1111};
1112
1113bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1114{
1115        struct throtl_data *td = q->td;
1116        struct throtl_grp *tg;
1117        bool rw = bio_data_dir(bio), update_disptime = true;
1118        struct blkcg *blkcg;
1119        bool throttled = false;
1120
1121        if (bio->bi_rw & REQ_THROTTLED) {
1122                bio->bi_rw &= ~REQ_THROTTLED;
1123                goto out;
1124        }
1125
1126        /*
1127         * A throtl_grp pointer retrieved under rcu can be used to access
1128         * basic fields like stats and io rates. If a group has no rules,
1129         * just update the dispatch stats in lockless manner and return.
1130         */
1131        rcu_read_lock();
1132        blkcg = bio_blkcg(bio);
1133        tg = throtl_lookup_tg(td, blkcg);
1134        if (tg) {
1135                if (tg_no_rule_group(tg, rw)) {
1136                        throtl_update_dispatch_stats(tg_to_blkg(tg),
1137                                                     bio->bi_size, bio->bi_rw);
1138                        goto out_unlock_rcu;
1139                }
1140        }
1141
1142        /*
1143         * Either group has not been allocated yet or it is not an unlimited
1144         * IO group
1145         */
1146        spin_lock_irq(q->queue_lock);
1147        tg = throtl_lookup_create_tg(td, blkcg);
1148        if (unlikely(!tg))
1149                goto out_unlock;
1150
1151        if (tg->nr_queued[rw]) {
1152                /*
1153                 * There is already another bio queued in same dir. No
1154                 * need to update dispatch time.
1155                 */
1156                update_disptime = false;
1157                goto queue_bio;
1158
1159        }
1160
1161        /* Bio is with-in rate limit of group */
1162        if (tg_may_dispatch(td, tg, bio, NULL)) {
1163                throtl_charge_bio(tg, bio);
1164
1165                /*
1166                 * We need to trim slice even when bios are not being queued
1167                 * otherwise it might happen that a bio is not queued for
1168                 * a long time and slice keeps on extending and trim is not
1169                 * called for a long time. Now if limits are reduced suddenly
1170                 * we take into account all the IO dispatched so far at new
1171                 * low rate and * newly queued IO gets a really long dispatch
1172                 * time.
1173                 *
1174                 * So keep on trimming slice even if bio is not queued.
1175                 */
1176                throtl_trim_slice(td, tg, rw);
1177                goto out_unlock;
1178        }
1179
1180queue_bio:
1181        throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1182                        " iodisp=%u iops=%u queued=%d/%d",
1183                        rw == READ ? 'R' : 'W',
1184                        tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1185                        tg->io_disp[rw], tg->iops[rw],
1186                        tg->nr_queued[READ], tg->nr_queued[WRITE]);
1187
1188        bio_associate_current(bio);
1189        throtl_add_bio_tg(q->td, tg, bio);
1190        throttled = true;
1191
1192        if (update_disptime) {
1193                tg_update_disptime(td, tg);
1194                throtl_schedule_next_dispatch(td);
1195        }
1196
1197out_unlock:
1198        spin_unlock_irq(q->queue_lock);
1199out_unlock_rcu:
1200        rcu_read_unlock();
1201out:
1202        return throttled;
1203}
1204
1205/**
1206 * blk_throtl_drain - drain throttled bios
1207 * @q: request_queue to drain throttled bios for
1208 *
1209 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1210 */
1211void blk_throtl_drain(struct request_queue *q)
1212        __releases(q->queue_lock) __acquires(q->queue_lock)
1213{
1214        struct throtl_data *td = q->td;
1215        struct throtl_rb_root *st = &td->tg_service_tree;
1216        struct throtl_grp *tg;
1217        struct bio_list bl;
1218        struct bio *bio;
1219
1220        queue_lockdep_assert_held(q);
1221
1222        bio_list_init(&bl);
1223
1224        while ((tg = throtl_rb_first(st))) {
1225                throtl_dequeue_tg(td, tg);
1226
1227                while ((bio = bio_list_peek(&tg->bio_lists[READ])))
1228                        tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1229                while ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
1230                        tg_dispatch_one_bio(td, tg, bio_data_dir(bio), &bl);
1231        }
1232        spin_unlock_irq(q->queue_lock);
1233
1234        while ((bio = bio_list_pop(&bl)))
1235                generic_make_request(bio);
1236
1237        spin_lock_irq(q->queue_lock);
1238}
1239
1240int blk_throtl_init(struct request_queue *q)
1241{
1242        struct throtl_data *td;
1243        int ret;
1244
1245        td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1246        if (!td)
1247                return -ENOMEM;
1248
1249        td->tg_service_tree = THROTL_RB_ROOT;
1250        td->limits_changed = false;
1251        INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1252
1253        q->td = td;
1254        td->queue = q;
1255
1256        /* activate policy */
1257        ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1258        if (ret)
1259                kfree(td);
1260        return ret;
1261}
1262
1263void blk_throtl_exit(struct request_queue *q)
1264{
1265        BUG_ON(!q->td);
1266        throtl_shutdown_wq(q);
1267        blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1268        kfree(q->td);
1269}
1270
1271static int __init throtl_init(void)
1272{
1273        kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1274        if (!kthrotld_workqueue)
1275                panic("Failed to create kthrotld\n");
1276
1277        return blkcg_policy_register(&blkcg_policy_throtl);
1278}
1279
1280module_init(throtl_init);
1281
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.