block/blk-throttle.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Interface for controlling IO bandwidth on a request queue
   4  *
   5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/slab.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blktrace_api.h>
  13 #include "blk.h"
  14 #include "blk-cgroup-rwstat.h"
  15 #include "blk-stat.h"
  16 #include "blk-throttle.h"
  17
  18 /* Max dispatch from a group in 1 round */
  19 #define THROTL_GRP_QUANTUM 8
  20
  21 /* Total max dispatch from all groups in one round */
  22 #define THROTL_QUANTUM 32
  23
  24 /* Throttling is performed over a slice and after that slice is renewed */
  25 #define DFL_THROTL_SLICE_HD (HZ / 10)
  26 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  27 #define MAX_THROTL_SLICE (HZ)
  28 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  29 #define MIN_THROTL_BPS (320 * 1024)
  30 #define MIN_THROTL_IOPS (10)
  31 #define DFL_LATENCY_TARGET (-1L)
  32 #define DFL_IDLE_THRESHOLD (0)
  33 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  34 #define LATENCY_FILTERED_SSD (0)
  35 /*
  36  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  37  * help determine if its IO is impacted by others, hence we ignore the IO
  38  */
  39 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  40
  41 /* A workqueue to queue throttle related work */
  42 static struct workqueue_struct *kthrotld_workqueue;
  43
  44 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  45
  46 /* We measure latency for request size from <= 4k to >= 1M */
  47 #define LATENCY_BUCKET_SIZE 9
  48
  49 struct latency_bucket {
  50         unsigned long total_latency; /* ns / 1024 */
  51         int samples;
  52 };
  53
  54 struct avg_latency_bucket {
  55         unsigned long latency; /* ns / 1024 */
  56         bool valid;
  57 };
  58
  59 struct throtl_data
  60 {
  61         /* service tree for active throtl groups */
  62         struct throtl_service_queue service_queue;
  63
  64         struct request_queue *queue;
  65
  66         /* Total Number of queued bios on READ and WRITE lists */
  67         unsigned int nr_queued[2];
  68
  69         unsigned int throtl_slice;
  70
  71         /* Work for dispatching throttled bios */
  72         struct work_struct dispatch_work;
  73         unsigned int limit_index;
  74         bool limit_valid[LIMIT_CNT];
  75
  76         unsigned long low_upgrade_time;
  77         unsigned long low_downgrade_time;
  78
  79         unsigned int scale;
  80
  81         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
  82         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
  83         struct latency_bucket __percpu *latency_buckets[2];
  84         unsigned long last_calculate_time;
  85         unsigned long filtered_latency;
  86
  87         bool track_bio_latency;
  88 };
  89
  90 static void throtl_pending_timer_fn(struct timer_list *t);
  91
  92 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
  93 {
  94         return pd_to_blkg(&tg->pd);
  95 }
  96
  97 /**
  98  * sq_to_tg - return the throl_grp the specified service queue belongs to
  99  * @sq: the throtl_service_queue of interest
 100  *
 101  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 102  * embedded in throtl_data, %NULL is returned.
 103  */
 104 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 105 {
 106         if (sq && sq->parent_sq)
 107                 return container_of(sq, struct throtl_grp, service_queue);
 108         else
 109                 return NULL;
 110 }
 111
 112 /**
 113  * sq_to_td - return throtl_data the specified service queue belongs to
 114  * @sq: the throtl_service_queue of interest
 115  *
 116  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 117  * Determine the associated throtl_data accordingly and return it.
 118  */
 119 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 120 {
 121         struct throtl_grp *tg = sq_to_tg(sq);
 122
 123         if (tg)
 124                 return tg->td;
 125         else
 126                 return container_of(sq, struct throtl_data, service_queue);
 127 }
 128
 129 /*
 130  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 131  * make the IO dispatch more smooth.
 132  * Scale up: linearly scale up according to elapsed time since upgrade. For
 133  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 134  *           limit hits .max limit
 135  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 136  */
 137 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 138 {
 139         /* arbitrary value to avoid too big scale */
 140         if (td->scale < 4096 && time_after_eq(jiffies,
 141             td->low_upgrade_time + td->scale * td->throtl_slice))
 142                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 143
 144         return low + (low >> 1) * td->scale;
 145 }
 146
 147 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 148 {
 149         struct blkcg_gq *blkg = tg_to_blkg(tg);
 150         struct throtl_data *td;
 151         uint64_t ret;
 152
 153         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 154                 return U64_MAX;
 155
 156         td = tg->td;
 157         ret = tg->bps[rw][td->limit_index];
 158         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 159                 /* intermediate node or iops isn't 0 */
 160                 if (!list_empty(&blkg->blkcg->css.children) ||
 161                     tg->iops[rw][td->limit_index])
 162                         return U64_MAX;
 163                 else
 164                         return MIN_THROTL_BPS;
 165         }
 166
 167         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 168             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 169                 uint64_t adjusted;
 170
 171                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 172                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 173         }
 174         return ret;
 175 }
 176
 177 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 178 {
 179         struct blkcg_gq *blkg = tg_to_blkg(tg);
 180         struct throtl_data *td;
 181         unsigned int ret;
 182
 183         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 184                 return UINT_MAX;
 185
 186         td = tg->td;
 187         ret = tg->iops[rw][td->limit_index];
 188         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 189                 /* intermediate node or bps isn't 0 */
 190                 if (!list_empty(&blkg->blkcg->css.children) ||
 191                     tg->bps[rw][td->limit_index])
 192                         return UINT_MAX;
 193                 else
 194                         return MIN_THROTL_IOPS;
 195         }
 196
 197         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 198             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 199                 uint64_t adjusted;
 200
 201                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 202                 if (adjusted > UINT_MAX)
 203                         adjusted = UINT_MAX;
 204                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 205         }
 206         return ret;
 207 }
 208
 209 #define request_bucket_index(sectors) \
 210         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 211
 212 /**
 213  * throtl_log - log debug message via blktrace
 214  * @sq: the service_queue being reported
 215  * @fmt: printf format string
 216  * @args: printf args
 217  *
 218  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 219  * throtl_grp; otherwise, just "throtl".
 220  */
 221 #define throtl_log(sq, fmt, args...)    do {                            \
 222         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 223         struct throtl_data *__td = sq_to_td((sq));                      \
 224                                                                         \
 225         (void)__td;                                                     \
 226         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 227                 break;                                                  \
 228         if ((__tg)) {                                                   \
 229                 blk_add_cgroup_trace_msg(__td->queue,                   \
 230                         &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
 231         } else {                                                        \
 232                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 233         }                                                               \
 234 } while (0)
 235
 236 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 237 {
 238         /* assume it's one sector */
 239         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 240                 return 512;
 241         return bio->bi_iter.bi_size;
 242 }
 243
 244 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 245 {
 246         INIT_LIST_HEAD(&qn->node);
 247         bio_list_init(&qn->bios);
 248         qn->tg = tg;
 249 }
 250
 251 /**
 252  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 253  * @bio: bio being added
 254  * @qn: qnode to add bio to
 255  * @queued: the service_queue->queued[] list @qn belongs to
 256  *
 257  * Add @bio to @qn and put @qn on @queued if it's not already on.
 258  * @qn->tg's reference count is bumped when @qn is activated.  See the
 259  * comment on top of throtl_qnode definition for details.
 260  */
 261 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 262                                  struct list_head *queued)
 263 {
 264         bio_list_add(&qn->bios, bio);
 265         if (list_empty(&qn->node)) {
 266                 list_add_tail(&qn->node, queued);
 267                 blkg_get(tg_to_blkg(qn->tg));
 268         }
 269 }
 270
 271 /**
 272  * throtl_peek_queued - peek the first bio on a qnode list
 273  * @queued: the qnode list to peek
 274  */
 275 static struct bio *throtl_peek_queued(struct list_head *queued)
 276 {
 277         struct throtl_qnode *qn;
 278         struct bio *bio;
 279
 280         if (list_empty(queued))
 281                 return NULL;
 282
 283         qn = list_first_entry(queued, struct throtl_qnode, node);
 284         bio = bio_list_peek(&qn->bios);
 285         WARN_ON_ONCE(!bio);
 286         return bio;
 287 }
 288
 289 /**
 290  * throtl_pop_queued - pop the first bio form a qnode list
 291  * @queued: the qnode list to pop a bio from
 292  * @tg_to_put: optional out argument for throtl_grp to put
 293  *
 294  * Pop the first bio from the qnode list @queued.  After popping, the first
 295  * qnode is removed from @queued if empty or moved to the end of @queued so
 296  * that the popping order is round-robin.
 297  *
 298  * When the first qnode is removed, its associated throtl_grp should be put
 299  * too.  If @tg_to_put is NULL, this function automatically puts it;
 300  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 301  * responsible for putting it.
