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: %llu %llu %u %u\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         unsigned 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 (tg->io_disp[rw] + 1 <= io_allowed) {
 842                 return 0;
 843         }
 844
 845         /* Calc approx time to dispatch */
 846         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 847         return jiffy_wait;
 848 }
 849
 850 static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
 851                                 u64 bps_limit)
 852 {
 853         bool rw = bio_data_dir(bio);
 854         u64 bytes_allowed, 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 (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 873                 return 0;
 874         }
 875
 876         /* Calc approx time to dispatch */
 877         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 878         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 879
 880         if (!jiffy_wait)
 881                 jiffy_wait = 1;
 882
 883         /*
 884          * This wait time is without taking into consideration the rounding
 885          * up we did. Add that time also.
 886          */
 887         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 888         return jiffy_wait;
 889 }
 890
 891 /*
 892  * Returns whether one can dispatch a bio or not. Also returns approx number
 893  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 894  */
 895 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 896                             unsigned long *wait)
 897 {
 898         bool rw = bio_data_dir(bio);
 899         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 900         u64 bps_limit = tg_bps_limit(tg, rw);
 901         u32 iops_limit = tg_iops_limit(tg, rw);
 902
 903         /*
 904          * Currently whole state machine of group depends on first bio
 905          * queued in the group bio list. So one should not be calling
 906          * this function with a different bio if there are other bios
 907          * queued.
 908          */
 909         BUG_ON(tg->service_queue.nr_queued[rw] &&
 910                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 911
 912         /* If tg->bps = -1, then BW is unlimited */
 913         if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
 914             tg->flags & THROTL_TG_CANCELING) {
 915                 if (wait)
 916                         *wait = 0;
 917                 return true;
 918         }
 919
 920         /*
 921          * If previous slice expired, start a new one otherwise renew/extend
 922          * existing slice to make sure it is at least throtl_slice interval
 923          * long since now. New slice is started only for empty throttle group.
 924          * If there is queued bio, that means there should be an active
 925          * slice and it should be extended instead.
 926          */
 927         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 928                 throtl_start_new_slice(tg, rw, true);
 929         else {
 930                 if (time_before(tg->slice_end[rw],
 931                     jiffies + tg->td->throtl_slice))
 932                         throtl_extend_slice(tg, rw,
 933                                 jiffies + tg->td->throtl_slice);
 934         }
 935
 936         bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
 937         iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
 938         if (bps_wait + iops_wait == 0) {
 939                 if (wait)
 940                         *wait = 0;
 941                 return true;
 942         }
 943
 944         max_wait = max(bps_wait, iops_wait);
 945
 946         if (wait)
 947                 *wait = max_wait;
 948
 949         if (time_before(tg->slice_end[rw], jiffies + max_wait))
 950                 throtl_extend_slice(tg, rw, jiffies + max_wait);
 951
 952         return false;
 953 }
 954
 955 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 956 {
 957         bool rw = bio_data_dir(bio);
 958         unsigned int bio_size = throtl_bio_data_size(bio);
 959
 960         /* Charge the bio to the group */
 961         if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
 962                 tg->bytes_disp[rw] += bio_size;
 963                 tg->last_bytes_disp[rw] += bio_size;
 964         }
 965
 966         tg->io_disp[rw]++;
 967         tg->last_io_disp[rw]++;
 968 }
 969
 970 /**
 971  * throtl_add_bio_tg - add a bio to the specified throtl_grp
 972  * @bio: bio to add
 973  * @qn: qnode to use
 974  * @tg: the target throtl_grp
 975  *
 976  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 977  * tg->qnode_on_self[] is used.
 978  */
 979 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 980                               struct throtl_grp *tg)
 981 {
 982         struct throtl_service_queue *sq = &tg->service_queue;
 983         bool rw = bio_data_dir(bio);
 984
 985         if (!qn)
 986                 qn = &tg->qnode_on_self[rw];
 987
 988         /*
 989          * If @tg doesn't currently have any bios queued in the same
 990          * direction, queueing @bio can change when @tg should be
 991          * dispatched.  Mark that @tg was empty.  This is automatically
 992          * cleared on the next tg_update_disptime().
 993          */
 994         if (!sq->nr_queued[rw])
 995                 tg->flags |= THROTL_TG_WAS_EMPTY;
 996
 997         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
 998
 999         sq->nr_queued[rw]++;
1000         throtl_enqueue_tg(tg);
1001 }
1002
1003 static void tg_update_disptime(struct throtl_grp *tg)
1004 {
1005         struct throtl_service_queue *sq = &tg->service_queue;
1006         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1007         struct bio *bio;
1008
1009         bio = throtl_peek_queued(&sq->queued[READ]);
1010         if (bio)
1011                 tg_may_dispatch(tg, bio, &read_wait);
1012
1013         bio = throtl_peek_queued(&sq->queued[WRITE]);
1014         if (bio)
1015                 tg_may_dispatch(tg, bio, &write_wait);
1016
1017         min_wait = min(read_wait, write_wait);
1018         disptime = jiffies + min_wait;
1019
1020         /* Update dispatch time */
1021         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
1022         tg->disptime = disptime;
1023         tg_service_queue_add(tg);
1024
1025         /* see throtl_add_bio_tg() */
1026         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1027 }
1028
1029 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1030                                         struct throtl_grp *parent_tg, bool rw)
1031 {
1032         if (throtl_slice_used(parent_tg, rw)) {
1033                 throtl_start_new_slice_with_credit(parent_tg, rw,
1034                                 child_tg->slice_start[rw]);
1035         }
1036
1037 }
1038
1039 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1040 {
1041         struct throtl_service_queue *sq = &tg->service_queue;
1042         struct throtl_service_queue *parent_sq = sq->parent_sq;
1043         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1044         struct throtl_grp *tg_to_put = NULL;
1045         struct bio *bio;
1046
1047         /*
1048          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1049          * from @tg may put its reference and @parent_sq might end up
1050          * getting released prematurely.  Remember the tg to put and put it
1051          * after @bio is transferred to @parent_sq.
