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;
  28
  29/*
  30 * To implement hierarchical throttling, throtl_grps form a tree and bios
  31 * are dispatched upwards level by level until they reach the top and get
  32 * issued.  When dispatching bios from the children and local group at each
  33 * level, if the bios are dispatched into a single bio_list, there's a risk
  34 * of a local or child group which can queue many bios at once filling up
  35 * the list starving others.
  36 *
  37 * To avoid such starvation, dispatched bios are queued separately
  38 * according to where they came from.  When they are again dispatched to
  39 * the parent, they're popped in round-robin order so that no single source
  40 * hogs the dispatch window.
  41 *
  42 * throtl_qnode is used to keep the queued bios separated by their sources.
  43 * Bios are queued to throtl_qnode which in turn is queued to
  44 * throtl_service_queue and then dispatched in round-robin order.
  45 *
  46 * It's also used to track the reference counts on blkg's.  A qnode always
  47 * belongs to a throtl_grp and gets queued on itself or the parent, so
  48 * incrementing the reference of the associated throtl_grp when a qnode is
  49 * queued and decrementing when dequeued is enough to keep the whole blkg
  50 * tree pinned while bios are in flight.
  51 */
  52struct throtl_qnode {
  53        struct list_head        node;           /* service_queue->queued[] */
  54        struct bio_list         bios;           /* queued bios */
  55        struct throtl_grp       *tg;            /* tg this qnode belongs to */
  56};
  57
  58struct throtl_service_queue {
  59        struct throtl_service_queue *parent_sq; /* the parent service_queue */
  60
  61        /*
  62         * Bios queued directly to this service_queue or dispatched from
  63         * children throtl_grp's.
  64         */
  65        struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  66        unsigned int            nr_queued[2];   /* number of queued bios */
  67
  68        /*
  69         * RB tree of active children throtl_grp's, which are sorted by
  70         * their ->disptime.
  71         */
  72        struct rb_root          pending_tree;   /* RB tree of active tgs */
  73        struct rb_node          *first_pending; /* first node in the tree */
  74        unsigned int            nr_pending;     /* # queued in the tree */
  75        unsigned long           first_pending_disptime; /* disptime of the first tg */
  76        struct timer_list       pending_timer;  /* fires on first_pending_disptime */
  77};
  78
  79enum tg_state_flags {
  80        THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  81        THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  82};
  83
  84#define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  85
  86/* Per-cpu group stats */
  87struct tg_stats_cpu {
  88        /* total bytes transferred */
  89        struct blkg_rwstat              service_bytes;
  90        /* total IOs serviced, post merge */
  91        struct blkg_rwstat              serviced;
  92};
  93
  94struct throtl_grp {
  95        /* must be the first member */
  96        struct blkg_policy_data pd;
  97
  98        /* active throtl group service_queue member */
  99        struct rb_node rb_node;
 100
 101        /* throtl_data this group belongs to */
 102        struct throtl_data *td;
 103
 104        /* this group's service queue */
 105        struct throtl_service_queue service_queue;
 106
 107        /*
 108         * qnode_on_self is used when bios are directly queued to this
 109         * throtl_grp so that local bios compete fairly with bios
 110         * dispatched from children.  qnode_on_parent is used when bios are
 111         * dispatched from this throtl_grp into its parent and will compete
 112         * with the sibling qnode_on_parents and the parent's
 113         * qnode_on_self.
 114         */
 115        struct throtl_qnode qnode_on_self[2];
 116        struct throtl_qnode qnode_on_parent[2];
 117
 118        /*
 119         * Dispatch time in jiffies. This is the estimated time when group
 120         * will unthrottle and is ready to dispatch more bio. It is used as
 121         * key to sort active groups in service tree.
 122         */
 123        unsigned long disptime;
 124
 125        unsigned int flags;
 126
 127        /* are there any throtl rules between this group and td? */
 128        bool has_rules[2];
 129
 130        /* bytes per second rate limits */
 131        uint64_t bps[2];
 132
 133        /* IOPS limits */
 134        unsigned int iops[2];
 135
 136        /* Number of bytes disptached in current slice */
 137        uint64_t bytes_disp[2];
 138        /* Number of bio's dispatched in current slice */
 139        unsigned int io_disp[2];
 140
 141        /* When did we start a new slice */
 142        unsigned long slice_start[2];
 143        unsigned long slice_end[2];
 144
 145        /* Per cpu stats pointer */
 146        struct tg_stats_cpu __percpu *stats_cpu;
 147
 148        /* List of tgs waiting for per cpu stats memory to be allocated */
 149        struct list_head stats_alloc_node;
 150};
 151
 152struct throtl_data
 153{
 154        /* service tree for active throtl groups */
 155        struct throtl_service_queue service_queue;
 156
 157        struct request_queue *queue;
 158
 159        /* Total Number of queued bios on READ and WRITE lists */
 160        unsigned int nr_queued[2];
 161
 162        /*
 163         * number of total undestroyed groups
 164         */
 165        unsigned int nr_undestroyed_grps;
 166
 167        /* Work for dispatching throttled bios */
 168        struct work_struct dispatch_work;
 169};
 170
 171/* list and work item to allocate percpu group stats */
 172static DEFINE_SPINLOCK(tg_stats_alloc_lock);
 173static LIST_HEAD(tg_stats_alloc_list);
 174
 175static void tg_stats_alloc_fn(struct work_struct *);
 176static DECLARE_DELAYED_WORK(tg_stats_alloc_work, tg_stats_alloc_fn);
 177
 178static void throtl_pending_timer_fn(unsigned long arg);
 179
 180static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 181{
 182        return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 183}
 184
 185static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 186{
 187        return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 188}
 189
 190static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 191{
 192        return pd_to_blkg(&tg->pd);
 193}
 194
 195static inline struct throtl_grp *td_root_tg(struct throtl_data *td)
 196{
 197        return blkg_to_tg(td->queue->root_blkg);
 198}
 199
 200/**
 201 * sq_to_tg - return the throl_grp the specified service queue belongs to
 202 * @sq: the throtl_service_queue of interest
 203 *
 204 * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 205 * embedded in throtl_data, %NULL is returned.
 206 */
 207static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 208{
 209        if (sq && sq->parent_sq)
 210                return container_of(sq, struct throtl_grp, service_queue);
 211        else
 212                return NULL;
 213}
 214
 215/**
 216 * sq_to_td - return throtl_data the specified service queue belongs to
 217 * @sq: the throtl_service_queue of interest
 218 *
 219 * A service_queue can be embeded in either a throtl_grp or throtl_data.
 220 * Determine the associated throtl_data accordingly and return it.
 221 */
 222static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 223{
 224        struct throtl_grp *tg = sq_to_tg(sq);
 225
 226        if (tg)
 227                return tg->td;
 228        else
 229                return container_of(sq, struct throtl_data, service_queue);
 230}
 231
 232/**
 233 * throtl_log - log debug message via blktrace
 234 * @sq: the service_queue being reported
 235 * @fmt: printf format string
 236 * @args: printf args
 237 *
 238 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 239 * throtl_grp; otherwise, just "throtl".
