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