1052          */
1053         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1054         sq->nr_queued[rw]--;
1055
1056         throtl_charge_bio(tg, bio);
1057
1058         /*
1059          * If our parent is another tg, we just need to transfer @bio to
1060          * the parent using throtl_add_bio_tg().  If our parent is
1061          * @td->service_queue, @bio is ready to be issued.  Put it on its
1062          * bio_lists[] and decrease total number queued.  The caller is
1063          * responsible for issuing these bios.
1064          */
1065         if (parent_tg) {
1066                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1067                 start_parent_slice_with_credit(tg, parent_tg, rw);
1068         } else {
1069                 bio_set_flag(bio, BIO_BPS_THROTTLED);
1070                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1071                                      &parent_sq->queued[rw]);
1072                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1073                 tg->td->nr_queued[rw]--;
1074         }
1075
1076         throtl_trim_slice(tg, rw);
1077
1078         if (tg_to_put)
1079                 blkg_put(tg_to_blkg(tg_to_put));
1080 }
1081
1082 static int throtl_dispatch_tg(struct throtl_grp *tg)
1083 {
1084         struct throtl_service_queue *sq = &tg->service_queue;
1085         unsigned int nr_reads = 0, nr_writes = 0;
1086         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1087         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1088         struct bio *bio;
1089
1090         /* Try to dispatch 75% READS and 25% WRITES */
1091
1092         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1093                tg_may_dispatch(tg, bio, NULL)) {
1094
1095                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1096                 nr_reads++;
1097
1098                 if (nr_reads >= max_nr_reads)
1099                         break;
1100         }
1101
1102         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1103                tg_may_dispatch(tg, bio, NULL)) {
1104
1105                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1106                 nr_writes++;
1107
1108                 if (nr_writes >= max_nr_writes)
1109                         break;
1110         }
1111
1112         return nr_reads + nr_writes;
1113 }
1114
1115 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1116 {
1117         unsigned int nr_disp = 0;
1118
1119         while (1) {
1120                 struct throtl_grp *tg;
1121                 struct throtl_service_queue *sq;
1122
1123                 if (!parent_sq->nr_pending)
1124                         break;
1125
1126                 tg = throtl_rb_first(parent_sq);
1127                 if (!tg)
1128                         break;
1129
1130                 if (time_before(jiffies, tg->disptime))
1131                         break;
1132
1133                 nr_disp += throtl_dispatch_tg(tg);
1134
1135                 sq = &tg->service_queue;
1136                 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
1137                         tg_update_disptime(tg);
1138                 else
1139                         throtl_dequeue_tg(tg);
1140
1141                 if (nr_disp >= THROTL_QUANTUM)
1142                         break;
1143         }
1144
1145         return nr_disp;
1146 }
1147
1148 static bool throtl_can_upgrade(struct throtl_data *td,
1149         struct throtl_grp *this_tg);
1150 /**
1151  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1152  * @t: the pending_timer member of the throtl_service_queue being serviced
1153  *
1154  * This timer is armed when a child throtl_grp with active bio's become
1155  * pending and queued on the service_queue's pending_tree and expires when
1156  * the first child throtl_grp should be dispatched.  This function
1157  * dispatches bio's from the children throtl_grps to the parent
1158  * service_queue.
1159  *
1160  * If the parent's parent is another throtl_grp, dispatching is propagated
1161  * by either arming its pending_timer or repeating dispatch directly.  If
1162  * the top-level service_tree is reached, throtl_data->dispatch_work is
1163  * kicked so that the ready bio's are issued.
1164  */
1165 static void throtl_pending_timer_fn(struct timer_list *t)
1166 {
1167         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1168         struct throtl_grp *tg = sq_to_tg(sq);
1169         struct throtl_data *td = sq_to_td(sq);
1170         struct throtl_service_queue *parent_sq;
1171         struct request_queue *q;
1172         bool dispatched;
1173         int ret;
1174
1175         /* throtl_data may be gone, so figure out request queue by blkg */
1176         if (tg)
1177                 q = tg->pd.blkg->q;
1178         else
1179                 q = td->queue;
1180
1181         spin_lock_irq(&q->queue_lock);
1182
1183         if (!q->root_blkg)
1184                 goto out_unlock;
1185
1186         if (throtl_can_upgrade(td, NULL))
1187                 throtl_upgrade_state(td);
1188
1189 again:
1190         parent_sq = sq->parent_sq;
1191         dispatched = false;
1192
1193         while (true) {
1194                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1195                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1196                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1197
1198                 ret = throtl_select_dispatch(sq);
1199                 if (ret) {
1200                         throtl_log(sq, "bios disp=%u", ret);
1201                         dispatched = true;
1202                 }
1203
1204                 if (throtl_schedule_next_dispatch(sq, false))
1205                         break;
1206
1207                 /* this dispatch windows is still open, relax and repeat */
1208                 spin_unlock_irq(&q->queue_lock);
1209                 cpu_relax();
1210                 spin_lock_irq(&q->queue_lock);
1211         }
1212
1213         if (!dispatched)
1214                 goto out_unlock;
1215
1216         if (parent_sq) {
1217                 /* @parent_sq is another throl_grp, propagate dispatch */
1218                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1219                         tg_update_disptime(tg);
1220                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1221                                 /* window is already open, repeat dispatching */
1222                                 sq = parent_sq;
1223                                 tg = sq_to_tg(sq);
1224                                 goto again;
1225                         }
1226                 }
1227         } else {
1228                 /* reached the top-level, queue issuing */
1229                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1230         }
1231 out_unlock:
1232         spin_unlock_irq(&q->queue_lock);
1233 }
1234
1235 /**
1236  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1237  * @work: work item being executed
1238  *
1239  * This function is queued for execution when bios reach the bio_lists[]
1240  * of throtl_data->service_queue.  Those bios are ready and issued by this
1241  * function.