 302  */
 303 static struct bio *throtl_pop_queued(struct list_head *queued,
 304                                      struct throtl_grp **tg_to_put)
 305 {
 306         struct throtl_qnode *qn;
 307         struct bio *bio;
 308
 309         if (list_empty(queued))
 310                 return NULL;
 311
 312         qn = list_first_entry(queued, struct throtl_qnode, node);
 313         bio = bio_list_pop(&qn->bios);
 314         WARN_ON_ONCE(!bio);
 315
 316         if (bio_list_empty(&qn->bios)) {
 317                 list_del_init(&qn->node);
 318                 if (tg_to_put)
 319                         *tg_to_put = qn->tg;
 320                 else
 321                         blkg_put(tg_to_blkg(qn->tg));
 322         } else {
 323                 list_move_tail(&qn->node, queued);
 324         }
 325
 326         return bio;
 327 }
 328
 329 /* init a service_queue, assumes the caller zeroed it */
 330 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 331 {
 332         INIT_LIST_HEAD(&sq->queued[READ]);
 333         INIT_LIST_HEAD(&sq->queued[WRITE]);
 334         sq->pending_tree = RB_ROOT_CACHED;
 335         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 336 }
 337
 338 static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk,
 339                 struct blkcg *blkcg, gfp_t gfp)
 340 {
 341         struct throtl_grp *tg;
 342         int rw;
 343
 344         tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id);
 345         if (!tg)
 346                 return NULL;
 347
 348         if (blkg_rwstat_init(&tg->stat_bytes, gfp))
 349                 goto err_free_tg;
 350
 351         if (blkg_rwstat_init(&tg->stat_ios, gfp))
 352                 goto err_exit_stat_bytes;
 353
 354         throtl_service_queue_init(&tg->service_queue);
 355
 356         for (rw = READ; rw <= WRITE; rw++) {
 357                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 358                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 359         }
 360
 361         RB_CLEAR_NODE(&tg->rb_node);
 362         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 363         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 364         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 365         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 366         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 367         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 368         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 369         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 370         /* LIMIT_LOW will have default value 0 */
 371
 372         tg->latency_target = DFL_LATENCY_TARGET;
 373         tg->latency_target_conf = DFL_LATENCY_TARGET;
 374         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 375         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 376
 377         return &tg->pd;
 378
 379 err_exit_stat_bytes:
 380         blkg_rwstat_exit(&tg->stat_bytes);
 381 err_free_tg:
 382         kfree(tg);
 383         return NULL;
 384 }
 385
 386 static void throtl_pd_init(struct blkg_policy_data *pd)
 387 {
 388         struct throtl_grp *tg = pd_to_tg(pd);
 389         struct blkcg_gq *blkg = tg_to_blkg(tg);
 390         struct throtl_data *td = blkg->q->td;
 391         struct throtl_service_queue *sq = &tg->service_queue;
 392
 393         /*
 394          * If on the default hierarchy, we switch to properly hierarchical
 395          * behavior where limits on a given throtl_grp are applied to the
 396          * whole subtree rather than just the group itself.  e.g. If 16M
 397          * read_bps limit is set on a parent group, summary bps of
 398          * parent group and its subtree groups can't exceed 16M for the
 399          * device.
 400          *
 401          * If not on the default hierarchy, the broken flat hierarchy
 402          * behavior is retained where all throtl_grps are treated as if
 403          * they're all separate root groups right below throtl_data.
 404          * Limits of a group don't interact with limits of other groups
 405          * regardless of the position of the group in the hierarchy.
 406          */
 407         sq->parent_sq = &td->service_queue;
 408         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 409                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 410         tg->td = td;
 411 }
 412
 413 /*
 414  * Set has_rules[] if @tg or any of its parents have limits configured.
 415  * This doesn't require walking up to the top of the hierarchy as the
 416  * parent's has_rules[] is guaranteed to be correct.
 417  */
 418 static void tg_update_has_rules(struct throtl_grp *tg)
 419 {
 420         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 421         struct throtl_data *td = tg->td;
 422         int rw;
 423
 424         for (rw = READ; rw <= WRITE; rw++) {
 425                 tg->has_rules_iops[rw] =
 426                         (parent_tg && parent_tg->has_rules_iops[rw]) ||
 427                         (td->limit_valid[td->limit_index] &&
 428                           tg_iops_limit(tg, rw) != UINT_MAX);
 429                 tg->has_rules_bps[rw] =
 430                         (parent_tg && parent_tg->has_rules_bps[rw]) ||
 431                         (td->limit_valid[td->limit_index] &&
 432                          (tg_bps_limit(tg, rw) != U64_MAX));
 433         }
 434 }
 435
 436 static void throtl_pd_online(struct blkg_policy_data *pd)
 437 {
 438         struct throtl_grp *tg = pd_to_tg(pd);
 439         /*
 440          * We don't want new groups to escape the limits of its ancestors.
 441          * Update has_rules[] after a new group is brought online.
 442          */
 443         tg_update_has_rules(tg);
 444 }
 445
 446 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 447 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 448 {
 449         struct cgroup_subsys_state *pos_css;
 450         struct blkcg_gq *blkg;
 451         bool low_valid = false;
 452
 453         rcu_read_lock();
 454         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 455                 struct throtl_grp *tg = blkg_to_tg(blkg);
 456
 457                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 458                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 459                         low_valid = true;
 460                         break;
 461                 }
 462         }
 463         rcu_read_unlock();
 464
 465         td->limit_valid[LIMIT_LOW] = low_valid;
 466 }
 467 #else
 468 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
 469 {
 470 }
 471 #endif
 472
 473 static void throtl_upgrade_state(struct throtl_data *td);
 474 static void throtl_pd_offline(struct blkg_policy_data *pd)
 475 {
 476         struct throtl_grp *tg = pd_to_tg(pd);
 477
 478         tg->bps[READ][LIMIT_LOW] = 0;
 479         tg->bps[WRITE][LIMIT_LOW] = 0;
 480         tg->iops[READ][LIMIT_LOW] = 0;
 481         tg->iops[WRITE][LIMIT_LOW] = 0;
 482
 483         blk_throtl_update_limit_valid(tg->td);
 484
 485         if (!tg->td->limit_valid[tg->td->limit_index])
 486                 throtl_upgrade_state(tg->td);
 487 }
 488
 489 static void throtl_pd_free(struct blkg_policy_data *pd)
 490 {
 491         struct throtl_grp *tg = pd_to_tg(pd);
 492
 493         del_timer_sync(&tg->service_queue.pending_timer);
 494         blkg_rwstat_exit(&tg->stat_bytes);
 495         blkg_rwstat_exit(&tg->stat_ios);
 496         kfree(tg);
 497 }
 498
 499 static struct throtl_grp *
 500 throtl_rb_first(struct throtl_service_queue *parent_sq)
 501 {
 502         struct rb_node *n;
 503
 504         n = rb_first_cached(&parent_sq->pending_tree);
 505         WARN_ON_ONCE(!n);
 506         if (!n)
 507                 return NULL;
 508         return rb_entry_tg(n);
 509 }
 510
 511 static void throtl_rb_erase(struct rb_node *n,
 512                             struct throtl_service_queue *parent_sq)
 513 {
 514         rb_erase_cached(n, &parent_sq->pending_tree);
 515         RB_CLEAR_NODE(n);
 516 }
 517
 518 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 519 {
 520         struct throtl_grp *tg;
 521
 522         tg = throtl_rb_first(parent_sq);
 523         if (!tg)
 524                 return;
 525
 526         parent_sq->first_pending_disptime = tg->disptime;
 527 }
 528
 529 static void tg_service_queue_add(struct throtl_grp *tg)
 530 {
 531         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 532         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 533         struct rb_node *parent = NULL;
 534         struct throtl_grp *__tg;
 535         unsigned long key = tg->disptime;
 536         bool leftmost = true;
 537
 538         while (*node != NULL) {
 539                 parent = *node;
 540                 __tg = rb_entry_tg(parent);
 541
 542                 if (time_before(key, __tg->disptime))
 543                         node = &parent->rb_left;
 544                 else {
 545                         node = &parent->rb_right;
 546                         leftmost = false;
 547                 }
 548         }
 549
 550         rb_link_node(&tg->rb_node, parent, node);
 551         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 552                                leftmost);
 553 }
 554
 555 static void throtl_enqueue_tg(struct throtl_grp *tg)
 556 {
 557         if (!(tg->flags & THROTL_TG_PENDING)) {
 558                 tg_service_queue_add(tg);
 559                 tg->flags |= THROTL_TG_PENDING;
 560                 tg->service_queue.parent_sq->nr_pending++;
 561         }
 562 }
 563
 564 static void throtl_dequeue_tg(struct throtl_grp *tg)
 565 {
 566         if (tg->flags & THROTL_TG_PENDING) {
 567                 struct throtl_service_queue *parent_sq =
 568                         tg->service_queue.parent_sq;
 569
 570                 throtl_rb_erase(&tg->rb_node, parent_sq);
 571                 --parent_sq->nr_pending;
 572                 tg->flags &= ~THROTL_TG_PENDING;
 573         }
 574 }
 575
 576 /* Call with queue lock held */
 577 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 578                                           unsigned long expires)
 579 {
 580         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 581
 582         /*
 583          * Since we are adjusting the throttle limit dynamically, the sleep
 584          * time calculated according to previous limit might be invalid. It's
 585          * possible the cgroup sleep time is very long and no other cgroups
 586          * have IO running so notify the limit changes. Make sure the cgroup
 587          * doesn't sleep too long to avoid the missed notification.
 588          */
 589         if (time_after(expires, max_expire))
 590                 expires = max_expire;
 591         mod_timer(&sq->pending_timer, expires);
 592         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 593                    expires - jiffies, jiffies);
 594 }
 595
 596 /**
 597  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 598  * @sq: the service_queue to schedule dispatch for
 599  * @force: force scheduling
 600  *
 601  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 602  * dispatch time of the first pending child.  Returns %true if either timer
 603  * is armed or there's no pending child left.  %false if the current
 604  * dispatch window is still open and the caller should continue
 605  * dispatching.
 606  *
 607  * If @force is %true, the dispatch timer is always scheduled and this
 608  * function is guaranteed to return %true.  This is to be used when the
 609  * caller can't dispatch itself and needs to invoke pending_timer
 610  * unconditionally.  Note that forced scheduling is likely to induce short
 611  * delay before dispatch starts even if @sq->first_pending_disptime is not
 612  * in the future and thus shouldn't be used in hot paths.
