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