1242  */
1243 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1244 {
1245         struct throtl_data *td = container_of(work, struct throtl_data,
1246                                               dispatch_work);
1247         struct throtl_service_queue *td_sq = &td->service_queue;
1248         struct request_queue *q = td->queue;
1249         struct bio_list bio_list_on_stack;
1250         struct bio *bio;
1251         struct blk_plug plug;
1252         int rw;
1253
1254         bio_list_init(&bio_list_on_stack);
1255
1256         spin_lock_irq(&q->queue_lock);
1257         for (rw = READ; rw <= WRITE; rw++)
1258                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1259                         bio_list_add(&bio_list_on_stack, bio);
1260         spin_unlock_irq(&q->queue_lock);
1261
1262         if (!bio_list_empty(&bio_list_on_stack)) {
1263                 blk_start_plug(&plug);
1264                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1265                         submit_bio_noacct_nocheck(bio);
1266                 blk_finish_plug(&plug);
1267         }
1268 }
1269
1270 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1271                               int off)
1272 {
1273         struct throtl_grp *tg = pd_to_tg(pd);
1274         u64 v = *(u64 *)((void *)tg + off);
1275
1276         if (v == U64_MAX)
1277                 return 0;
1278         return __blkg_prfill_u64(sf, pd, v);
1279 }
1280
1281 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1282                                int off)
1283 {
1284         struct throtl_grp *tg = pd_to_tg(pd);
1285         unsigned int v = *(unsigned int *)((void *)tg + off);
1286
1287         if (v == UINT_MAX)
1288                 return 0;
1289         return __blkg_prfill_u64(sf, pd, v);
1290 }
1291
1292 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1293 {
1294         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1295                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1296         return 0;
1297 }
1298
1299 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1300 {
1301         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1302                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1303         return 0;
1304 }
1305
1306 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1307 {
1308         struct throtl_service_queue *sq = &tg->service_queue;
1309         struct cgroup_subsys_state *pos_css;
1310         struct blkcg_gq *blkg;
1311
1312         throtl_log(&tg->service_queue,
1313                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1314                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1315                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1316
1317         /*
1318          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1319          * considered to have rules if either the tg itself or any of its
1320          * ancestors has rules.  This identifies groups without any
1321          * restrictions in the whole hierarchy and allows them to bypass
1322          * blk-throttle.
1323          */
1324         blkg_for_each_descendant_pre(blkg, pos_css,
1325                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1326                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1327                 struct throtl_grp *parent_tg;
1328
1329                 tg_update_has_rules(this_tg);
1330                 /* ignore root/second level */
1331                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1332                     !blkg->parent->parent)
1333                         continue;
1334                 parent_tg = blkg_to_tg(blkg->parent);
1335                 /*
1336                  * make sure all children has lower idle time threshold and
1337                  * higher latency target
1338                  */
1339                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1340                                 parent_tg->idletime_threshold);
1341                 this_tg->latency_target = max(this_tg->latency_target,
1342                                 parent_tg->latency_target);
1343         }
1344
1345         /*
1346          * We're already holding queue_lock and know @tg is valid.  Let's
1347          * apply the new config directly.
1348          *
1349          * Restart the slices for both READ and WRITES. It might happen
1350          * that a group's limit are dropped suddenly and we don't want to
1351          * account recently dispatched IO with new low rate.
1352          */
1353         throtl_start_new_slice(tg, READ, false);
1354         throtl_start_new_slice(tg, WRITE, false);
1355
1356         if (tg->flags & THROTL_TG_PENDING) {
1357                 tg_update_disptime(tg);
1358                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1359         }
1360 }
1361
1362 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1363                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1364 {
1365         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1366         struct blkg_conf_ctx ctx;
1367         struct throtl_grp *tg;
1368         int ret;
1369         u64 v;
1370
1371         blkg_conf_init(&ctx, buf);
1372
1373         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1374         if (ret)
1375                 goto out_finish;
1376
1377         ret = -EINVAL;
1378         if (sscanf(ctx.body, "%llu", &v) != 1)
1379                 goto out_finish;
1380         if (!v)
1381                 v = U64_MAX;
1382
1383         tg = blkg_to_tg(ctx.blkg);
1384         tg_update_carryover(tg);
1385
1386         if (is_u64)
1387                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1388         else
1389                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1390
1391         tg_conf_updated(tg, false);
1392         ret = 0;
1393 out_finish:
1394         blkg_conf_exit(&ctx);
1395         return ret ?: nbytes;
1396 }
1397
1398 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1399                                char *buf, size_t nbytes, loff_t off)
1400 {
1401         return tg_set_conf(of, buf, nbytes, off, true);
1402 }
1403
1404 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1405                                 char *buf, size_t nbytes, loff_t off)
1406 {
1407         return tg_set_conf(of, buf, nbytes, off, false);
1408 }
1409
1410 static int tg_print_rwstat(struct seq_file *sf, void *v)
1411 {
1412         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1413                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1414                           seq_cft(sf)->private, true);
1415         return 0;
1416 }
1417
1418 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1419                                       struct blkg_policy_data *pd, int off)