 613  */
 614 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 615                                           bool force)
 616 {
 617         /* any pending children left? */
 618         if (!sq->nr_pending)
 619                 return true;
 620
 621         update_min_dispatch_time(sq);
 622
 623         /* is the next dispatch time in the future? */
 624         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 625                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 626                 return true;
 627         }
 628
 629         /* tell the caller to continue dispatching */
 630         return false;
 631 }
 632
 633 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 634                 bool rw, unsigned long start)
 635 {
 636         tg->bytes_disp[rw] = 0;
 637         tg->io_disp[rw] = 0;
 638         tg->carryover_bytes[rw] = 0;
 639         tg->carryover_ios[rw] = 0;
 640
 641         /*
 642          * Previous slice has expired. We must have trimmed it after last
 643          * bio dispatch. That means since start of last slice, we never used
 644          * that bandwidth. Do try to make use of that bandwidth while giving
 645          * credit.
 646          */
 647         if (time_after(start, tg->slice_start[rw]))
 648                 tg->slice_start[rw] = start;
 649
 650         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 651         throtl_log(&tg->service_queue,
 652                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 653                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 654                    tg->slice_end[rw], jiffies);
 655 }
 656
 657 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
 658                                           bool clear_carryover)
 659 {
 660         tg->bytes_disp[rw] = 0;
 661         tg->io_disp[rw] = 0;
 662         tg->slice_start[rw] = jiffies;
 663         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 664         if (clear_carryover) {
 665                 tg->carryover_bytes[rw] = 0;
 666                 tg->carryover_ios[rw] = 0;
 667         }
 668
 669         throtl_log(&tg->service_queue,
 670                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 671                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 672                    tg->slice_end[rw], jiffies);
 673 }
 674
 675 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 676                                         unsigned long jiffy_end)
 677 {
 678         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 679 }
 680
 681 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 682                                        unsigned long jiffy_end)
 683 {
 684         throtl_set_slice_end(tg, rw, jiffy_end);
 685         throtl_log(&tg->service_queue,
 686                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 687                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 688                    tg->slice_end[rw], jiffies);
 689 }
 690
 691 /* Determine if previously allocated or extended slice is complete or not */
 692 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 693 {
 694         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 695                 return false;
 696
 697         return true;
 698 }
 699
 700 /* Trim the used slices and adjust slice start accordingly */
 701 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 702 {
 703         unsigned long nr_slices, time_elapsed, io_trim;
 704         u64 bytes_trim, tmp;
 705
 706         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 707
 708         /*
 709          * If bps are unlimited (-1), then time slice don't get
 710          * renewed. Don't try to trim the slice if slice is used. A new
 711          * slice will start when appropriate.
 712          */
 713         if (throtl_slice_used(tg, rw))
 714                 return;
 715
 716         /*
 717          * A bio has been dispatched. Also adjust slice_end. It might happen
 718          * that initially cgroup limit was very low resulting in high
 719          * slice_end, but later limit was bumped up and bio was dispatched
 720          * sooner, then we need to reduce slice_end. A high bogus slice_end
 721          * is bad because it does not allow new slice to start.
 722          */
 723
 724         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 725
 726         time_elapsed = jiffies - tg->slice_start[rw];
 727
 728         nr_slices = time_elapsed / tg->td->throtl_slice;
 729
 730         if (!nr_slices)
 731                 return;
 732         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 733         do_div(tmp, HZ);
 734         bytes_trim = tmp;
 735
 736         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 737                 HZ;
 738
 739         if (!bytes_trim && !io_trim)
 740                 return;
 741
 742         if (tg->bytes_disp[rw] >= bytes_trim)
 743                 tg->bytes_disp[rw] -= bytes_trim;
 744         else
 745                 tg->bytes_disp[rw] = 0;
 746
 747         if (tg->io_disp[rw] >= io_trim)
 748                 tg->io_disp[rw] -= io_trim;
 749         else
 750                 tg->io_disp[rw] = 0;
 751
 752         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 753
 754         throtl_log(&tg->service_queue,
 755                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 756                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 757                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 758 }
 759
 760 static unsigned int calculate_io_allowed(u32 iops_limit,
 761                                          unsigned long jiffy_elapsed)
 762 {
 763         unsigned int io_allowed;
 764         u64 tmp;
 765
 766         /*
 767          * jiffy_elapsed should not be a big value as minimum iops can be
 768          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 769          * will allow dispatch after 1 second and after that slice should
 770          * have been trimmed.
 771          */
 772
 773         tmp = (u64)iops_limit * jiffy_elapsed;
 774         do_div(tmp, HZ);
 775
 776         if (tmp > UINT_MAX)
 777                 io_allowed = UINT_MAX;
 778         else
 779                 io_allowed = tmp;
 780
 781         return io_allowed;
 782 }
 783
 784 static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
 785 {
 786         return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
 787 }
 788
 789 static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
 790 {
 791         unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
 792         u64 bps_limit = tg_bps_limit(tg, rw);
 793         u32 iops_limit = tg_iops_limit(tg, rw);
 794
 795         /*
 796          * If config is updated while bios are still throttled, calculate and
 797          * accumulate how many bytes/ios are waited across changes. And
 798          * carryover_bytes/ios will be used to calculate new wait time under new
 799          * configuration.
 800          */
 801         if (bps_limit != U64_MAX)
 802                 tg->carryover_bytes[rw] +=
 803                         calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
 804                         tg->bytes_disp[rw];
 805         if (iops_limit != UINT_MAX)
 806                 tg->carryover_ios[rw] +=
 807                         calculate_io_allowed(iops_limit, jiffy_elapsed) -
 808                         tg->io_disp[rw];
 809 }
 810
 811 static void tg_update_carryover(struct throtl_grp *tg)
 812 {
 813         if (tg->service_queue.nr_queued[READ])
 814                 __tg_update_carryover(tg, READ);
 815         if (tg->service_queue.nr_queued[WRITE])
 816                 __tg_update_carryover(tg, WRITE);
 817
 818         /* see comments in struct throtl_grp for meaning of these fields. */
 819         throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
 820                    tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
 821                    tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
 822 }
 823
 824 static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
 825                                  u32 iops_limit)
 826 {
 827         bool rw = bio_data_dir(bio);
 828         int io_allowed;
 829         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 830
 831         if (iops_limit == UINT_MAX) {
 832                 return 0;
 833         }
 834
 835         jiffy_elapsed = jiffies - tg->slice_start[rw];
 836
 837         /* Round up to the next throttle slice, wait time must be nonzero */
 838         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 839         io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
 840                      tg->carryover_ios[rw];
 841         if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
 842                 return 0;
 843
 844         /* Calc approx time to dispatch */
 845         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 846         return jiffy_wait;
 847 }
 848
 849 static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
 850                                 u64 bps_limit)
 851 {
 852         bool rw = bio_data_dir(bio);
 853         long long bytes_allowed;
 854         u64 extra_bytes;
 855         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 856         unsigned int bio_size = throtl_bio_data_size(bio);
 857
 858         /* no need to throttle if this bio's bytes have been accounted */
 859         if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
 860                 return 0;
 861         }
 862
 863         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 864
 865         /* Slice has just started. Consider one slice interval */
 866         if (!jiffy_elapsed)
 867                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 868
 869         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 870         bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
 871                         tg->carryover_bytes[rw];
 872         if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
 873                 return 0;
 874
 875         /* Calc approx time to dispatch */
 876         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 877         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 878
 879         if (!jiffy_wait)
 880                 jiffy_wait = 1;
 881
 882         /*
 883          * This wait time is without taking into consideration the rounding
 884          * up we did. Add that time also.
 885          */
 886         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 887         return jiffy_wait;
 888 }
 889
 890 /*
 891  * Returns whether one can dispatch a bio or not. Also returns approx number
 892  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 893  */
 894 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 895                             unsigned long *wait)
 896 {
 897         bool rw = bio_data_dir(bio);
 898         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 899         u64 bps_limit = tg_bps_limit(tg, rw);
 900         u32 iops_limit = tg_iops_limit(tg, rw);
 901
 902         /*
 903          * Currently whole state machine of group depends on first bio
 904          * queued in the group bio list. So one should not be calling
 905          * this function with a different bio if there are other bios
 906          * queued.
 907          */
 908         BUG_ON(tg->service_queue.nr_queued[rw] &&
 909                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 910
 911         /* If tg->bps = -1, then BW is unlimited */
 912         if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
 913             tg->flags & THROTL_TG_CANCELING) {
 914                 if (wait)
 915                         *wait = 0;
 916                 return true;
 917         }
 918
 919         /*
 920          * If previous slice expired, start a new one otherwise renew/extend
 921          * existing slice to make sure it is at least throtl_slice interval
 922          * long since now. New slice is started only for empty throttle group.
 923          * If there is queued bio, that means there should be an active
 924          * slice and it should be extended instead.