 240 *
 241 * TODO: this should be made a function and name formatting should happen
 242 * after testing whether blktrace is enabled.
 243 */
 244#define throtl_log(sq, fmt, args...)    do {                            \
 245        struct throtl_grp *__tg = sq_to_tg((sq));                       \
 246        struct throtl_data *__td = sq_to_td((sq));                      \
 247                                                                        \
 248        (void)__td;                                                     \
 249        if ((__tg)) {                                                   \
 250                char __pbuf[128];                                       \
 251                                                                        \
 252                blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf));    \
 253                blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
 254        } else {                                                        \
 255                blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 256        }                                                               \
 257} while (0)
 258
 259/*
 260 * Worker for allocating per cpu stat for tgs. This is scheduled on the
 261 * system_wq once there are some groups on the alloc_list waiting for
 262 * allocation.
 263 */
 264static void tg_stats_alloc_fn(struct work_struct *work)
 265{
 266        static struct tg_stats_cpu *stats_cpu;  /* this fn is non-reentrant */
 267        struct delayed_work *dwork = to_delayed_work(work);
 268        bool empty = false;
 269
 270alloc_stats:
 271        if (!stats_cpu) {
 272                stats_cpu = alloc_percpu(struct tg_stats_cpu);
 273                if (!stats_cpu) {
 274                        /* allocation failed, try again after some time */
 275                        schedule_delayed_work(dwork, msecs_to_jiffies(10));
 276                        return;
 277                }
 278        }
 279
 280        spin_lock_irq(&tg_stats_alloc_lock);
 281
 282        if (!list_empty(&tg_stats_alloc_list)) {
 283                struct throtl_grp *tg = list_first_entry(&tg_stats_alloc_list,
 284                                                         struct throtl_grp,
 285                                                         stats_alloc_node);
 286                swap(tg->stats_cpu, stats_cpu);
 287                list_del_init(&tg->stats_alloc_node);
 288        }
 289
 290        empty = list_empty(&tg_stats_alloc_list);
 291        spin_unlock_irq(&tg_stats_alloc_lock);
 292        if (!empty)
 293                goto alloc_stats;
 294}
 295
 296static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 297{
 298        INIT_LIST_HEAD(&qn->node);
 299        bio_list_init(&qn->bios);
 300        qn->tg = tg;
 301}
 302
 303/**
 304 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 305 * @bio: bio being added
 306 * @qn: qnode to add bio to
 307 * @queued: the service_queue->queued[] list @qn belongs to
 308 *
 309 * Add @bio to @qn and put @qn on @queued if it's not already on.
 310 * @qn->tg's reference count is bumped when @qn is activated.  See the
 311 * comment on top of throtl_qnode definition for details.
 312 */
 313static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 314                                 struct list_head *queued)
 315{
 316        bio_list_add(&qn->bios, bio);
 317        if (list_empty(&qn->node)) {
 318                list_add_tail(&qn->node, queued);
 319                blkg_get(tg_to_blkg(qn->tg));
 320        }
 321}
 322
 323/**
 324 * throtl_peek_queued - peek the first bio on a qnode list
 325 * @queued: the qnode list to peek
 326 */
 327static struct bio *throtl_peek_queued(struct list_head *queued)
 328{
 329        struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 330        struct bio *bio;
 331
 332        if (list_empty(queued))
 333                return NULL;
 334
 335        bio = bio_list_peek(&qn->bios);
 336        WARN_ON_ONCE(!bio);
 337        return bio;
 338}
 339
 340/**
 341 * throtl_pop_queued - pop the first bio form a qnode list
 342 * @queued: the qnode list to pop a bio from
 343 * @tg_to_put: optional out argument for throtl_grp to put
 344 *
 345 * Pop the first bio from the qnode list @queued.  After popping, the first
 346 * qnode is removed from @queued if empty or moved to the end of @queued so
 347 * that the popping order is round-robin.
 348 *
 349 * When the first qnode is removed, its associated throtl_grp should be put
 350 * too.  If @tg_to_put is NULL, this function automatically puts it;
 351 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 352 * responsible for putting it.
 353 */
 354static struct bio *throtl_pop_queued(struct list_head *queued,
 355                                     struct throtl_grp **tg_to_put)
 356{
 357        struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 358        struct bio *bio;
 359
 360        if (list_empty(queued))
 361                return NULL;
 362
 363        bio = bio_list_pop(&qn->bios);
 364        WARN_ON_ONCE(!bio);
 365
 366        if (bio_list_empty(&qn->bios)) {
 367                list_del_init(&qn->node);
 368                if (tg_to_put)
 369                        *tg_to_put = qn->tg;
 370                else
 371                        blkg_put(tg_to_blkg(qn->tg));
 372        } else {
 373                list_move_tail(&qn->node, queued);
 374        }
 375
 376        return bio;
 377}
 378
 379/* init a service_queue, assumes the caller zeroed it */
 380static void throtl_service_queue_init(struct throtl_service_queue *sq,
 381                                      struct throtl_service_queue *parent_sq)
 382{
 383        INIT_LIST_HEAD(&sq->queued[0]);
 384        INIT_LIST_HEAD(&sq->queued[1]);
 385        sq->pending_tree = RB_ROOT;
 386        sq->parent_sq = parent_sq;
 387        setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
 388                    (unsigned long)sq);
 389}
 390
 391static void throtl_service_queue_exit(struct throtl_service_queue *sq)
 392{
 393        del_timer_sync(&sq->pending_timer);
 394}
 395
 396static void throtl_pd_init(struct blkcg_gq *blkg)
 397{
 398        struct throtl_grp *tg = blkg_to_tg(blkg);
 399        struct throtl_data *td = blkg->q->td;
 400        struct throtl_service_queue *parent_sq;
 401        unsigned long flags;
 402        int rw;
 403
 404        /*
 405         * If sane_hierarchy is enabled, we switch to properly hierarchical
 406         * behavior where limits on a given throtl_grp are applied to the
 407         * whole subtree rather than just the group itself.  e.g. If 16M
 408         * read_bps limit is set on the root group, the whole system can't
 409         * exceed 16M for the device.
 410         *
 411         * If sane_hierarchy is not enabled, the broken flat hierarchy
 412         * behavior is retained where all throtl_grps are treated as if
 413         * they're all separate root groups right below throtl_data.
 414         * Limits of a group don't interact with limits of other groups
 415         * regardless of the position of the group in the hierarchy.