1420 {
1421         struct blkg_rwstat_sample sum;
1422
1423         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1424                                   &sum);
1425         return __blkg_prfill_rwstat(sf, pd, &sum);
1426 }
1427
1428 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1429 {
1430         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1431                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1432                           seq_cft(sf)->private, true);
1433         return 0;
1434 }
1435
1436 static struct cftype throtl_legacy_files[] = {
1437         {
1438                 .name = "throttle.read_bps_device",
1439                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1440                 .seq_show = tg_print_conf_u64,
1441                 .write = tg_set_conf_u64,
1442         },
1443         {
1444                 .name = "throttle.write_bps_device",
1445                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1446                 .seq_show = tg_print_conf_u64,
1447                 .write = tg_set_conf_u64,
1448         },
1449         {
1450                 .name = "throttle.read_iops_device",
1451                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1452                 .seq_show = tg_print_conf_uint,
1453                 .write = tg_set_conf_uint,
1454         },
1455         {
1456                 .name = "throttle.write_iops_device",
1457                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1458                 .seq_show = tg_print_conf_uint,
1459                 .write = tg_set_conf_uint,
1460         },
1461         {
1462                 .name = "throttle.io_service_bytes",
1463                 .private = offsetof(struct throtl_grp, stat_bytes),
1464                 .seq_show = tg_print_rwstat,
1465         },
1466         {
1467                 .name = "throttle.io_service_bytes_recursive",
1468                 .private = offsetof(struct throtl_grp, stat_bytes),
1469                 .seq_show = tg_print_rwstat_recursive,
1470         },
1471         {
1472                 .name = "throttle.io_serviced",
1473                 .private = offsetof(struct throtl_grp, stat_ios),
1474                 .seq_show = tg_print_rwstat,
1475         },
1476         {
1477                 .name = "throttle.io_serviced_recursive",
1478                 .private = offsetof(struct throtl_grp, stat_ios),
1479                 .seq_show = tg_print_rwstat_recursive,
1480         },
1481         { }     /* terminate */
1482 };
1483
1484 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1485                          int off)
1486 {
1487         struct throtl_grp *tg = pd_to_tg(pd);
1488         const char *dname = blkg_dev_name(pd->blkg);
1489         char bufs[4][21] = { "max", "max", "max", "max" };
1490         u64 bps_dft;
1491         unsigned int iops_dft;
1492         char idle_time[26] = "";
1493         char latency_time[26] = "";
1494
1495         if (!dname)
1496                 return 0;
1497
1498         if (off == LIMIT_LOW) {
1499                 bps_dft = 0;
1500                 iops_dft = 0;
1501         } else {
1502                 bps_dft = U64_MAX;
1503                 iops_dft = UINT_MAX;
1504         }
1505
1506         if (tg->bps_conf[READ][off] == bps_dft &&
1507             tg->bps_conf[WRITE][off] == bps_dft &&
1508             tg->iops_conf[READ][off] == iops_dft &&
1509             tg->iops_conf[WRITE][off] == iops_dft &&
1510             (off != LIMIT_LOW ||
1511              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1512               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1513                 return 0;
1514
1515         if (tg->bps_conf[READ][off] != U64_MAX)
1516                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1517                         tg->bps_conf[READ][off]);
1518         if (tg->bps_conf[WRITE][off] != U64_MAX)
1519                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1520                         tg->bps_conf[WRITE][off]);
1521         if (tg->iops_conf[READ][off] != UINT_MAX)
1522                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1523                         tg->iops_conf[READ][off]);
1524         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1525                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1526                         tg->iops_conf[WRITE][off]);
1527         if (off == LIMIT_LOW) {
1528                 if (tg->idletime_threshold_conf == ULONG_MAX)
1529                         strcpy(idle_time, " idle=max");
1530                 else
1531                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1532                                 tg->idletime_threshold_conf);
1533
1534                 if (tg->latency_target_conf == ULONG_MAX)
1535                         strcpy(latency_time, " latency=max");
1536                 else
1537                         snprintf(latency_time, sizeof(latency_time),
1538                                 " latency=%lu", tg->latency_target_conf);
1539         }
1540
1541         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1542                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1543                    latency_time);
1544         return 0;
1545 }
1546
1547 static int tg_print_limit(struct seq_file *sf, void *v)
1548 {
1549         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1550                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1551         return 0;
1552 }
1553
1554 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1555                           char *buf, size_t nbytes, loff_t off)
1556 {
1557         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1558         struct blkg_conf_ctx ctx;
1559         struct throtl_grp *tg;
1560         u64 v[4];
1561         unsigned long idle_time;
1562         unsigned long latency_time;
1563         int ret;
1564         int index = of_cft(of)->private;
1565
1566         blkg_conf_init(&ctx, buf);
1567
1568         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1569         if (ret)
1570                 goto out_finish;
1571
1572         tg = blkg_to_tg(ctx.blkg);
1573         tg_update_carryover(tg);
1574
1575         v[0] = tg->bps_conf[READ][index];
1576         v[1] = tg->bps_conf[WRITE][index];
1577         v[2] = tg->iops_conf[READ][index];
1578         v[3] = tg->iops_conf[WRITE][index];
1579
1580         idle_time = tg->idletime_threshold_conf;
1581         latency_time = tg->latency_target_conf;
1582         while (true) {
1583                 char tok[27];   /* wiops=18446744073709551616 */
1584                 char *p;
1585                 u64 val = U64_MAX;
1586                 int len;
1587