 925          */
 926         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 927                 throtl_start_new_slice(tg, rw, true);
 928         else {
 929                 if (time_before(tg->slice_end[rw],
 930                     jiffies + tg->td->throtl_slice))
 931                         throtl_extend_slice(tg, rw,
 932                                 jiffies + tg->td->throtl_slice);
 933         }
 934
 935         bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
 936         iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
 937         if (bps_wait + iops_wait == 0) {
 938                 if (wait)
 939                         *wait = 0;
 940                 return true;
 941         }
 942
 943         max_wait = max(bps_wait, iops_wait);
 944
 945         if (wait)
 946                 *wait = max_wait;
 947
 948         if (time_before(tg->slice_end[rw], jiffies + max_wait))
 949                 throtl_extend_slice(tg, rw, jiffies + max_wait);
 950
 951         return false;
 952 }
 953
 954 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 955 {
 956         bool rw = bio_data_dir(bio);
 957         unsigned int bio_size = throtl_bio_data_size(bio);
 958
 959         /* Charge the bio to the group */
 960         if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
 961                 tg->bytes_disp[rw] += bio_size;
 962                 tg->last_bytes_disp[rw] += bio_size;
 963         }
 964
 965         tg->io_disp[rw]++;
 966         tg->last_io_disp[rw]++;
 967 }
 968
 969 /**
 970  * throtl_add_bio_tg - add a bio to the specified throtl_grp
 971  * @bio: bio to add
 972  * @qn: qnode to use
 973  * @tg: the target throtl_grp
 974  *
 975  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 976  * tg->qnode_on_self[] is used.
 977  */
 978 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 979                               struct throtl_grp *tg)
 980 {
 981         struct throtl_service_queue *sq = &tg->service_queue;
 982         bool rw = bio_data_dir(bio);
 983
 984         if (!qn)
 985                 qn = &tg->qnode_on_self[rw];
 986
 987         /*
 988          * If @tg doesn't currently have any bios queued in the same
 989          * direction, queueing @bio can change when @tg should be
 990          * dispatched.  Mark that @tg was empty.  This is automatically
 991          * cleared on the next tg_update_disptime().
 992          */
 993         if (!sq->nr_queued[rw])
 994                 tg->flags |= THROTL_TG_WAS_EMPTY;
 995
 996         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
 997
 998         sq->nr_queued[rw]++;
 999         throtl_enqueue_tg(tg);
1000 }
1001
1002 static void tg_update_disptime(struct throtl_grp *tg)
1003 {
1004         struct throtl_service_queue *sq = &tg->service_queue;
1005         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1006         struct bio *bio;
1007
1008         bio = throtl_peek_queued(&sq->queued[READ]);
1009         if (bio)
1010                 tg_may_dispatch(tg, bio, &read_wait);
1011
1012         bio = throtl_peek_queued(&sq->queued[WRITE]);
1013         if (bio)
1014                 tg_may_dispatch(tg, bio, &write_wait);
1015
1016         min_wait = min(read_wait, write_wait);
1017         disptime = jiffies + min_wait;
1018
1019         /* Update dispatch time */
1020         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
1021         tg->disptime = disptime;
1022         tg_service_queue_add(tg);
1023
1024         /* see throtl_add_bio_tg() */
1025         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1026 }
1027
1028 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1029                                         struct throtl_grp *parent_tg, bool rw)
1030 {
1031         if (throtl_slice_used(parent_tg, rw)) {
1032                 throtl_start_new_slice_with_credit(parent_tg, rw,
1033                                 child_tg->slice_start[rw]);
1034         }
1035
1036 }
1037
1038 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1039 {
1040         struct throtl_service_queue *sq = &tg->service_queue;
1041         struct throtl_service_queue *parent_sq = sq->parent_sq;
1042         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1043         struct throtl_grp *tg_to_put = NULL;
1044         struct bio *bio;
1045
1046         /*
1047          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1048          * from @tg may put its reference and @parent_sq might end up
1049          * getting released prematurely.  Remember the tg to put and put it
1050          * after @bio is transferred to @parent_sq.
1051          */
1052         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1053         sq->nr_queued[rw]--;
1054
1055         throtl_charge_bio(tg, bio);
1056
1057         /*
1058          * If our parent is another tg, we just need to transfer @bio to
1059          * the parent using throtl_add_bio_tg().  If our parent is
1060          * @td->service_queue, @bio is ready to be issued.  Put it on its
1061          * bio_lists[] and decrease total number queued.  The caller is
1062          * responsible for issuing these bios.
1063          */
1064         if (parent_tg) {
1065                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1066                 start_parent_slice_with_credit(tg, parent_tg, rw);
1067         } else {
1068                 bio_set_flag(bio, BIO_BPS_THROTTLED);
1069                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1070                                      &parent_sq->queued[rw]);
1071                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1072                 tg->td->nr_queued[rw]--;
1073         }
1074
1075         throtl_trim_slice(tg, rw);
1076
1077         if (tg_to_put)
1078                 blkg_put(tg_to_blkg(tg_to_put));
1079 }
1080
1081 static int throtl_dispatch_tg(struct throtl_grp *tg)
1082 {
1083         struct throtl_service_queue *sq = &tg->service_queue;
1084         unsigned int nr_reads = 0, nr_writes = 0;
1085         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1086         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1087         struct bio *bio;
1088
1089         /* Try to dispatch 75% READS and 25% WRITES */
1090
1091         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1092                tg_may_dispatch(tg, bio, NULL)) {
1093
1094                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1095                 nr_reads++;
1096
1097                 if (nr_reads >= max_nr_reads)
1098                         break;
1099         }
1100
1101         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1102                tg_may_dispatch(tg, bio, NULL)) {
1103
1104                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1105                 nr_writes++;
1106
1107                 if (nr_writes >= max_nr_writes)
1108                         break;
1109         }
1110
1111         return nr_reads + nr_writes;
1112 }
1113
1114 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1115 {
1116         unsigned int nr_disp = 0;
1117
1118         while (1) {
1119                 struct throtl_grp *tg;
1120                 struct throtl_service_queue *sq;
1121
1122                 if (!parent_sq->nr_pending)
1123                         break;
1124
1125                 tg = throtl_rb_first(parent_sq);
1126                 if (!tg)
1127                         break;
1128
1129                 if (time_before(jiffies, tg->disptime))
1130                         break;
1131
1132                 nr_disp += throtl_dispatch_tg(tg);
1133
1134                 sq = &tg->service_queue;
1135                 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
1136                         tg_update_disptime(tg);
1137                 else
1138                         throtl_dequeue_tg(tg);
1139
1140                 if (nr_disp >= THROTL_QUANTUM)
1141                         break;
1142         }
1143
1144         return nr_disp;
1145 }
1146
1147 static bool throtl_can_upgrade(struct throtl_data *td,
1148         struct throtl_grp *this_tg);
1149 /**
1150  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1151  * @t: the pending_timer member of the throtl_service_queue being serviced
1152  *
1153  * This timer is armed when a child throtl_grp with active bio's become
1154  * pending and queued on the service_queue's pending_tree and expires when
1155  * the first child throtl_grp should be dispatched.  This function
1156  * dispatches bio's from the children throtl_grps to the parent
1157  * service_queue.
1158  *
1159  * If the parent's parent is another throtl_grp, dispatching is propagated
1160  * by either arming its pending_timer or repeating dispatch directly.  If
1161  * the top-level service_tree is reached, throtl_data->dispatch_work is
1162  * kicked so that the ready bio's are issued.
1163  */
1164 static void throtl_pending_timer_fn(struct timer_list *t)
1165 {
1166         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1167         struct throtl_grp *tg = sq_to_tg(sq);
1168         struct throtl_data *td = sq_to_td(sq);
1169         struct throtl_service_queue *parent_sq;
1170         struct request_queue *q;
1171         bool dispatched;
1172         int ret;
1173
1174         /* throtl_data may be gone, so figure out request queue by blkg */
1175         if (tg)
1176                 q = tg->pd.blkg->q;
1177         else
1178                 q = td->queue;
1179
1180         spin_lock_irq(&q->queue_lock);
1181
1182         if (!q->root_blkg)
1183                 goto out_unlock;
1184
1185         if (throtl_can_upgrade(td, NULL))
1186                 throtl_upgrade_state(td);
1187
1188 again:
1189         parent_sq = sq->parent_sq;
1190         dispatched = false;
1191
1192         while (true) {
1193                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1194                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1195                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1196
1197                 ret = throtl_select_dispatch(sq);
1198                 if (ret) {
1199                         throtl_log(sq, "bios disp=%u", ret);
1200                         dispatched = true;
1201                 }
1202
1203                 if (throtl_schedule_next_dispatch(sq, false))
1204                         break;
1205
1206                 /* this dispatch windows is still open, relax and repeat */
1207                 spin_unlock_irq(&q->queue_lock);
1208                 cpu_relax();
1209                 spin_lock_irq(&q->queue_lock);
1210         }
1211
1212         if (!dispatched)
1213                 goto out_unlock;
1214
1215         if (parent_sq) {
1216                 /* @parent_sq is another throl_grp, propagate dispatch */
1217                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1218                         tg_update_disptime(tg);
1219                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1220                                 /* window is already open, repeat dispatching */
1221                                 sq = parent_sq;
1222                                 tg = sq_to_tg(sq);
1223                                 goto again;
1224                         }
1225                 }
1226         } else {
1227                 /* reached the top-level, queue issuing */
1228                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1229         }
1230 out_unlock:
1231         spin_unlock_irq(&q->queue_lock);
1232 }
1233
1234 /**
1235  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1236  * @work: work item being executed
1237  *
1238  * This function is queued for execution when bios reach the bio_lists[]
1239  * of throtl_data->service_queue.  Those bios are ready and issued by this
1240  * function.