 416         */
 417        parent_sq = &td->service_queue;
 418
 419        if (cgroup_sane_behavior(blkg->blkcg->css.cgroup) && blkg->parent)
 420                parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 421
 422        throtl_service_queue_init(&tg->service_queue, parent_sq);
 423
 424        for (rw = READ; rw <= WRITE; rw++) {
 425                throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 426                throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 427        }
 428
 429        RB_CLEAR_NODE(&tg->rb_node);
 430        tg->td = td;
 431
 432        tg->bps[READ] = -1;
 433        tg->bps[WRITE] = -1;
 434        tg->iops[READ] = -1;
 435        tg->iops[WRITE] = -1;
 436
 437        /*
 438         * Ugh... We need to perform per-cpu allocation for tg->stats_cpu
 439         * but percpu allocator can't be called from IO path.  Queue tg on
 440         * tg_stats_alloc_list and allocate from work item.
 441         */
 442        spin_lock_irqsave(&tg_stats_alloc_lock, flags);
 443        list_add(&tg->stats_alloc_node, &tg_stats_alloc_list);
 444        schedule_delayed_work(&tg_stats_alloc_work, 0);
 445        spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
 446}
 447
 448/*
 449 * Set has_rules[] if @tg or any of its parents have limits configured.
 450 * This doesn't require walking up to the top of the hierarchy as the
 451 * parent's has_rules[] is guaranteed to be correct.
 452 */
 453static void tg_update_has_rules(struct throtl_grp *tg)
 454{
 455        struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 456        int rw;
 457
 458        for (rw = READ; rw <= WRITE; rw++)
 459                tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 460                                    (tg->bps[rw] != -1 || tg->iops[rw] != -1);
 461}
 462
 463static void throtl_pd_online(struct blkcg_gq *blkg)
 464{
 465        /*
 466         * We don't want new groups to escape the limits of its ancestors.
 467         * Update has_rules[] after a new group is brought online.
 468         */
 469        tg_update_has_rules(blkg_to_tg(blkg));
 470}
 471
 472static void throtl_pd_exit(struct blkcg_gq *blkg)
 473{
 474        struct throtl_grp *tg = blkg_to_tg(blkg);
 475        unsigned long flags;
 476
 477        spin_lock_irqsave(&tg_stats_alloc_lock, flags);
 478        list_del_init(&tg->stats_alloc_node);
 479        spin_unlock_irqrestore(&tg_stats_alloc_lock, flags);
 480
 481        free_percpu(tg->stats_cpu);
 482
 483        throtl_service_queue_exit(&tg->service_queue);
 484}
 485
 486static void throtl_pd_reset_stats(struct blkcg_gq *blkg)
 487{
 488        struct throtl_grp *tg = blkg_to_tg(blkg);
 489        int cpu;
 490
 491        if (tg->stats_cpu == NULL)
 492                return;
 493
 494        for_each_possible_cpu(cpu) {
 495                struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
 496
 497                blkg_rwstat_reset(&sc->service_bytes);
 498                blkg_rwstat_reset(&sc->serviced);
 499        }
 500}
 501
 502static struct throtl_grp *throtl_lookup_tg(struct throtl_data *td,
 503                                           struct blkcg *blkcg)
 504{
 505        /*
 506         * This is the common case when there are no blkcgs.  Avoid lookup
 507         * in this case
 508         */
 509        if (blkcg == &blkcg_root)
 510                return td_root_tg(td);
 511
 512        return blkg_to_tg(blkg_lookup(blkcg, td->queue));
 513}
 514
 515static struct throtl_grp *throtl_lookup_create_tg(struct throtl_data *td,
 516                                                  struct blkcg *blkcg)
 517{
 518        struct request_queue *q = td->queue;
 519        struct throtl_grp *tg = NULL;
 520
 521        /*
 522         * This is the common case when there are no blkcgs.  Avoid lookup
 523         * in this case
 524         */
 525        if (blkcg == &blkcg_root) {
 526                tg = td_root_tg(td);
 527        } else {
 528                struct blkcg_gq *blkg;
 529
 530                blkg = blkg_lookup_create(blkcg, q);
 531
 532                /* if %NULL and @q is alive, fall back to root_tg */
 533                if (!IS_ERR(blkg))
 534                        tg = blkg_to_tg(blkg);
 535                else if (!blk_queue_dying(q))
 536                        tg = td_root_tg(td);
 537        }
 538
 539        return tg;
 540}
 541
 542static struct throtl_grp *
 543throtl_rb_first(struct throtl_service_queue *parent_sq)
 544{
 545        /* Service tree is empty */
 546        if (!parent_sq->nr_pending)
 547                return NULL;
 548
 549        if (!parent_sq->first_pending)
 550                parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
 551
 552        if (parent_sq->first_pending)
 553                return rb_entry_tg(parent_sq->first_pending);
 554
 555        return NULL;
 556}
 557
 558static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 559{
 560        rb_erase(n, root);
 561        RB_CLEAR_NODE(n);
 562}
 563
 564static void throtl_rb_erase(struct rb_node *n,
 565                            struct throtl_service_queue *parent_sq)
 566{
 567        if (parent_sq->first_pending == n)
 568                parent_sq->first_pending = NULL;
 569        rb_erase_init(n, &parent_sq->pending_tree);
 570        --parent_sq->nr_pending;
 571}
 572
 573static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 574{
 575        struct throtl_grp *tg;
 576
 577        tg = throtl_rb_first(parent_sq);
 578        if (!tg)
 579                return;
 580
 581        parent_sq->first_pending_disptime = tg->disptime;
 582}
 583
 584static void tg_service_queue_add(struct throtl_grp *tg)
 585{
 586        struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 587        struct rb_node **node = &parent_sq->pending_tree.rb_node;
 588        struct rb_node *parent = NULL;
 589        struct throtl_grp *__tg;
 590        unsigned long key = tg->disptime;
 591        int left = 1;
 592
 593        while (*node != NULL) {
 594                parent = *node;
 595                __tg = rb_entry_tg(parent);
 596
 597                if (time_before(key, __tg->disptime))
 598                        node = &parent->rb_left;
 599                else {
 600                        node = &parent->rb_right;
 601                        left = 0;
 602                }
 603        }
 604
 605        if (left)
 606                parent_sq->first_pending = &tg->rb_node;
 607
 608        rb_link_node(&tg->rb_node, parent, node);
 609        rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
 610}
 611
 612static void __throtl_enqueue_tg(struct throtl_grp *tg)
 613{
 614        tg_service_queue_add(tg);
 615        tg->flags |= THROTL_TG_PENDING;
 616        tg->service_queue.parent_sq->nr_pending++;
 617}
 618
 619static void throtl_enqueue_tg(struct throtl_grp *tg)
 620{
 621        if (!(tg->flags & THROTL_TG_PENDING))
 622                __throtl_enqueue_tg(tg);
 623}
 624
 625static void __throtl_dequeue_tg(struct throtl_grp *tg)
 626{
 627        throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 628        tg->flags &= ~THROTL_TG_PENDING;
 629}
 630
 631static void throtl_dequeue_tg(struct throtl_grp *tg)
 632{
 633        if (tg->flags & THROTL_TG_PENDING)
 634                __throtl_dequeue_tg(tg);
 635}
 636
 637/* Call with queue lock held */
 638static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 639                                          unsigned long expires)
 640{
 641        mod_timer(&sq->pending_timer, expires);
 642        throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 643                   expires - jiffies, jiffies);
 644}
 645
 646/**
 647 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 648 * @sq: the service_queue to schedule dispatch for
 649 * @force: force scheduling
 650 *
 651 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 652 * dispatch time of the first pending child.  Returns %true if either timer
 653 * is armed or there's no pending child left.  %false if the current
 654 * dispatch window is still open and the caller should continue
 655 * dispatching.