1588                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1589                         break;
1590                 if (tok[0] == '\0')
1591                         break;
1592                 ctx.body += len;
1593
1594                 ret = -EINVAL;
1595                 p = tok;
1596                 strsep(&p, "=");
1597                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1598                         goto out_finish;
1599
1600                 ret = -ERANGE;
1601                 if (!val)
1602                         goto out_finish;
1603
1604                 ret = -EINVAL;
1605                 if (!strcmp(tok, "rbps") && val > 1)
1606                         v[0] = val;
1607                 else if (!strcmp(tok, "wbps") && val > 1)
1608                         v[1] = val;
1609                 else if (!strcmp(tok, "riops") && val > 1)
1610                         v[2] = min_t(u64, val, UINT_MAX);
1611                 else if (!strcmp(tok, "wiops") && val > 1)
1612                         v[3] = min_t(u64, val, UINT_MAX);
1613                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1614                         idle_time = val;
1615                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1616                         latency_time = val;
1617                 else
1618                         goto out_finish;
1619         }
1620
1621         tg->bps_conf[READ][index] = v[0];
1622         tg->bps_conf[WRITE][index] = v[1];
1623         tg->iops_conf[READ][index] = v[2];
1624         tg->iops_conf[WRITE][index] = v[3];
1625
1626         if (index == LIMIT_MAX) {
1627                 tg->bps[READ][index] = v[0];
1628                 tg->bps[WRITE][index] = v[1];
1629                 tg->iops[READ][index] = v[2];
1630                 tg->iops[WRITE][index] = v[3];
1631         }
1632         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1633                 tg->bps_conf[READ][LIMIT_MAX]);
1634         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1635                 tg->bps_conf[WRITE][LIMIT_MAX]);
1636         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1637                 tg->iops_conf[READ][LIMIT_MAX]);
1638         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1639                 tg->iops_conf[WRITE][LIMIT_MAX]);
1640         tg->idletime_threshold_conf = idle_time;
1641         tg->latency_target_conf = latency_time;
1642
1643         /* force user to configure all settings for low limit  */
1644         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1645               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1646             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1647             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1648                 tg->bps[READ][LIMIT_LOW] = 0;
1649                 tg->bps[WRITE][LIMIT_LOW] = 0;
1650                 tg->iops[READ][LIMIT_LOW] = 0;
1651                 tg->iops[WRITE][LIMIT_LOW] = 0;
1652                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1653                 tg->latency_target = DFL_LATENCY_TARGET;
1654         } else if (index == LIMIT_LOW) {
1655                 tg->idletime_threshold = tg->idletime_threshold_conf;
1656                 tg->latency_target = tg->latency_target_conf;
1657         }
1658
1659         blk_throtl_update_limit_valid(tg->td);
1660         if (tg->td->limit_valid[LIMIT_LOW]) {
1661                 if (index == LIMIT_LOW)
1662                         tg->td->limit_index = LIMIT_LOW;
1663         } else
1664                 tg->td->limit_index = LIMIT_MAX;
1665         tg_conf_updated(tg, index == LIMIT_LOW &&
1666                 tg->td->limit_valid[LIMIT_LOW]);
1667         ret = 0;
1668 out_finish:
1669         blkg_conf_exit(&ctx);
1670         return ret ?: nbytes;
1671 }
1672
1673 static struct cftype throtl_files[] = {
1674 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1675         {
1676                 .name = "low",
1677                 .flags = CFTYPE_NOT_ON_ROOT,
1678                 .seq_show = tg_print_limit,
1679                 .write = tg_set_limit,
1680                 .private = LIMIT_LOW,
1681         },
1682 #endif
1683         {
1684                 .name = "max",
1685                 .flags = CFTYPE_NOT_ON_ROOT,
1686                 .seq_show = tg_print_limit,
1687                 .write = tg_set_limit,
1688                 .private = LIMIT_MAX,
1689         },
1690         { }     /* terminate */
1691 };
1692
1693 static void throtl_shutdown_wq(struct request_queue *q)
1694 {
1695         struct throtl_data *td = q->td;
1696
1697         cancel_work_sync(&td->dispatch_work);
1698 }
1699
1700 struct blkcg_policy blkcg_policy_throtl = {
1701         .dfl_cftypes            = throtl_files,
1702         .legacy_cftypes         = throtl_legacy_files,
1703
1704         .pd_alloc_fn            = throtl_pd_alloc,
1705         .pd_init_fn             = throtl_pd_init,
1706         .pd_online_fn           = throtl_pd_online,
1707         .pd_offline_fn          = throtl_pd_offline,
1708         .pd_free_fn             = throtl_pd_free,
1709 };
1710
1711 void blk_throtl_cancel_bios(struct gendisk *disk)
1712 {
1713         struct request_queue *q = disk->queue;
1714         struct cgroup_subsys_state *pos_css;
1715         struct blkcg_gq *blkg;
1716
1717         spin_lock_irq(&q->queue_lock);
1718         /*
1719          * queue_lock is held, rcu lock is not needed here technically.
1720          * However, rcu lock is still held to emphasize that following
1721          * path need RCU protection and to prevent warning from lockdep.
1722          */
1723         rcu_read_lock();
1724         blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1725                 struct throtl_grp *tg = blkg_to_tg(blkg);
1726                 struct throtl_service_queue *sq = &tg->service_queue;
1727
1728                 /*
1729                  * Set the flag to make sure throtl_pending_timer_fn() won't
1730                  * stop until all throttled bios are dispatched.
1731                  */
1732                 tg->flags |= THROTL_TG_CANCELING;
1733
1734                 /*
1735                  * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1736                  * will be inserted to service queue without THROTL_TG_PENDING
1737                  * set in tg_update_disptime below. Then IO dispatched from
1738                  * child in tg_dispatch_one_bio will trigger double insertion
1739                  * and corrupt the tree.
1740                  */
1741                 if (!(tg->flags & THROTL_TG_PENDING))
1742                         continue;
1743
1744                 /*
1745                  * Update disptime after setting the above flag to make sure
1746                  * throtl_select_dispatch() won't exit without dispatching.