1241  */
1242 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1243 {
1244         struct throtl_data *td = container_of(work, struct throtl_data,
1245                                               dispatch_work);
1246         struct throtl_service_queue *td_sq = &td->service_queue;
1247         struct request_queue *q = td->queue;
1248         struct bio_list bio_list_on_stack;
1249         struct bio *bio;
1250         struct blk_plug plug;
1251         int rw;
1252
1253         bio_list_init(&bio_list_on_stack);
1254
1255         spin_lock_irq(&q->queue_lock);
1256         for (rw = READ; rw <= WRITE; rw++)
1257                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1258                         bio_list_add(&bio_list_on_stack, bio);
1259         spin_unlock_irq(&q->queue_lock);
1260
1261         if (!bio_list_empty(&bio_list_on_stack)) {
1262                 blk_start_plug(&plug);
1263                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1264                         submit_bio_noacct_nocheck(bio);
1265                 blk_finish_plug(&plug);
1266         }
1267 }
1268
1269 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1270                               int off)
1271 {
1272         struct throtl_grp *tg = pd_to_tg(pd);
1273         u64 v = *(u64 *)((void *)tg + off);
1274
1275         if (v == U64_MAX)
1276                 return 0;
1277         return __blkg_prfill_u64(sf, pd, v);
1278 }
1279
1280 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1281                                int off)
1282 {
1283         struct throtl_grp *tg = pd_to_tg(pd);
1284         unsigned int v = *(unsigned int *)((void *)tg + off);
1285
1286         if (v == UINT_MAX)
1287                 return 0;
1288         return __blkg_prfill_u64(sf, pd, v);
1289 }
1290
1291 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1292 {
1293         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1294                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1295         return 0;
1296 }
1297
1298 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1299 {
1300         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1301                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1302         return 0;
1303 }
1304
1305 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1306 {
1307         struct throtl_service_queue *sq = &tg->service_queue;
1308         struct cgroup_subsys_state *pos_css;
1309         struct blkcg_gq *blkg;
1310
1311         throtl_log(&tg->service_queue,
1312                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1313                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1314                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1315
1316         /*
1317          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1318          * considered to have rules if either the tg itself or any of its
1319          * ancestors has rules.  This identifies groups without any
1320          * restrictions in the whole hierarchy and allows them to bypass
1321          * blk-throttle.
1322          */
1323         blkg_for_each_descendant_pre(blkg, pos_css,
1324                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1325                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1326                 struct throtl_grp *parent_tg;
1327
1328                 tg_update_has_rules(this_tg);
1329                 /* ignore root/second level */
1330                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1331                     !blkg->parent->parent)
1332                         continue;
1333                 parent_tg = blkg_to_tg(blkg->parent);
1334                 /*
1335                  * make sure all children has lower idle time threshold and
1336                  * higher latency target
1337                  */
1338                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1339                                 parent_tg->idletime_threshold);
1340                 this_tg->latency_target = max(this_tg->latency_target,
1341                                 parent_tg->latency_target);
1342         }
1343
1344         /*
1345          * We're already holding queue_lock and know @tg is valid.  Let's
1346          * apply the new config directly.
1347          *
1348          * Restart the slices for both READ and WRITES. It might happen
1349          * that a group's limit are dropped suddenly and we don't want to
1350          * account recently dispatched IO with new low rate.
1351          */
1352         throtl_start_new_slice(tg, READ, false);
1353         throtl_start_new_slice(tg, WRITE, false);
1354
1355         if (tg->flags & THROTL_TG_PENDING) {
1356                 tg_update_disptime(tg);
1357                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1358         }
1359 }
1360
1361 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1362                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1363 {
1364         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1365         struct blkg_conf_ctx ctx;
1366         struct throtl_grp *tg;
1367         int ret;
1368         u64 v;
1369
1370         blkg_conf_init(&ctx, buf);
1371
1372         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1373         if (ret)
1374                 goto out_finish;
1375
1376         ret = -EINVAL;
1377         if (sscanf(ctx.body, "%llu", &v) != 1)
1378                 goto out_finish;
1379         if (!v)
1380                 v = U64_MAX;
1381
1382         tg = blkg_to_tg(ctx.blkg);
1383         tg_update_carryover(tg);
1384
1385         if (is_u64)
1386                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1387         else
1388                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1389
1390         tg_conf_updated(tg, false);
1391         ret = 0;
1392 out_finish:
1393         blkg_conf_exit(&ctx);
1394         return ret ?: nbytes;
1395 }
1396
1397 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1398                                char *buf, size_t nbytes, loff_t off)
1399 {
1400         return tg_set_conf(of, buf, nbytes, off, true);
1401 }
1402
1403 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1404                                 char *buf, size_t nbytes, loff_t off)
1405 {
1406         return tg_set_conf(of, buf, nbytes, off, false);
1407 }
1408
1409 static int tg_print_rwstat(struct seq_file *sf, void *v)
1410 {
1411         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1412                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1413                           seq_cft(sf)->private, true);
1414         return 0;
1415 }
1416
1417 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1418                                       struct blkg_policy_data *pd, int off)
1419 {
1420         struct blkg_rwstat_sample sum;
1421
1422         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1423                                   &sum);
1424         return __blkg_prfill_rwstat(sf, pd, &sum);
1425 }
1426
1427 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1428 {
1429         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1430                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1431                           seq_cft(sf)->private, true);
1432         return 0;
1433 }
1434
1435 static struct cftype throtl_legacy_files[] = {
1436         {
1437                 .name = "throttle.read_bps_device",
1438                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1439                 .seq_show = tg_print_conf_u64,
1440                 .write = tg_set_conf_u64,
1441         },
1442         {
1443                 .name = "throttle.write_bps_device",
1444                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1445                 .seq_show = tg_print_conf_u64,
1446                 .write = tg_set_conf_u64,
1447         },
1448         {
1449                 .name = "throttle.read_iops_device",
1450                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1451                 .seq_show = tg_print_conf_uint,
1452                 .write = tg_set_conf_uint,
1453         },
1454         {
1455                 .name = "throttle.write_iops_device",
1456                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1457                 .seq_show = tg_print_conf_uint,
1458                 .write = tg_set_conf_uint,
1459         },
1460         {
1461                 .name = "throttle.io_service_bytes",
1462                 .private = offsetof(struct throtl_grp, stat_bytes),
1463                 .seq_show = tg_print_rwstat,
1464         },
1465         {
1466                 .name = "throttle.io_service_bytes_recursive",
1467                 .private = offsetof(struct throtl_grp, stat_bytes),
1468                 .seq_show = tg_print_rwstat_recursive,
1469         },
1470         {
1471                 .name = "throttle.io_serviced",
1472                 .private = offsetof(struct throtl_grp, stat_ios),
1473                 .seq_show = tg_print_rwstat,
1474         },
1475         {
1476                 .name = "throttle.io_serviced_recursive",
1477                 .private = offsetof(struct throtl_grp, stat_ios),
1478                 .seq_show = tg_print_rwstat_recursive,
1479         },
1480         { }     /* terminate */
1481 };
1482
1483 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1484                          int off)
1485 {
1486         struct throtl_grp *tg = pd_to_tg(pd);
1487         const char *dname = blkg_dev_name(pd->blkg);
1488         char bufs[4][21] = { "max", "max", "max", "max" };
1489         u64 bps_dft;
1490         unsigned int iops_dft;
1491         char idle_time[26] = "";
1492         char latency_time[26] = "";
1493
1494         if (!dname)
1495                 return 0;
1496
1497         if (off == LIMIT_LOW) {
1498                 bps_dft = 0;
1499                 iops_dft = 0;
1500         } else {
1501                 bps_dft = U64_MAX;
1502                 iops_dft = UINT_MAX;
1503         }
1504
1505         if (tg->bps_conf[READ][off] == bps_dft &&
1506             tg->bps_conf[WRITE][off] == bps_dft &&
1507             tg->iops_conf[READ][off] == iops_dft &&
1508             tg->iops_conf[WRITE][off] == iops_dft &&
1509             (off != LIMIT_LOW ||
1510              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1511               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1512                 return 0;
1513
1514         if (tg->bps_conf[READ][off] != U64_MAX)
1515                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1516                         tg->bps_conf[READ][off]);
1517         if (tg->bps_conf[WRITE][off] != U64_MAX)
1518                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1519                         tg->bps_conf[WRITE][off]);
1520         if (tg->iops_conf[READ][off] != UINT_MAX)
1521                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1522                         tg->iops_conf[READ][off]);
1523         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1524                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1525                         tg->iops_conf[WRITE][off]);
1526         if (off == LIMIT_LOW) {
1527                 if (tg->idletime_threshold_conf == ULONG_MAX)
1528                         strcpy(idle_time, " idle=max");
1529                 else
1530                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1531                                 tg->idletime_threshold_conf);
1532
1533                 if (tg->latency_target_conf == ULONG_MAX)
1534                         strcpy(latency_time, " latency=max");
1535                 else
1536                         snprintf(latency_time, sizeof(latency_time),
1537                                 " latency=%lu", tg->latency_target_conf);
1538         }
1539