 656 *
 657 * If @force is %true, the dispatch timer is always scheduled and this
 658 * function is guaranteed to return %true.  This is to be used when the
 659 * caller can't dispatch itself and needs to invoke pending_timer
 660 * unconditionally.  Note that forced scheduling is likely to induce short
 661 * delay before dispatch starts even if @sq->first_pending_disptime is not
 662 * in the future and thus shouldn't be used in hot paths.
 663 */
 664static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 665                                          bool force)
 666{
 667        /* any pending children left? */
 668        if (!sq->nr_pending)
 669                return true;
 670
 671        update_min_dispatch_time(sq);
 672
 673        /* is the next dispatch time in the future? */
 674        if (force || time_after(sq->first_pending_disptime, jiffies)) {
 675                throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 676                return true;
 677        }
 678
 679        /* tell the caller to continue dispatching */
 680        return false;
 681}
 682
 683static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 684                bool rw, unsigned long start)
 685{
 686        tg->bytes_disp[rw] = 0;
 687        tg->io_disp[rw] = 0;
 688
 689        /*
 690         * Previous slice has expired. We must have trimmed it after last
 691         * bio dispatch. That means since start of last slice, we never used
 692         * that bandwidth. Do try to make use of that bandwidth while giving
 693         * credit.
 694         */
 695        if (time_after_eq(start, tg->slice_start[rw]))
 696                tg->slice_start[rw] = start;
 697
 698        tg->slice_end[rw] = jiffies + throtl_slice;
 699        throtl_log(&tg->service_queue,
 700                   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 701                   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 702                   tg->slice_end[rw], jiffies);
 703}
 704
 705static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 706{
 707        tg->bytes_disp[rw] = 0;
 708        tg->io_disp[rw] = 0;
 709        tg->slice_start[rw] = jiffies;
 710        tg->slice_end[rw] = jiffies + throtl_slice;
 711        throtl_log(&tg->service_queue,
 712                   "[%c] new slice start=%lu end=%lu jiffies=%lu",
 713                   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 714                   tg->slice_end[rw], jiffies);
 715}
 716
 717static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 718                                        unsigned long jiffy_end)
 719{
 720        tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 721}
 722
 723static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 724                                       unsigned long jiffy_end)
 725{
 726        tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 727        throtl_log(&tg->service_queue,
 728                   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 729                   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 730                   tg->slice_end[rw], jiffies);
 731}
 732
 733/* Determine if previously allocated or extended slice is complete or not */
 734static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 735{
 736        if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 737                return 0;
 738
 739        return 1;
 740}
 741
 742/* Trim the used slices and adjust slice start accordingly */
 743static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 744{
 745        unsigned long nr_slices, time_elapsed, io_trim;
 746        u64 bytes_trim, tmp;
 747
 748        BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 749
 750        /*
 751         * If bps are unlimited (-1), then time slice don't get
 752         * renewed. Don't try to trim the slice if slice is used. A new
 753         * slice will start when appropriate.
 754         */
 755        if (throtl_slice_used(tg, rw))
 756                return;
 757
 758        /*
 759         * A bio has been dispatched. Also adjust slice_end. It might happen
 760         * that initially cgroup limit was very low resulting in high
 761         * slice_end, but later limit was bumped up and bio was dispached
 762         * sooner, then we need to reduce slice_end. A high bogus slice_end
 763         * is bad because it does not allow new slice to start.
 764         */
 765
 766        throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
 767
 768        time_elapsed = jiffies - tg->slice_start[rw];
 769
 770        nr_slices = time_elapsed / throtl_slice;
 771
 772        if (!nr_slices)
 773                return;
 774        tmp = tg->bps[rw] * throtl_slice * nr_slices;
 775        do_div(tmp, HZ);
 776        bytes_trim = tmp;
 777
 778        io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
 779
 780        if (!bytes_trim && !io_trim)
 781                return;
 782
 783        if (tg->bytes_disp[rw] >= bytes_trim)
 784                tg->bytes_disp[rw] -= bytes_trim;
 785        else
 786                tg->bytes_disp[rw] = 0;
 787
 788        if (tg->io_disp[rw] >= io_trim)
 789                tg->io_disp[rw] -= io_trim;
 790        else
 791                tg->io_disp[rw] = 0;
 792
 793        tg->slice_start[rw] += nr_slices * throtl_slice;
 794
 795        throtl_log(&tg->service_queue,
 796                   "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 797                   rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 798                   tg->slice_start[rw], tg->slice_end[rw], jiffies);
 799}
 800
 801static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 802                                  unsigned long *wait)
 803{
 804        bool rw = bio_data_dir(bio);
 805        unsigned int io_allowed;
 806        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 807        u64 tmp;
 808
 809        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 810
 811        /* Slice has just started. Consider one slice interval */
 812        if (!jiffy_elapsed)
 813                jiffy_elapsed_rnd = throtl_slice;
 814
 815        jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 816
 817        /*
 818         * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 819         * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 820         * will allow dispatch after 1 second and after that slice should
 821         * have been trimmed.
 822         */
 823
 824        tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
 825        do_div(tmp, HZ);
 826
 827        if (tmp > UINT_MAX)
 828                io_allowed = UINT_MAX;
 829        else
 830                io_allowed = tmp;
 831
 832        if (tg->io_disp[rw] + 1 <= io_allowed) {
 833                if (wait)
 834                        *wait = 0;
 835                return 1;
 836        }
 837
 838        /* Calc approx time to dispatch */
 839        jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
 840
 841        if (jiffy_wait > jiffy_elapsed)
 842                jiffy_wait = jiffy_wait - jiffy_elapsed;
 843        else
 844                jiffy_wait = 1;
 845
 846        if (wait)
 847                *wait = jiffy_wait;
 848        return 0;
 849}
 850
 851static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 852                                 unsigned long *wait)
 853{
 854        bool rw = bio_data_dir(bio);
 855        u64 bytes_allowed, extra_bytes, tmp;
 856        unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 857
 858        jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 859
 860        /* Slice has just started. Consider one slice interval */
 861        if (!jiffy_elapsed)
 862                jiffy_elapsed_rnd = throtl_slice;
 863
 864        jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 865
 866        tmp = tg->bps[rw] * jiffy_elapsed_rnd;
 867        do_div(tmp, HZ);
 868        bytes_allowed = tmp;
 869
 870        if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
 871                if (wait)
 872                        *wait = 0;
 873                return 1;
 874        }
 875
 876        /* Calc approx time to dispatch */
 877        extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
 878        jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
 879
 880        if (!jiffy_wait)
 881                jiffy_wait = 1;
 882
 883        /*
 884         * This wait time is without taking into consideration the rounding
 885         * up we did. Add that time also.