1747                  */
1748                 tg_update_disptime(tg);
1749
1750                 throtl_schedule_pending_timer(sq, jiffies + 1);
1751         }
1752         rcu_read_unlock();
1753         spin_unlock_irq(&q->queue_lock);
1754 }
1755
1756 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1757 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1758 {
1759         unsigned long rtime = jiffies, wtime = jiffies;
1760
1761         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1762                 rtime = tg->last_low_overflow_time[READ];
1763         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1764                 wtime = tg->last_low_overflow_time[WRITE];
1765         return min(rtime, wtime);
1766 }
1767
1768 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1769 {
1770         struct throtl_service_queue *parent_sq;
1771         struct throtl_grp *parent = tg;
1772         unsigned long ret = __tg_last_low_overflow_time(tg);
1773
1774         while (true) {
1775                 parent_sq = parent->service_queue.parent_sq;
1776                 parent = sq_to_tg(parent_sq);
1777                 if (!parent)
1778                         break;
1779
1780                 /*
1781                  * The parent doesn't have low limit, it always reaches low
1782                  * limit. Its overflow time is useless for children
1783                  */
1784                 if (!parent->bps[READ][LIMIT_LOW] &&
1785                     !parent->iops[READ][LIMIT_LOW] &&
1786                     !parent->bps[WRITE][LIMIT_LOW] &&
1787                     !parent->iops[WRITE][LIMIT_LOW])
1788                         continue;
1789                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1790                         ret = __tg_last_low_overflow_time(parent);
1791         }
1792         return ret;
1793 }
1794
1795 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1796 {
1797         /*
1798          * cgroup is idle if:
1799          * - single idle is too long, longer than a fixed value (in case user
1800          *   configure a too big threshold) or 4 times of idletime threshold
1801          * - average think time is more than threshold
1802          * - IO latency is largely below threshold
1803          */
1804         unsigned long time;
1805         bool ret;
1806
1807         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1808         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1809               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1810               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1811               tg->avg_idletime > tg->idletime_threshold ||
1812               (tg->latency_target && tg->bio_cnt &&
1813                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1814         throtl_log(&tg->service_queue,
1815                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1816                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1817                 tg->bio_cnt, ret, tg->td->scale);
1818         return ret;
1819 }
1820
1821 static bool throtl_low_limit_reached(struct throtl_grp *tg, int rw)
1822 {
1823         struct throtl_service_queue *sq = &tg->service_queue;
1824         bool limit = tg->bps[rw][LIMIT_LOW] || tg->iops[rw][LIMIT_LOW];
1825
1826         /*
1827          * if low limit is zero, low limit is always reached.
1828          * if low limit is non-zero, we can check if there is any request
1829          * is queued to determine if low limit is reached as we throttle
1830          * request according to limit.
1831          */
1832         return !limit || sq->nr_queued[rw];
1833 }
1834
1835 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1836 {
1837         /*
1838          * cgroup reaches low limit when low limit of READ and WRITE are
1839          * both reached, it's ok to upgrade to next limit if cgroup reaches
1840          * low limit
1841          */
1842         if (throtl_low_limit_reached(tg, READ) &&
1843             throtl_low_limit_reached(tg, WRITE))
1844                 return true;
1845
1846         if (time_after_eq(jiffies,
1847                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1848             throtl_tg_is_idle(tg))
1849                 return true;
1850         return false;
1851 }
1852
1853 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1854 {
1855         while (true) {
1856                 if (throtl_tg_can_upgrade(tg))
1857                         return true;
1858                 tg = sq_to_tg(tg->service_queue.parent_sq);
1859                 if (!tg || !tg_to_blkg(tg)->parent)
1860                         return false;
1861         }
1862         return false;
1863 }
1864
1865 static bool throtl_can_upgrade(struct throtl_data *td,
1866         struct throtl_grp *this_tg)
1867 {
1868         struct cgroup_subsys_state *pos_css;
1869         struct blkcg_gq *blkg;
1870
1871         if (td->limit_index != LIMIT_LOW)
1872                 return false;
1873
1874         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1875                 return false;
1876
1877         rcu_read_lock();
1878         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1879                 struct throtl_grp *tg = blkg_to_tg(blkg);
1880
1881                 if (tg == this_tg)
1882                         continue;
1883                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1884                         continue;
1885                 if (!throtl_hierarchy_can_upgrade(tg)) {
1886                         rcu_read_unlock();
1887                         return false;
1888                 }
1889         }
1890         rcu_read_unlock();
1891         return true;
1892 }
1893
1894 static void throtl_upgrade_check(struct throtl_grp *tg)
1895 {
1896         unsigned long now = jiffies;
1897
1898         if (tg->td->limit_index != LIMIT_LOW)
1899                 return;
1900
1901         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1902                 return;
1903
1904         tg->last_check_time = now;
1905
1906         if (!time_after_eq(now,
1907              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1908                 return;
1909
1910         if (throtl_can_upgrade(tg->td, NULL))
1911                 throtl_upgrade_state(tg->td);
1912 }
1913
1914 static void throtl_upgrade_state(struct throtl_data *td)
1915 {
1916         struct cgroup_subsys_state *pos_css;
1917         struct blkcg_gq *blkg;
1918
1919         throtl_log(&td->service_queue, "upgrade to max");
1920         td->limit_index = LIMIT_MAX;
1921         td->low_upgrade_time = jiffies;
1922         td->scale = 0;
1923         rcu_read_lock();
1924         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1925                 struct throtl_grp *tg = blkg_to_tg(blkg);
1926                 struct throtl_service_queue *sq = &tg->service_queue;
1927
1928                 tg->disptime = jiffies - 1;
1929                 throtl_select_dispatch(sq);
1930                 throtl_schedule_next_dispatch(sq, true);
1931         }
1932         rcu_read_unlock();
1933         throtl_select_dispatch(&td->service_queue);
1934         throtl_schedule_next_dispatch(&td->service_queue, true);
1935         queue_work(kthrotld_workqueue, &td->dispatch_work);
1936 }
1937
1938 static void throtl_downgrade_state(struct throtl_data *td)
1939 {
1940         td->scale /= 2;
1941
1942         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1943         if (td->scale) {
1944                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1945                 return;
