block/blk-throttle.c

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