 886         */
 887        jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 888        if (wait)
 889                *wait = jiffy_wait;
 890        return 0;
 891}
 892
 893/*
 894 * Returns whether one can dispatch a bio or not. Also returns approx number
 895 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 896 */
 897static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 898                            unsigned long *wait)
 899{
 900        bool rw = bio_data_dir(bio);
 901        unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 902
 903        /*
 904         * Currently whole state machine of group depends on first bio
 905         * queued in the group bio list. So one should not be calling
 906         * this function with a different bio if there are other bios
 907         * queued.
 908         */
 909        BUG_ON(tg->service_queue.nr_queued[rw] &&
 910               bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 911
 912        /* If tg->bps = -1, then BW is unlimited */
 913        if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
 914                if (wait)
 915                        *wait = 0;
 916                return 1;
 917        }
 918
 919        /*
 920         * If previous slice expired, start a new one otherwise renew/extend
 921         * existing slice to make sure it is at least throtl_slice interval
 922         * long since now.
 923         */
 924        if (throtl_slice_used(tg, rw))
 925                throtl_start_new_slice(tg, rw);
 926        else {
 927                if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
 928                        throtl_extend_slice(tg, rw, jiffies + throtl_slice);
 929        }
 930
 931        if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
 932            tg_with_in_iops_limit(tg, bio, &iops_wait)) {
 933                if (wait)
 934                        *wait = 0;
 935                return 1;
 936        }
 937
 938        max_wait = max(bps_wait, iops_wait);
 939
 940        if (wait)
 941                *wait = max_wait;
 942
 943        if (time_before(tg->slice_end[rw], jiffies + max_wait))
 944                throtl_extend_slice(tg, rw, jiffies + max_wait);
 945
 946        return 0;
 947}
 948
 949static void throtl_update_dispatch_stats(struct blkcg_gq *blkg, u64 bytes,
 950                                         int rw)
 951{
 952        struct throtl_grp *tg = blkg_to_tg(blkg);
 953        struct tg_stats_cpu *stats_cpu;
 954        unsigned long flags;
 955
 956        /* If per cpu stats are not allocated yet, don't do any accounting. */
 957        if (tg->stats_cpu == NULL)
 958                return;
 959
 960        /*
 961         * Disabling interrupts to provide mutual exclusion between two
 962         * writes on same cpu. It probably is not needed for 64bit. Not
 963         * optimizing that case yet.
 964         */
 965        local_irq_save(flags);
 966
 967        stats_cpu = this_cpu_ptr(tg->stats_cpu);
 968
 969        blkg_rwstat_add(&stats_cpu->serviced, rw, 1);
 970        blkg_rwstat_add(&stats_cpu->service_bytes, rw, bytes);
 971
 972        local_irq_restore(flags);
 973}
 974
 975static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 976{
 977        bool rw = bio_data_dir(bio);
 978
 979        /* Charge the bio to the group */
 980        tg->bytes_disp[rw] += bio->bi_size;
 981        tg->io_disp[rw]++;
 982
 983        /*
 984         * REQ_THROTTLED is used to prevent the same bio to be throttled
 985         * more than once as a throttled bio will go through blk-throtl the
 986         * second time when it eventually gets issued.  Set it when a bio
 987         * is being charged to a tg.
 988         *
 989         * Dispatch stats aren't recursive and each @bio should only be
 990         * accounted by the @tg it was originally associated with.  Let's
 991         * update the stats when setting REQ_THROTTLED for the first time
 992         * which is guaranteed to be for the @bio's original tg.
 993         */
 994        if (!(bio->bi_rw & REQ_THROTTLED)) {
 995                bio->bi_rw |= REQ_THROTTLED;
 996                throtl_update_dispatch_stats(tg_to_blkg(tg), bio->bi_size,
 997                                             bio->bi_rw);
 998        }
 999}
1000
1001/**
1002 * throtl_add_bio_tg - add a bio to the specified throtl_grp
1003 * @bio: bio to add
1004 * @qn: qnode to use
1005 * @tg: the target throtl_grp
1006 *
1007 * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1008 * tg->qnode_on_self[] is used.
1009 */
1010static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1011                              struct throtl_grp *tg)
1012{
1013        struct throtl_service_queue *sq = &tg->service_queue;
1014        bool rw = bio_data_dir(bio);
1015
1016        if (!qn)
1017                qn = &tg->qnode_on_self[rw];
1018
1019        /*
1020         * If @tg doesn't currently have any bios queued in the same
1021         * direction, queueing @bio can change when @tg should be
1022         * dispatched.  Mark that @tg was empty.  This is automatically
1023         * cleaered on the next tg_update_disptime().
1024         */
1025        if (!sq->nr_queued[rw])
1026                tg->flags |= THROTL_TG_WAS_EMPTY;
1027
1028        throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1029
1030        sq->nr_queued[rw]++;
1031        throtl_enqueue_tg(tg);
1032}
1033
1034static void tg_update_disptime(struct throtl_grp *tg)
1035{
1036        struct throtl_service_queue *sq = &tg->service_queue;
1037        unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1038        struct bio *bio;
1039
1040        if ((bio = throtl_peek_queued(&sq->queued[READ])))
1041                tg_may_dispatch(tg, bio, &read_wait);
1042
1043        if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1044                tg_may_dispatch(tg, bio, &write_wait);
1045
1046        min_wait = min(read_wait, write_wait);
1047        disptime = jiffies + min_wait;
1048
1049        /* Update dispatch time */
1050        throtl_dequeue_tg(tg);
1051        tg->disptime = disptime;
1052        throtl_enqueue_tg(tg);
1053
1054        /* see throtl_add_bio_tg() */
1055        tg->flags &= ~THROTL_TG_WAS_EMPTY;
1056}
1057
1058static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1059                                        struct throtl_grp *parent_tg, bool rw)
1060{
1061        if (throtl_slice_used(parent_tg, rw)) {
1062                throtl_start_new_slice_with_credit(parent_tg, rw,
1063                                child_tg->slice_start[rw]);
1064        }
1065
1066}
1067
1068static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1069{
1070        struct throtl_service_queue *sq = &tg->service_queue;
1071        struct throtl_service_queue *parent_sq = sq->parent_sq;
1072        struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1073        struct throtl_grp *tg_to_put = NULL;
1074        struct bio *bio;
1075
1076        /*
1077         * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1078         * from @tg may put its reference and @parent_sq might end up
1079         * getting released prematurely.  Remember the tg to put and put it
1080         * after @bio is transferred to @parent_sq.
1081         */
1082        bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1083        sq->nr_queued[rw]--;
1084
1085        throtl_charge_bio(tg, bio);
1086
1087        /*
1088         * If our parent is another tg, we just need to transfer @bio to
1089         * the parent using throtl_add_bio_tg().  If our parent is
1090         * @td->service_queue, @bio is ready to be issued.  Put it on its
1091         * bio_lists[] and decrease total number queued.  The caller is
1092         * responsible for issuing these bios.