1946         }
1947
1948         td->limit_index = LIMIT_LOW;
1949         td->low_downgrade_time = jiffies;
1950 }
1951
1952 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1953 {
1954         struct throtl_data *td = tg->td;
1955         unsigned long now = jiffies;
1956
1957         /*
1958          * If cgroup is below low limit, consider downgrade and throttle other
1959          * cgroups
1960          */
1961         if (time_after_eq(now, tg_last_low_overflow_time(tg) +
1962                                         td->throtl_slice) &&
1963             (!throtl_tg_is_idle(tg) ||
1964              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1965                 return true;
1966         return false;
1967 }
1968
1969 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1970 {
1971         struct throtl_data *td = tg->td;
1972
1973         if (time_before(jiffies, td->low_upgrade_time + td->throtl_slice))
1974                 return false;
1975
1976         while (true) {
1977                 if (!throtl_tg_can_downgrade(tg))
1978                         return false;
1979                 tg = sq_to_tg(tg->service_queue.parent_sq);
1980                 if (!tg || !tg_to_blkg(tg)->parent)
1981                         break;
1982         }
1983         return true;
1984 }
1985
1986 static void throtl_downgrade_check(struct throtl_grp *tg)
1987 {
1988         uint64_t bps;
1989         unsigned int iops;
1990         unsigned long elapsed_time;
1991         unsigned long now = jiffies;
1992
1993         if (tg->td->limit_index != LIMIT_MAX ||
1994             !tg->td->limit_valid[LIMIT_LOW])
1995                 return;
1996         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1997                 return;
1998         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1999                 return;
2000
2001         elapsed_time = now - tg->last_check_time;
2002         tg->last_check_time = now;
2003
2004         if (time_before(now, tg_last_low_overflow_time(tg) +
2005                         tg->td->throtl_slice))
2006                 return;
2007
2008         if (tg->bps[READ][LIMIT_LOW]) {
2009                 bps = tg->last_bytes_disp[READ] * HZ;
2010                 do_div(bps, elapsed_time);
2011                 if (bps >= tg->bps[READ][LIMIT_LOW])
2012                         tg->last_low_overflow_time[READ] = now;
2013         }
2014
2015         if (tg->bps[WRITE][LIMIT_LOW]) {
2016                 bps = tg->last_bytes_disp[WRITE] * HZ;
2017                 do_div(bps, elapsed_time);
2018                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2019                         tg->last_low_overflow_time[WRITE] = now;
2020         }
2021
2022         if (tg->iops[READ][LIMIT_LOW]) {
2023                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2024                 if (iops >= tg->iops[READ][LIMIT_LOW])
2025                         tg->last_low_overflow_time[READ] = now;
2026         }
2027
2028         if (tg->iops[WRITE][LIMIT_LOW]) {
2029                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2030                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2031                         tg->last_low_overflow_time[WRITE] = now;
2032         }
2033
2034         /*
2035          * If cgroup is below low limit, consider downgrade and throttle other
2036          * cgroups
2037          */
2038         if (throtl_hierarchy_can_downgrade(tg))
2039                 throtl_downgrade_state(tg->td);
2040
2041         tg->last_bytes_disp[READ] = 0;
2042         tg->last_bytes_disp[WRITE] = 0;
2043         tg->last_io_disp[READ] = 0;
2044         tg->last_io_disp[WRITE] = 0;
2045 }
2046
2047 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2048 {
2049         unsigned long now;
2050         unsigned long last_finish_time = tg->last_finish_time;
2051
2052         if (last_finish_time == 0)
2053                 return;
2054
2055         now = ktime_get_ns() >> 10;
2056         if (now <= last_finish_time ||
2057             last_finish_time == tg->checked_last_finish_time)
2058                 return;
2059
2060         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2061         tg->checked_last_finish_time = last_finish_time;
2062 }
2063
2064 static void throtl_update_latency_buckets(struct throtl_data *td)
2065 {
2066         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2067         int i, cpu, rw;
2068         unsigned long last_latency[2] = { 0 };
2069         unsigned long latency[2];
2070
2071         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2072                 return;
2073         if (time_before(jiffies, td->last_calculate_time + HZ))
2074                 return;
2075         td->last_calculate_time = jiffies;
2076
2077         memset(avg_latency, 0, sizeof(avg_latency));
2078         for (rw = READ; rw <= WRITE; rw++) {
2079                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2080                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2081
2082                         for_each_possible_cpu(cpu) {
2083                                 struct latency_bucket *bucket;
2084
2085                                 /* this isn't race free, but ok in practice */
2086                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2087                                         cpu);
2088                                 tmp->total_latency += bucket[i].total_latency;
2089                                 tmp->samples += bucket[i].samples;
2090                                 bucket[i].total_latency = 0;
2091                                 bucket[i].samples = 0;
2092                         }
2093
2094                         if (tmp->samples >= 32) {
2095                                 int samples = tmp->samples;
2096
2097                                 latency[rw] = tmp->total_latency;
2098
2099                                 tmp->total_latency = 0;
2100                                 tmp->samples = 0;
2101                                 latency[rw] /= samples;
2102                                 if (latency[rw] == 0)
2103                                         continue;
2104                                 avg_latency[rw][i].latency = latency[rw];
2105                         }
2106                 }
2107         }
2108
2109         for (rw = READ; rw <= WRITE; rw++) {
2110                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2111                         if (!avg_latency[rw][i].latency) {
2112                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2113                                         td->avg_buckets[rw][i].latency =
2114                                                 last_latency[rw];
2115                                 continue;
2116                         }
2117
2118                         if (!td->avg_buckets[rw][i].valid)
2119                                 latency[rw] = avg_latency[rw][i].latency;
2120                         else
2121                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2122                                         avg_latency[rw][i].latency) >> 3;
2123
2124                         td->avg_buckets[rw][i].latency = max(latency[rw],
2125                                 last_latency[rw]);
2126                         td->avg_buckets[rw][i].valid = true;
2127                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2128                 }
2129         }
2130
2131         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2132                 throtl_log(&td->service_queue,
