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