1540         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1541                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1542                    latency_time);
1543         return 0;
1544 }
1545
1546 static int tg_print_limit(struct seq_file *sf, void *v)
1547 {
1548         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1549                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1550         return 0;
1551 }
1552
1553 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1554                           char *buf, size_t nbytes, loff_t off)
1555 {
1556         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1557         struct blkg_conf_ctx ctx;
1558         struct throtl_grp *tg;
1559         u64 v[4];
1560         unsigned long idle_time;
1561         unsigned long latency_time;
1562         int ret;
1563         int index = of_cft(of)->private;
1564
1565         blkg_conf_init(&ctx, buf);
1566
1567         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1568         if (ret)
1569                 goto out_finish;
1570
1571         tg = blkg_to_tg(ctx.blkg);
1572         tg_update_carryover(tg);
1573
1574         v[0] = tg->bps_conf[READ][index];
1575         v[1] = tg->bps_conf[WRITE][index];
1576         v[2] = tg->iops_conf[READ][index];
1577         v[3] = tg->iops_conf[WRITE][index];
1578
1579         idle_time = tg->idletime_threshold_conf;
1580         latency_time = tg->latency_target_conf;
1581         while (true) {
1582                 char tok[27];   /* wiops=18446744073709551616 */
1583                 char *p;
1584                 u64 val = U64_MAX;
1585                 int len;
1586
1587                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1588                         break;
1589                 if (tok[0] == '\0')
1590                         break;
1591                 ctx.body += len;
1592
1593                 ret = -EINVAL;
1594                 p = tok;
1595                 strsep(&p, "=");
1596                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1597                         goto out_finish;
1598
1599                 ret = -ERANGE;
1600                 if (!val)
1601                         goto out_finish;
1602
1603                 ret = -EINVAL;
1604                 if (!strcmp(tok, "rbps") && val > 1)
1605                         v[0] = val;
1606                 else if (!strcmp(tok, "wbps") && val > 1)
1607                         v[1] = val;
1608                 else if (!strcmp(tok, "riops") && val > 1)
1609                         v[2] = min_t(u64, val, UINT_MAX);
1610                 else if (!strcmp(tok, "wiops") && val > 1)
1611                         v[3] = min_t(u64, val, UINT_MAX);
1612                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1613                         idle_time = val;
1614                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1615                         latency_time = val;
1616                 else
1617                         goto out_finish;
1618         }
1619
1620         tg->bps_conf[READ][index] = v[0];
1621         tg->bps_conf[WRITE][index] = v[1];
1622         tg->iops_conf[READ][index] = v[2];
1623         tg->iops_conf[WRITE][index] = v[3];
1624
1625         if (index == LIMIT_MAX) {
1626                 tg->bps[READ][index] = v[0];
1627                 tg->bps[WRITE][index] = v[1];
1628                 tg->iops[READ][index] = v[2];
1629                 tg->iops[WRITE][index] = v[3];
1630         }
1631         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1632                 tg->bps_conf[READ][LIMIT_MAX]);
1633         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1634                 tg->bps_conf[WRITE][LIMIT_MAX]);
1635         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1636                 tg->iops_conf[READ][LIMIT_MAX]);
1637         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1638                 tg->iops_conf[WRITE][LIMIT_MAX]);
1639         tg->idletime_threshold_conf = idle_time;
1640         tg->latency_target_conf = latency_time;
1641
1642         /* force user to configure all settings for low limit  */
1643         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1644               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1645             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1646             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1647                 tg->bps[READ][LIMIT_LOW] = 0;
1648                 tg->bps[WRITE][LIMIT_LOW] = 0;
1649                 tg->iops[READ][LIMIT_LOW] = 0;
1650                 tg->iops[WRITE][LIMIT_LOW] = 0;
1651                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1652                 tg->latency_target = DFL_LATENCY_TARGET;
1653         } else if (index == LIMIT_LOW) {
1654                 tg->idletime_threshold = tg->idletime_threshold_conf;
1655                 tg->latency_target = tg->latency_target_conf;
1656         }
1657
1658         blk_throtl_update_limit_valid(tg->td);
1659         if (tg->td->limit_valid[LIMIT_LOW]) {
1660                 if (index == LIMIT_LOW)
1661                         tg->td->limit_index = LIMIT_LOW;
1662         } else
1663                 tg->td->limit_index = LIMIT_MAX;
1664         tg_conf_updated(tg, index == LIMIT_LOW &&
1665                 tg->td->limit_valid[LIMIT_LOW]);
1666         ret = 0;
1667 out_finish:
1668         blkg_conf_exit(&ctx);
1669         return ret ?: nbytes;
1670 }
1671
1672 static struct cftype throtl_files[] = {
1673 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1674         {
1675                 .name = "low",
1676                 .flags = CFTYPE_NOT_ON_ROOT,
1677                 .seq_show = tg_print_limit,
1678                 .write = tg_set_limit,
1679                 .private = LIMIT_LOW,
1680         },
1681 #endif
1682         {
1683                 .name = "max",
1684                 .flags = CFTYPE_NOT_ON_ROOT,
1685                 .seq_show = tg_print_limit,
1686                 .write = tg_set_limit,
1687                 .private = LIMIT_MAX,
1688         },
1689         { }     /* terminate */
1690 };
1691
1692 static void throtl_shutdown_wq(struct request_queue *q)
1693 {
1694         struct throtl_data *td = q->td;
1695
1696         cancel_work_sync(&td->dispatch_work);
1697 }
1698
1699 struct blkcg_policy blkcg_policy_throtl = {
1700         .dfl_cftypes            = throtl_files,
1701         .legacy_cftypes         = throtl_legacy_files,
1702
1703         .pd_alloc_fn            = throtl_pd_alloc,
1704         .pd_init_fn             = throtl_pd_init,
1705         .pd_online_fn           = throtl_pd_online,
1706         .pd_offline_fn          = throtl_pd_offline,
1707         .pd_free_fn             = throtl_pd_free,
1708 };
1709
1710 void blk_throtl_cancel_bios(struct gendisk *disk)
1711 {
1712         struct request_queue *q = disk->queue;
1713         struct cgroup_subsys_state *pos_css;
1714         struct blkcg_gq *blkg;
1715
1716         spin_lock_irq(&q->queue_lock);
1717         /*
1718          * queue_lock is held, rcu lock is not needed here technically.
1719          * However, rcu lock is still held to emphasize that following
1720          * path need RCU protection and to prevent warning from lockdep.
1721          */
1722         rcu_read_lock();
1723         blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1724                 struct throtl_grp *tg = blkg_to_tg(blkg);
1725                 struct throtl_service_queue *sq = &tg->service_queue;
1726
1727                 /*
1728                  * Set the flag to make sure throtl_pending_timer_fn() won't
1729                  * stop until all throttled bios are dispatched.
1730                  */
1731                 tg->flags |= THROTL_TG_CANCELING;
1732
1733                 /*
1734                  * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1735                  * will be inserted to service queue without THROTL_TG_PENDING
1736                  * set in tg_update_disptime below. Then IO dispatched from
1737                  * child in tg_dispatch_one_bio will trigger double insertion
1738                  * and corrupt the tree.
1739                  */
1740                 if (!(tg->flags & THROTL_TG_PENDING))
1741                         continue;
1742
1743                 /*
1744                  * Update disptime after setting the above flag to make sure
1745                  * throtl_select_dispatch() won't exit without dispatching.
1746                  */
1747                 tg_update_disptime(tg);
1748
1749                 throtl_schedule_pending_timer(sq, jiffies + 1);
1750         }
1751         rcu_read_unlock();
1752         spin_unlock_irq(&q->queue_lock);
1753 }
1754
1755 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1756 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1757 {
1758         unsigned long rtime = jiffies, wtime = jiffies;
1759
1760         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1761                 rtime = tg->last_low_overflow_time[READ];
1762         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1763                 wtime = tg->last_low_overflow_time[WRITE];
1764         return min(rtime, wtime);
1765 }
1766
1767 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1768 {
1769         struct throtl_service_queue *parent_sq;
1770         struct throtl_grp *parent = tg;
1771         unsigned long ret = __tg_last_low_overflow_time(tg);
1772
1773         while (true) {
1774                 parent_sq = parent->service_queue.parent_sq;
1775                 parent = sq_to_tg(parent_sq);
1776                 if (!parent)
1777                         break;
1778
1779                 /*
1780                  * The parent doesn't have low limit, it always reaches low
1781                  * limit. Its overflow time is useless for children
1782                  */
1783                 if (!parent->bps[READ][LIMIT_LOW] &&
1784                     !parent->iops[READ][LIMIT_LOW] &&
1785                     !parent->bps[WRITE][LIMIT_LOW] &&
1786                     !parent->iops[WRITE][LIMIT_LOW])
1787                         continue;
1788                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1789                         ret = __tg_last_low_overflow_time(parent);
1790         }
1791         return ret;
1792 }
1793
1794 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1795 {
1796         /*
1797          * cgroup is idle if:
1798          * - single idle is too long, longer than a fixed value (in case user
1799          *   configure a too big threshold) or 4 times of idletime threshold
1800          * - average think time is more than threshold
1801          * - IO latency is largely below threshold
1802          */
1803         unsigned long time;
1804         bool ret;
1805
1806         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1807         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1808               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1809               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1810               tg->avg_idletime > tg->idletime_threshold ||
1811               (tg->latency_target && tg->bio_cnt &&
1812                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1813         throtl_log(&tg->service_queue,
1814                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1815                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1816                 tg->bio_cnt, ret, tg->td->scale);
1817         return ret;
1818 }
1819
1820 static bool throtl_low_limit_reached(struct throtl_grp *tg, int rw)
1821 {
1822         struct throtl_service_queue *sq = &tg->service_queue;
1823         bool limit = tg->bps[rw][LIMIT_LOW] || tg->iops[rw][LIMIT_LOW];
1824
1825         /*
1826          * if low limit is zero, low limit is always reached.
1827          * if low limit is non-zero, we can check if there is any request
1828          * is queued to determine if low limit is reached as we throttle
1829          * request according to limit.