1093         */
1094        if (parent_tg) {
1095                throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1096                start_parent_slice_with_credit(tg, parent_tg, rw);
1097        } else {
1098                throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1099                                     &parent_sq->queued[rw]);
1100                BUG_ON(tg->td->nr_queued[rw] <= 0);
1101                tg->td->nr_queued[rw]--;
1102        }
1103
1104        throtl_trim_slice(tg, rw);
1105
1106        if (tg_to_put)
1107                blkg_put(tg_to_blkg(tg_to_put));
1108}
1109
1110static int throtl_dispatch_tg(struct throtl_grp *tg)
1111{
1112        struct throtl_service_queue *sq = &tg->service_queue;
1113        unsigned int nr_reads = 0, nr_writes = 0;
1114        unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1115        unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1116        struct bio *bio;
1117
1118        /* Try to dispatch 75% READS and 25% WRITES */
1119
1120        while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1121               tg_may_dispatch(tg, bio, NULL)) {
1122
1123                tg_dispatch_one_bio(tg, bio_data_dir(bio));
1124                nr_reads++;
1125
1126                if (nr_reads >= max_nr_reads)
1127                        break;
1128        }
1129
1130        while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1131               tg_may_dispatch(tg, bio, NULL)) {
1132
1133                tg_dispatch_one_bio(tg, bio_data_dir(bio));
1134                nr_writes++;
1135
1136                if (nr_writes >= max_nr_writes)
1137                        break;
1138        }
1139
1140        return nr_reads + nr_writes;
1141}
1142
1143static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1144{
1145        unsigned int nr_disp = 0;
1146
1147        while (1) {
1148                struct throtl_grp *tg = throtl_rb_first(parent_sq);
1149                struct throtl_service_queue *sq = &tg->service_queue;
1150
1151                if (!tg)
1152                        break;
1153
1154                if (time_before(jiffies, tg->disptime))
1155                        break;
1156
1157                throtl_dequeue_tg(tg);
1158
1159                nr_disp += throtl_dispatch_tg(tg);
1160
1161                if (sq->nr_queued[0] || sq->nr_queued[1])
1162                        tg_update_disptime(tg);
1163
1164                if (nr_disp >= throtl_quantum)
1165                        break;
1166        }
1167
1168        return nr_disp;
1169}
1170
1171/**
1172 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1173 * @arg: the throtl_service_queue being serviced
1174 *
1175 * This timer is armed when a child throtl_grp with active bio's become
1176 * pending and queued on the service_queue's pending_tree and expires when
1177 * the first child throtl_grp should be dispatched.  This function
1178 * dispatches bio's from the children throtl_grps to the parent
1179 * service_queue.
1180 *
1181 * If the parent's parent is another throtl_grp, dispatching is propagated
1182 * by either arming its pending_timer or repeating dispatch directly.  If
1183 * the top-level service_tree is reached, throtl_data->dispatch_work is
1184 * kicked so that the ready bio's are issued.
1185 */
1186static void throtl_pending_timer_fn(unsigned long arg)
1187{
1188        struct throtl_service_queue *sq = (void *)arg;
1189        struct throtl_grp *tg = sq_to_tg(sq);
1190        struct throtl_data *td = sq_to_td(sq);
1191        struct request_queue *q = td->queue;
1192        struct throtl_service_queue *parent_sq;
1193        bool dispatched;
1194        int ret;
1195
1196        spin_lock_irq(q->queue_lock);
1197again:
1198        parent_sq = sq->parent_sq;
1199        dispatched = false;
1200
1201        while (true) {
1202                throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1203                           sq->nr_queued[READ] + sq->nr_queued[WRITE],
1204                           sq->nr_queued[READ], sq->nr_queued[WRITE]);
1205
1206                ret = throtl_select_dispatch(sq);
1207                if (ret) {
1208                        throtl_log(sq, "bios disp=%u", ret);
1209                        dispatched = true;
1210                }
1211
1212                if (throtl_schedule_next_dispatch(sq, false))
1213                        break;
1214
1215                /* this dispatch windows is still open, relax and repeat */
1216                spin_unlock_irq(q->queue_lock);
1217                cpu_relax();
1218                spin_lock_irq(q->queue_lock);
1219        }
1220
1221        if (!dispatched)
1222                goto out_unlock;
1223
1224        if (parent_sq) {
1225                /* @parent_sq is another throl_grp, propagate dispatch */
1226                if (tg->flags & THROTL_TG_WAS_EMPTY) {
1227                        tg_update_disptime(tg);
1228                        if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1229                                /* window is already open, repeat dispatching */
1230                                sq = parent_sq;
1231                                tg = sq_to_tg(sq);
1232                                goto again;
1233                        }
1234                }
1235        } else {
1236                /* reached the top-level, queue issueing */
1237                queue_work(kthrotld_workqueue, &td->dispatch_work);
1238        }
1239out_unlock:
1240        spin_unlock_irq(q->queue_lock);
1241}
1242
1243/**
1244 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1245 * @work: work item being executed
1246 *
1247 * This function is queued for execution when bio's reach the bio_lists[]
1248 * of throtl_data->service_queue.  Those bio's are ready and issued by this
1249 * function.
