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