1830          */
1831         return !limit || sq->nr_queued[rw];
1832 }
1833
1834 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1835 {
1836         /*
1837          * cgroup reaches low limit when low limit of READ and WRITE are
1838          * both reached, it's ok to upgrade to next limit if cgroup reaches
1839          * low limit
1840          */
1841         if (throtl_low_limit_reached(tg, READ) &&
1842             throtl_low_limit_reached(tg, WRITE))
1843                 return true;
1844
1845         if (time_after_eq(jiffies,
1846                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1847             throtl_tg_is_idle(tg))
1848                 return true;
1849         return false;
1850 }
1851
1852 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1853 {
1854         while (true) {
1855                 if (throtl_tg_can_upgrade(tg))
1856                         return true;
1857                 tg = sq_to_tg(tg->service_queue.parent_sq);
1858                 if (!tg || !tg_to_blkg(tg)->parent)
1859                         return false;
1860         }
1861         return false;
1862 }
1863
1864 static bool throtl_can_upgrade(struct throtl_data *td,
1865         struct throtl_grp *this_tg)
1866 {
1867         struct cgroup_subsys_state *pos_css;
1868         struct blkcg_gq *blkg;
1869
1870         if (td->limit_index != LIMIT_LOW)
1871                 return false;
1872
1873         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1874                 return false;
1875
1876         rcu_read_lock();
1877         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1878                 struct throtl_grp *tg = blkg_to_tg(blkg);
1879
1880                 if (tg == this_tg)
1881                         continue;
1882                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1883                         continue;
1884                 if (!throtl_hierarchy_can_upgrade(tg)) {
1885                         rcu_read_unlock();
1886                         return false;
1887                 }
1888         }
1889         rcu_read_unlock();
1890         return true;
1891 }
1892
1893 static void throtl_upgrade_check(struct throtl_grp *tg)
1894 {
1895         unsigned long now = jiffies;
1896
1897         if (tg->td->limit_index != LIMIT_LOW)
1898                 return;
1899
1900         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1901                 return;
1902
1903         tg->last_check_time = now;
1904
1905         if (!time_after_eq(now,
1906              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1907                 return;
1908
1909         if (throtl_can_upgrade(tg->td, NULL))
1910                 throtl_upgrade_state(tg->td);
1911 }
1912
1913 static void throtl_upgrade_state(struct throtl_data *td)
1914 {
1915         struct cgroup_subsys_state *pos_css;
1916         struct blkcg_gq *blkg;
1917
1918         throtl_log(&td->service_queue, "upgrade to max");
1919         td->limit_index = LIMIT_MAX;
1920         td->low_upgrade_time = jiffies;
1921         td->scale = 0;
1922         rcu_read_lock();
1923         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1924                 struct throtl_grp *tg = blkg_to_tg(blkg);
1925                 struct throtl_service_queue *sq = &tg->service_queue;
1926
1927                 tg->disptime = jiffies - 1;
1928                 throtl_select_dispatch(sq);
1929                 throtl_schedule_next_dispatch(sq, true);
1930         }
1931         rcu_read_unlock();
1932         throtl_select_dispatch(&td->service_queue);
1933         throtl_schedule_next_dispatch(&td->service_queue, true);
1934         queue_work(kthrotld_workqueue, &td->dispatch_work);
1935 }
1936
1937 static void throtl_downgrade_state(struct throtl_data *td)
1938 {
1939         td->scale /= 2;
1940
1941         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1942         if (td->scale) {
1943                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1944                 return;
1945         }
1946
1947         td->limit_index = LIMIT_LOW;
1948         td->low_downgrade_time = jiffies;
1949 }
1950
1951 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1952 {
1953         struct throtl_data *td = tg->td;
1954         unsigned long now = jiffies;
1955
1956         /*
1957          * If cgroup is below low limit, consider downgrade and throttle other
1958          * cgroups
1959          */
1960         if (time_after_eq(now, tg_last_low_overflow_time(tg) +
1961                                         td->throtl_slice) &&
1962             (!throtl_tg_is_idle(tg) ||
1963              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1964                 return true;
1965         return false;
1966 }
1967
1968 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1969 {
1970         struct throtl_data *td = tg->td;
1971
1972         if (time_before(jiffies, td->low_upgrade_time + td->throtl_slice))
1973                 return false;
1974
1975         while (true) {
1976                 if (!throtl_tg_can_downgrade(tg))
1977                         return false;
1978                 tg = sq_to_tg(tg->service_queue.parent_sq);
1979                 if (!tg || !tg_to_blkg(tg)->parent)
1980                         break;
1981         }
1982         return true;
1983 }
1984
1985 static void throtl_downgrade_check(struct throtl_grp *tg)
1986 {
1987         uint64_t bps;
1988         unsigned int iops;
1989         unsigned long elapsed_time;
1990         unsigned long now = jiffies;
1991
1992         if (tg->td->limit_index != LIMIT_MAX ||
1993             !tg->td->limit_valid[LIMIT_LOW])
1994                 return;
1995         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1996                 return;
1997         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1998                 return;
1999
2000         elapsed_time = now - tg->last_check_time;
2001         tg->last_check_time = now;
2002
2003         if (time_before(now, tg_last_low_overflow_time(tg) +
2004                         tg->td->throtl_slice))
2005                 return;
2006
2007         if (tg->bps[READ][LIMIT_LOW]) {
2008                 bps = tg->last_bytes_disp[READ] * HZ;
2009                 do_div(bps, elapsed_time);
2010                 if (bps >= tg->bps[READ][LIMIT_LOW])
2011                         tg->last_low_overflow_time[READ] = now;
2012         }
2013
2014         if (tg->bps[WRITE][LIMIT_LOW]) {
2015                 bps = tg->last_bytes_disp[WRITE] * HZ;
2016                 do_div(bps, elapsed_time);
2017                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2018                         tg->last_low_overflow_time[WRITE] = now;
2019         }
2020
2021         if (tg->iops[READ][LIMIT_LOW]) {
2022                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2023                 if (iops >= tg->iops[READ][LIMIT_LOW])
2024                         tg->last_low_overflow_time[READ] = now;
2025         }
2026
2027         if (tg->iops[WRITE][LIMIT_LOW]) {
2028                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2029                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2030                         tg->last_low_overflow_time[WRITE] = now;
2031         }
2032
2033         /*
2034          * If cgroup is below low limit, consider downgrade and throttle other
2035          * cgroups
2036          */
2037         if (throtl_hierarchy_can_downgrade(tg))
2038                 throtl_downgrade_state(tg->td);
2039
2040         tg->last_bytes_disp[READ] = 0;
2041         tg->last_bytes_disp[WRITE] = 0;
2042         tg->last_io_disp[READ] = 0;
2043         tg->last_io_disp[WRITE] = 0;
2044 }
2045
2046 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2047 {
2048         unsigned long now;
2049         unsigned long last_finish_time = tg->last_finish_time;
2050
2051         if (last_finish_time == 0)
2052                 return;
2053
2054         now = ktime_get_ns() >> 10;
2055         if (now <= last_finish_time ||
2056             last_finish_time == tg->checked_last_finish_time)
2057                 return;
2058
2059         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2060         tg->checked_last_finish_time = last_finish_time;
2061 }
2062
2063 static void throtl_update_latency_buckets(struct throtl_data *td)
2064 {
2065         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2066         int i, cpu, rw;
2067         unsigned long last_latency[2] = { 0 };
2068         unsigned long latency[2];
2069
2070         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2071                 return;
2072         if (time_before(jiffies, td->last_calculate_time + HZ))
2073                 return;
2074         td->last_calculate_time = jiffies;
2075
2076         memset(avg_latency, 0, sizeof(avg_latency));
2077         for (rw = READ; rw <= WRITE; rw++) {
2078                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2079                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2080
2081                         for_each_possible_cpu(cpu) {
2082                                 struct latency_bucket *bucket;
2083
2084                                 /* this isn't race free, but ok in practice */
2085                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2086                                         cpu);
2087                                 tmp->total_latency += bucket[i].total_latency;
2088                                 tmp->samples += bucket[i].samples;
2089                                 bucket[i].total_latency = 0;
2090                                 bucket[i].samples = 0;
2091                         }
2092
2093                         if (tmp->samples >= 32) {
2094                                 int samples = tmp->samples;
2095
2096                                 latency[rw] = tmp->total_latency;
2097
2098                                 tmp->total_latency = 0;
2099                                 tmp->samples = 0;
2100                                 latency[rw] /= samples;
2101                                 if (latency[rw] == 0)
2102                                         continue;
2103                                 avg_latency[rw][i].latency = latency[rw];
2104                         }
2105                 }
2106         }
2107
2108         for (rw = READ; rw <= WRITE; rw++) {
2109                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2110                         if (!avg_latency[rw][i].latency) {
2111                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2112                                         td->avg_buckets[rw][i].latency =
2113                                                 last_latency[rw];
2114                                 continue;
2115                         }
2116
2117                         if (!td->avg_buckets[rw][i].valid)
2118                                 latency[rw] = avg_latency[rw][i].latency;
2119                         else
2120                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2121                                         avg_latency[rw][i].latency) >> 3;
2122
2123                         td->avg_buckets[rw][i].latency = max(latency[rw],
2124                                 last_latency[rw]);
2125                         td->avg_buckets[rw][i].valid = true;
2126                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2127                 }
2128         }
2129
2130         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2131                 throtl_log(&td->service_queue,
2132                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2133                         "write latency=%ld, write valid=%d", i,
2134                         td->avg_buckets[READ][i].latency,
2135                         td->avg_buckets[READ][i].valid,
2136                         td->avg_buckets[WRITE][i].latency,
2137                         td->avg_buckets[WRITE][i].valid);
2138 }
2139 #else
2140 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2141 {
2142 }
2143
2144 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2145 {
2146 }
2147
2148 static void throtl_downgrade_check(struct throtl_grp *tg)
2149 {
2150 }
2151
2152 static void throtl_upgrade_check(struct throtl_grp *tg)
2153 {
2154 }
2155
2156 static bool throtl_can_upgrade(struct throtl_data *td,
2157         struct throtl_grp *this_tg)
2158 {
2159         return false;
2160 }
2161
2162 static void throtl_upgrade_state(struct throtl_data *td)
2163 {
2164 }
2165 #endif
2166
2167 bool __blk_throtl_bio(struct bio *bio)
2168 {
2169         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2170         struct blkcg_gq *blkg = bio->bi_blkg;
2171         struct throtl_qnode *qn = NULL;
2172         struct throtl_grp *tg = blkg_to_tg(blkg);
2173         struct throtl_service_queue *sq;
2174         bool rw = bio_data_dir(bio);
2175         bool throttled = false;
2176         struct throtl_data *td = tg->td;
2177
2178         rcu_read_lock();
2179
2180         spin_lock_irq(&q->queue_lock);
2181
2182         throtl_update_latency_buckets(td);
2183
2184         blk_throtl_update_idletime(tg);
2185
2186         sq = &tg->service_queue;
2187
2188 again:
2189         while (true) {
2190                 if (tg->last_low_overflow_time[rw] == 0)
2191                         tg->last_low_overflow_time[rw] = jiffies;
2192                 throtl_downgrade_check(tg);
2193                 throtl_upgrade_check(tg);
2194                 /* throtl is FIFO - if bios are already queued, should queue */
2195                 if (sq->nr_queued[rw])
2196                         break;
2197
2198                 /* if above limits, break to queue */
2199                 if (!tg_may_dispatch(tg, bio, NULL)) {
2200                         tg->last_low_overflow_time[rw] = jiffies;
2201                         if (throtl_can_upgrade(td, tg)) {
2202                                 throtl_upgrade_state(td);
2203                                 goto again;
2204                         }
2205                         break;
2206                 }
2207
2208                 /* within limits, let's charge and dispatch directly */
2209                 throtl_charge_bio(tg, bio);
2210
2211                 /*
2212                  * We need to trim slice even when bios are not being queued
2213                  * otherwise it might happen that a bio is not queued for
2214                  * a long time and slice keeps on extending and trim is not
2215                  * called for a long time. Now if limits are reduced suddenly
2216                  * we take into account all the IO dispatched so far at new
2217                  * low rate and * newly queued IO gets a really long dispatch
2218                  * time.