1250 */
1251void blk_throtl_dispatch_work_fn(struct work_struct *work)
1252{
1253        struct throtl_data *td = container_of(work, struct throtl_data,
1254                                              dispatch_work);
1255        struct throtl_service_queue *td_sq = &td->service_queue;
1256        struct request_queue *q = td->queue;
1257        struct bio_list bio_list_on_stack;
1258        struct bio *bio;
1259        struct blk_plug plug;
1260        int rw;
1261
1262        bio_list_init(&bio_list_on_stack);
1263
1264        spin_lock_irq(q->queue_lock);
1265        for (rw = READ; rw <= WRITE; rw++)
1266                while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1267                        bio_list_add(&bio_list_on_stack, bio);
1268        spin_unlock_irq(q->queue_lock);
1269
1270        if (!bio_list_empty(&bio_list_on_stack)) {
1271                blk_start_plug(&plug);
1272                while((bio = bio_list_pop(&bio_list_on_stack)))
1273                        generic_make_request(bio);
1274                blk_finish_plug(&plug);
1275        }
1276}
1277
1278static u64 tg_prfill_cpu_rwstat(struct seq_file *sf,
1279                                struct blkg_policy_data *pd, int off)
1280{
1281        struct throtl_grp *tg = pd_to_tg(pd);
1282        struct blkg_rwstat rwstat = { }, tmp;
1283        int i, cpu;
1284
1285        for_each_possible_cpu(cpu) {
1286                struct tg_stats_cpu *sc = per_cpu_ptr(tg->stats_cpu, cpu);
1287
1288                tmp = blkg_rwstat_read((void *)sc + off);
1289                for (i = 0; i < BLKG_RWSTAT_NR; i++)
1290                        rwstat.cnt[i] += tmp.cnt[i];
1291        }
1292
1293        return __blkg_prfill_rwstat(sf, pd, &rwstat);
1294}
1295
1296static int tg_print_cpu_rwstat(struct cgroup *cgrp, struct cftype *cft,
1297                               struct seq_file *sf)
1298{
1299        struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1300
1301        blkcg_print_blkgs(sf, blkcg, tg_prfill_cpu_rwstat, &blkcg_policy_throtl,
1302                          cft->private, true);
1303        return 0;
1304}
1305
1306static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1307                              int off)
1308{
1309        struct throtl_grp *tg = pd_to_tg(pd);
1310        u64 v = *(u64 *)((void *)tg + off);
1311
1312        if (v == -1)
1313                return 0;
1314        return __blkg_prfill_u64(sf, pd, v);
1315}
1316
1317static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1318                               int off)
1319{
1320        struct throtl_grp *tg = pd_to_tg(pd);
1321        unsigned int v = *(unsigned int *)((void *)tg + off);
1322
1323        if (v == -1)
1324                return 0;
1325        return __blkg_prfill_u64(sf, pd, v);
1326}
1327
1328static int tg_print_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1329                             struct seq_file *sf)
1330{
1331        blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_u64,
1332                          &blkcg_policy_throtl, cft->private, false);
1333        return 0;
1334}
1335
1336static int tg_print_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1337                              struct seq_file *sf)
1338{
1339        blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp), tg_prfill_conf_uint,
1340                          &blkcg_policy_throtl, cft->private, false);
1341        return 0;
1342}
1343
1344static int tg_set_conf(struct cgroup *cgrp, struct cftype *cft, const char *buf,
1345                       bool is_u64)
1346{
1347        struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1348        struct blkg_conf_ctx ctx;
1349        struct throtl_grp *tg;
1350        struct throtl_service_queue *sq;
1351        struct blkcg_gq *blkg;
1352        struct cgroup *pos_cgrp;
1353        int ret;
1354
1355        ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1356        if (ret)
1357                return ret;
1358
1359        tg = blkg_to_tg(ctx.blkg);
1360        sq = &tg->service_queue;
1361
1362        if (!ctx.v)
1363                ctx.v = -1;
1364
1365        if (is_u64)
1366                *(u64 *)((void *)tg + cft->private) = ctx.v;
1367        else
1368                *(unsigned int *)((void *)tg + cft->private) = ctx.v;
1369
1370        throtl_log(&tg->service_queue,
1371                   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1372                   tg->bps[READ], tg->bps[WRITE],
1373                   tg->iops[READ], tg->iops[WRITE]);
1374
1375        /*
1376         * Update has_rules[] flags for the updated tg's subtree.  A tg is
1377         * considered to have rules if either the tg itself or any of its
1378         * ancestors has rules.  This identifies groups without any
1379         * restrictions in the whole hierarchy and allows them to bypass
1380         * blk-throttle.
1381         */
1382        tg_update_has_rules(tg);
1383        blkg_for_each_descendant_pre(blkg, pos_cgrp, ctx.blkg)
1384                tg_update_has_rules(blkg_to_tg(blkg));
1385
1386        /*
1387         * We're already holding queue_lock and know @tg is valid.  Let's
1388         * apply the new config directly.
1389         *
1390         * Restart the slices for both READ and WRITES. It might happen
1391         * that a group's limit are dropped suddenly and we don't want to
1392         * account recently dispatched IO with new low rate.
1393         */
1394        throtl_start_new_slice(tg, 0);
1395        throtl_start_new_slice(tg, 1);
1396
1397        if (tg->flags & THROTL_TG_PENDING) {
1398                tg_update_disptime(tg);
1399                throtl_schedule_next_dispatch(sq->parent_sq, true);
1400        }
1401
1402        blkg_conf_finish(&ctx);
1403        return 0;
1404}
1405
1406static int tg_set_conf_u64(struct cgroup *cgrp, struct cftype *cft,
1407                           const char *buf)
1408{
1409        return tg_set_conf(cgrp, cft, buf, true);
1410}
1411
1412static int tg_set_conf_uint(struct cgroup *cgrp, struct cftype *cft,
1413                            const char *buf)
1414{
1415        return tg_set_conf(cgrp, cft, buf, false);
1416}
1417
1418static struct cftype throtl_files[] = {
1419        {
1420                .name = "throttle.read_bps_device",
1421                .private = offsetof(struct throtl_grp, bps[READ]),
1422                .read_seq_string = tg_print_conf_u64,
1423                .write_string = tg_set_conf_u64,
1424                .max_write_len = 256,
1425        },
1426        {
1427                .name = "throttle.write_bps_device",
1428                .private = offsetof(struct throtl_grp, bps[WRITE]),
1429                .read_seq_string = tg_print_conf_u64,
1430                .write_string = tg_set_conf_u64,
1431                .max_write_len = 256,
1432        },
1433        {
1434                .name = "throttle.read_iops_device",
1435                .private = offsetof(struct throtl_grp, iops[READ]),
1436                .read_seq_string = tg_print_conf_uint,
1437                .write_string = tg_set_conf_uint,
1438                .max_write_len = 256,
1439        },
1440        {
1441                .name = "throttle.write_iops_device",
1442                .private = offsetof(struct throtl_grp, iops[WRITE]),
1443                .read_seq_string = tg_print_conf_uint,
1444                .write_string = tg_set_conf_uint,
1445                .max_write_len = 256,
1446        },
1447        {
1448                .name = "throttle.io_service_bytes",
1449                .private = offsetof(struct tg_stats_cpu, service_bytes),
1450                .read_seq_string = tg_print_cpu_rwstat,
1451        },
1452        {
1453                .name = "throttle.io_serviced",
1454                .private = offsetof(struct tg_stats_cpu, serviced),
1455                .read_seq_string = tg_print_cpu_rwstat,
1456        },
1457        { }     /* terminate */
1458};
1459
1460static void throtl_shutdown_wq(struct request_queue *q)
1461{
1462        struct throtl_data *td = q->td;
1463
1464        cancel_work_sync(&td->dispatch_work);
1465}
1466
1467static struct blkcg_policy blkcg_policy_throtl = {
1468        .pd_size                = sizeof(struct throtl_grp),
1469        .cftypes                = throtl_files,
1470
1471        .pd_init_fn             = throtl_pd_init,
1472        .pd_online_fn           = throtl_pd_online,
1473        .pd_exit_fn             = throtl_pd_exit,
1474        .pd_reset_stats_fn      = throtl_pd_reset_stats,
1475};
1476
1477bool blk_throtl_bio(struct request_queue *q, struct bio *bio)
1478{
1479        struct throtl_data *td = q->td;
1480        struct throtl_qnode *qn = NULL;
1481        struct throtl_grp *tg;
1482        struct throtl_service_queue *sq;
1483        bool rw = bio_data_dir(bio);
1484        struct blkcg *blkcg;
1485        bool throttled = false;
1486
1487        /* see throtl_charge_bio() */
1488        if (bio->bi_rw & REQ_THROTTLED)
1489                goto out;
1490
1491        /*
1492         * A throtl_grp pointer retrieved under rcu can be used to access
1493         * basic fields like stats and io rates. If a group has no rules,
1494         * just update the dispatch stats in lockless manner and return.