2219                  *
2220                  * So keep on trimming slice even if bio is not queued.
2221                  */
2222                 throtl_trim_slice(tg, rw);
2223
2224                 /*
2225                  * @bio passed through this layer without being throttled.
2226                  * Climb up the ladder.  If we're already at the top, it
2227                  * can be executed directly.
2228                  */
2229                 qn = &tg->qnode_on_parent[rw];
2230                 sq = sq->parent_sq;
2231                 tg = sq_to_tg(sq);
2232                 if (!tg) {
2233                         bio_set_flag(bio, BIO_BPS_THROTTLED);
2234                         goto out_unlock;
2235                 }
2236         }
2237
2238         /* out-of-limit, queue to @tg */
2239         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2240                    rw == READ ? 'R' : 'W',
2241                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2242                    tg_bps_limit(tg, rw),
2243                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2244                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2245
2246         tg->last_low_overflow_time[rw] = jiffies;
2247
2248         td->nr_queued[rw]++;
2249         throtl_add_bio_tg(bio, qn, tg);
2250         throttled = true;
2251
2252         /*
2253          * Update @tg's dispatch time and force schedule dispatch if @tg
2254          * was empty before @bio.  The forced scheduling isn't likely to
2255          * cause undue delay as @bio is likely to be dispatched directly if
2256          * its @tg's disptime is not in the future.
2257          */
2258         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2259                 tg_update_disptime(tg);
2260                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2261         }
2262
2263 out_unlock:
2264 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2265         if (throttled || !td->track_bio_latency)
2266                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2267 #endif
2268         spin_unlock_irq(&q->queue_lock);
2269
2270         rcu_read_unlock();
2271         return throttled;
2272 }
2273
2274 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2275 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2276                                  enum req_op op, unsigned long time)
2277 {
2278         const bool rw = op_is_write(op);
2279         struct latency_bucket *latency;
2280         int index;
2281
2282         if (!td || td->limit_index != LIMIT_LOW ||
2283             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2284             !blk_queue_nonrot(td->queue))
2285                 return;
2286
2287         index = request_bucket_index(size);
2288
2289         latency = get_cpu_ptr(td->latency_buckets[rw]);
2290         latency[index].total_latency += time;
2291         latency[index].samples++;
2292         put_cpu_ptr(td->latency_buckets[rw]);
2293 }
2294
2295 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2296 {
2297         struct request_queue *q = rq->q;
2298         struct throtl_data *td = q->td;
2299
2300         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2301                              time_ns >> 10);
2302 }
2303
2304 void blk_throtl_bio_endio(struct bio *bio)
2305 {
2306         struct blkcg_gq *blkg;
2307         struct throtl_grp *tg;
2308         u64 finish_time_ns;
2309         unsigned long finish_time;
2310         unsigned long start_time;
2311         unsigned long lat;
2312         int rw = bio_data_dir(bio);
2313
2314         blkg = bio->bi_blkg;
2315         if (!blkg)
2316                 return;
2317         tg = blkg_to_tg(blkg);
2318         if (!tg->td->limit_valid[LIMIT_LOW])
2319                 return;
2320
2321         finish_time_ns = ktime_get_ns();
2322         tg->last_finish_time = finish_time_ns >> 10;
2323
2324         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2325         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2326         if (!start_time || finish_time <= start_time)
2327                 return;
2328
2329         lat = finish_time - start_time;
2330         /* this is only for bio based driver */
2331         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2332                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2333                                      bio_op(bio), lat);
2334
2335         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2336                 int bucket;
2337                 unsigned int threshold;
2338
2339                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2340                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2341                         tg->latency_target;
2342                 if (lat > threshold)
2343                         tg->bad_bio_cnt++;
2344                 /*
2345                  * Not race free, could get wrong count, which means cgroups
2346                  * will be throttled
2347                  */
2348                 tg->bio_cnt++;
2349         }
2350
2351         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2352                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2353                 tg->bio_cnt /= 2;
2354                 tg->bad_bio_cnt /= 2;
2355         }
2356 }
2357 #endif
2358
2359 int blk_throtl_init(struct gendisk *disk)
2360 {
2361         struct request_queue *q = disk->queue;
2362         struct throtl_data *td;
2363         int ret;
2364
2365         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2366         if (!td)
2367                 return -ENOMEM;
2368         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2369                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2370         if (!td->latency_buckets[READ]) {
2371                 kfree(td);
2372                 return -ENOMEM;
2373         }
2374         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2375                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2376         if (!td->latency_buckets[WRITE]) {
2377                 free_percpu(td->latency_buckets[READ]);
2378                 kfree(td);
2379                 return -ENOMEM;
2380         }
2381
2382         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2383         throtl_service_queue_init(&td->service_queue);
2384
2385         q->td = td;
2386         td->queue = q;
2387
2388         td->limit_valid[LIMIT_MAX] = true;
2389         td->limit_index = LIMIT_MAX;
2390         td->low_upgrade_time = jiffies;
2391         td->low_downgrade_time = jiffies;
2392
2393         /* activate policy */
2394         ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
2395         if (ret) {
2396                 free_percpu(td->latency_buckets[READ]);
2397                 free_percpu(td->latency_buckets[WRITE]);
2398                 kfree(td);
2399         }
2400         return ret;
2401 }
2402
2403 void blk_throtl_exit(struct gendisk *disk)
2404 {
2405         struct request_queue *q = disk->queue;
2406
2407         BUG_ON(!q->td);
2408         del_timer_sync(&q->td->service_queue.pending_timer);
2409         throtl_shutdown_wq(q);
2410         blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
2411         free_percpu(q->td->latency_buckets[READ]);
2412         free_percpu(q->td->latency_buckets[WRITE]);
2413         kfree(q->td);
2414 }
2415
2416 void blk_throtl_register(struct gendisk *disk)
2417 {
2418         struct request_queue *q = disk->queue;
2419         struct throtl_data *td;
2420         int i;
2421
2422         td = q->td;
2423         BUG_ON(!td);
2424
2425         if (blk_queue_nonrot(q)) {
2426                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2427                 td->filtered_latency = LATENCY_FILTERED_SSD;
2428         } else {
2429                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2430                 td->filtered_latency = LATENCY_FILTERED_HD;
2431                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2432                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2433                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2434                 }
2435         }
2436 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2437         /* if no low limit, use previous default */
2438         td->throtl_slice = DFL_THROTL_SLICE_HD;
2439
2440 #else
2441         td->track_bio_latency = !queue_is_mq(q);
2442         if (!td->track_bio_latency)
2443                 blk_stat_enable_accounting(q);
2444 #endif
2445 }
2446
2447 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2448 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2449 {
2450         if (!q->td)
2451                 return -EINVAL;
2452         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2453 }
2454
2455 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2456         const char *page, size_t count)
2457 {
2458         unsigned long v;
2459         unsigned long t;
2460
2461         if (!q->td)
2462                 return -EINVAL;
2463         if (kstrtoul(page, 10, &v))
2464                 return -EINVAL;
2465         t = msecs_to_jiffies(v);
2466         if (t == 0 || t > MAX_THROTL_SLICE)
2467                 return -EINVAL;
2468         q->td->throtl_slice = t;
2469         return count;
2470 }
2471 #endif
2472
2473 static int __init throtl_init(void)
2474 {
2475         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2476         if (!kthrotld_workqueue)
2477                 panic("Failed to create kthrotld\n");
2478
2479         return blkcg_policy_register(&blkcg_policy_throtl);
2480 }
2481
2482 module_init(throtl_init);