1495         */
1496        rcu_read_lock();
1497        blkcg = bio_blkcg(bio);
1498        tg = throtl_lookup_tg(td, blkcg);
1499        if (tg) {
1500                if (!tg->has_rules[rw]) {
1501                        throtl_update_dispatch_stats(tg_to_blkg(tg),
1502                                                     bio->bi_size, bio->bi_rw);
1503                        goto out_unlock_rcu;
1504                }
1505        }
1506
1507        /*
1508         * Either group has not been allocated yet or it is not an unlimited
1509         * IO group
1510         */
1511        spin_lock_irq(q->queue_lock);
1512        tg = throtl_lookup_create_tg(td, blkcg);
1513        if (unlikely(!tg))
1514                goto out_unlock;
1515
1516        sq = &tg->service_queue;
1517
1518        while (true) {
1519                /* throtl is FIFO - if bios are already queued, should queue */
1520                if (sq->nr_queued[rw])
1521                        break;
1522
1523                /* if above limits, break to queue */
1524                if (!tg_may_dispatch(tg, bio, NULL))
1525                        break;
1526
1527                /* within limits, let's charge and dispatch directly */
1528                throtl_charge_bio(tg, bio);
1529
1530                /*
1531                 * We need to trim slice even when bios are not being queued
1532                 * otherwise it might happen that a bio is not queued for
1533                 * a long time and slice keeps on extending and trim is not
1534                 * called for a long time. Now if limits are reduced suddenly
1535                 * we take into account all the IO dispatched so far at new
1536                 * low rate and * newly queued IO gets a really long dispatch
1537                 * time.
1538                 *
1539                 * So keep on trimming slice even if bio is not queued.
1540                 */
1541                throtl_trim_slice(tg, rw);
1542
1543                /*
1544                 * @bio passed through this layer without being throttled.
1545                 * Climb up the ladder.  If we''re already at the top, it
1546                 * can be executed directly.
1547                 */
1548                qn = &tg->qnode_on_parent[rw];
1549                sq = sq->parent_sq;
1550                tg = sq_to_tg(sq);
1551                if (!tg)
1552                        goto out_unlock;
1553        }
1554
1555        /* out-of-limit, queue to @tg */
1556        throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1557                   rw == READ ? 'R' : 'W',
1558                   tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1559                   tg->io_disp[rw], tg->iops[rw],
1560                   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1561
1562        bio_associate_current(bio);
1563        tg->td->nr_queued[rw]++;
1564        throtl_add_bio_tg(bio, qn, tg);
1565        throttled = true;
1566
1567        /*
1568         * Update @tg's dispatch time and force schedule dispatch if @tg
1569         * was empty before @bio.  The forced scheduling isn't likely to
1570         * cause undue delay as @bio is likely to be dispatched directly if
1571         * its @tg's disptime is not in the future.
1572         */
1573        if (tg->flags & THROTL_TG_WAS_EMPTY) {
1574                tg_update_disptime(tg);
1575                throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1576        }
1577
1578out_unlock:
1579        spin_unlock_irq(q->queue_lock);
1580out_unlock_rcu:
1581        rcu_read_unlock();
1582out:
1583        /*
1584         * As multiple blk-throtls may stack in the same issue path, we
1585         * don't want bios to leave with the flag set.  Clear the flag if
1586         * being issued.
1587         */
1588        if (!throttled)
1589                bio->bi_rw &= ~REQ_THROTTLED;
1590        return throttled;
1591}
1592
1593/*
1594 * Dispatch all bios from all children tg's queued on @parent_sq.  On
1595 * return, @parent_sq is guaranteed to not have any active children tg's
1596 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1597 */
1598static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1599{
1600        struct throtl_grp *tg;
1601
1602        while ((tg = throtl_rb_first(parent_sq))) {
1603                struct throtl_service_queue *sq = &tg->service_queue;
1604                struct bio *bio;
1605
1606                throtl_dequeue_tg(tg);
1607
1608                while ((bio = throtl_peek_queued(&sq->queued[READ])))
1609                        tg_dispatch_one_bio(tg, bio_data_dir(bio));
1610                while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1611                        tg_dispatch_one_bio(tg, bio_data_dir(bio));
1612        }
1613}
1614
1615/**
1616 * blk_throtl_drain - drain throttled bios
1617 * @q: request_queue to drain throttled bios for
1618 *
1619 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1620 */
1621void blk_throtl_drain(struct request_queue *q)
1622        __releases(q->queue_lock) __acquires(q->queue_lock)
1623{
1624        struct throtl_data *td = q->td;
1625        struct blkcg_gq *blkg;
1626        struct cgroup *pos_cgrp;
1627        struct bio *bio;
1628        int rw;
1629
1630        queue_lockdep_assert_held(q);
1631        rcu_read_lock();
1632
1633        /*
1634         * Drain each tg while doing post-order walk on the blkg tree, so
1635         * that all bios are propagated to td->service_queue.  It'd be
1636         * better to walk service_queue tree directly but blkg walk is
1637         * easier.
1638         */
1639        blkg_for_each_descendant_post(blkg, pos_cgrp, td->queue->root_blkg)
1640                tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1641
1642        tg_drain_bios(&td_root_tg(td)->service_queue);
1643
1644        /* finally, transfer bios from top-level tg's into the td */
1645        tg_drain_bios(&td->service_queue);
1646
1647        rcu_read_unlock();
1648        spin_unlock_irq(q->queue_lock);
1649
1650        /* all bios now should be in td->service_queue, issue them */
1651        for (rw = READ; rw <= WRITE; rw++)
1652                while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1653                                                NULL)))
1654                        generic_make_request(bio);
1655
1656        spin_lock_irq(q->queue_lock);
1657}
1658
1659int blk_throtl_init(struct request_queue *q)
1660{
1661        struct throtl_data *td;
1662        int ret;
1663
1664        td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1665        if (!td)
1666                return -ENOMEM;
1667
1668        INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1669        throtl_service_queue_init(&td->service_queue, NULL);
1670
1671        q->td = td;
1672        td->queue = q;
1673
1674        /* activate policy */
1675        ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1676        if (ret)
1677                kfree(td);
1678        return ret;
1679}
1680
1681void blk_throtl_exit(struct request_queue *q)
1682{
1683        BUG_ON(!q->td);
1684        throtl_shutdown_wq(q);
1685        blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1686        kfree(q->td);
1687}
1688
1689static int __init throtl_init(void)
1690{
1691        kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1692        if (!kthrotld_workqueue)
1693                panic("Failed to create kthrotld\n");
1694
1695        return blkcg_policy_register(&blkcg_policy_throtl);
1696}
1697
1698module_init(throtl_init);
1699
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.