2133                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2134                         "write latency=%ld, write valid=%d", i,
2135                         td->avg_buckets[READ][i].latency,
2136                         td->avg_buckets[READ][i].valid,
2137                         td->avg_buckets[WRITE][i].latency,
2138                         td->avg_buckets[WRITE][i].valid);
2139 }
2140 #else
2141 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2142 {
2143 }
2144
2145 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2146 {
2147 }
2148
2149 static void throtl_downgrade_check(struct throtl_grp *tg)
2150 {
2151 }
2152
2153 static void throtl_upgrade_check(struct throtl_grp *tg)
2154 {
2155 }
2156
2157 static bool throtl_can_upgrade(struct throtl_data *td,
2158         struct throtl_grp *this_tg)
2159 {
2160         return false;
2161 }
2162
2163 static void throtl_upgrade_state(struct throtl_data *td)
2164 {
2165 }
2166 #endif
2167
2168 bool __blk_throtl_bio(struct bio *bio)
2169 {
2170         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2171         struct blkcg_gq *blkg = bio->bi_blkg;
2172         struct throtl_qnode *qn = NULL;
2173         struct throtl_grp *tg = blkg_to_tg(blkg);
2174         struct throtl_service_queue *sq;
2175         bool rw = bio_data_dir(bio);
2176         bool throttled = false;
2177         struct throtl_data *td = tg->td;
2178
2179         rcu_read_lock();
2180
2181         if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2182                 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2183                                 bio->bi_iter.bi_size);
2184                 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2185         }
2186
2187         spin_lock_irq(&q->queue_lock);
2188
2189         throtl_update_latency_buckets(td);
2190
2191         blk_throtl_update_idletime(tg);
2192
2193         sq = &tg->service_queue;
2194
2195 again:
2196         while (true) {
2197                 if (tg->last_low_overflow_time[rw] == 0)
2198                         tg->last_low_overflow_time[rw] = jiffies;
2199                 throtl_downgrade_check(tg);
2200                 throtl_upgrade_check(tg);
2201                 /* throtl is FIFO - if bios are already queued, should queue */
2202                 if (sq->nr_queued[rw])
2203                         break;
2204
2205                 /* if above limits, break to queue */
2206                 if (!tg_may_dispatch(tg, bio, NULL)) {
2207                         tg->last_low_overflow_time[rw] = jiffies;
2208                         if (throtl_can_upgrade(td, tg)) {
2209                                 throtl_upgrade_state(td);
2210                                 goto again;
2211                         }
2212                         break;
2213                 }
2214
2215                 /* within limits, let's charge and dispatch directly */
2216                 throtl_charge_bio(tg, bio);
2217
2218                 /*
2219                  * We need to trim slice even when bios are not being queued
2220                  * otherwise it might happen that a bio is not queued for
2221                  * a long time and slice keeps on extending and trim is not
2222                  * called for a long time. Now if limits are reduced suddenly
2223                  * we take into account all the IO dispatched so far at new
2224                  * low rate and * newly queued IO gets a really long dispatch
2225                  * time.
2226                  *
2227                  * So keep on trimming slice even if bio is not queued.
2228                  */
2229                 throtl_trim_slice(tg, rw);
2230
2231                 /*
2232                  * @bio passed through this layer without being throttled.
2233                  * Climb up the ladder.  If we're already at the top, it
2234                  * can be executed directly.
2235                  */
2236                 qn = &tg->qnode_on_parent[rw];
2237                 sq = sq->parent_sq;
2238                 tg = sq_to_tg(sq);
2239                 if (!tg) {
2240                         bio_set_flag(bio, BIO_BPS_THROTTLED);
2241                         goto out_unlock;
2242                 }
2243         }
2244
2245         /* out-of-limit, queue to @tg */
2246         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2247                    rw == READ ? 'R' : 'W',
2248                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2249                    tg_bps_limit(tg, rw),
2250                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2251                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2252
2253         tg->last_low_overflow_time[rw] = jiffies;
2254
2255         td->nr_queued[rw]++;
2256         throtl_add_bio_tg(bio, qn, tg);
2257         throttled = true;
2258
2259         /*
2260          * Update @tg's dispatch time and force schedule dispatch if @tg
2261          * was empty before @bio.  The forced scheduling isn't likely to
2262          * cause undue delay as @bio is likely to be dispatched directly if
2263          * its @tg's disptime is not in the future.
2264          */
2265         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2266                 tg_update_disptime(tg);
2267                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2268         }
2269
2270 out_unlock:
2271 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2272         if (throttled || !td->track_bio_latency)
2273                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2274 #endif
2275         spin_unlock_irq(&q->queue_lock);
2276
2277         rcu_read_unlock();
2278         return throttled;
2279 }
2280
2281 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2282 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2283                                  enum req_op op, unsigned long time)
2284 {
2285         const bool rw = op_is_write(op);
2286         struct latency_bucket *latency;
2287         int index;
2288
2289         if (!td || td->limit_index != LIMIT_LOW ||
2290             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2291             !blk_queue_nonrot(td->queue))
2292                 return;
2293
2294         index = request_bucket_index(size);
2295
2296         latency = get_cpu_ptr(td->latency_buckets[rw]);
2297         latency[index].total_latency += time;
2298         latency[index].samples++;
2299         put_cpu_ptr(td->latency_buckets[rw]);
2300 }
2301
2302 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2303 {
2304         struct request_queue *q = rq->q;
2305         struct throtl_data *td = q->td;
2306
2307         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2308                              time_ns >> 10);
2309 }
2310
2311 void blk_throtl_bio_endio(struct bio *bio)
2312 {
2313         struct blkcg_gq *blkg;
2314         struct throtl_grp *tg;
2315         u64 finish_time_ns;
2316         unsigned long finish_time;
2317         unsigned long start_time;
2318         unsigned long lat;
2319         int rw = bio_data_dir(bio);
2320
2321         blkg = bio->bi_blkg;
2322         if (!blkg)
2323                 return;
2324         tg = blkg_to_tg(blkg);
2325         if (!tg->td->limit_valid[LIMIT_LOW])
2326                 return;
2327
2328         finish_time_ns = ktime_get_ns();
2329         tg->last_finish_time = finish_time_ns >> 10;
2330
2331         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2332         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2333         if (!start_time || finish_time <= start_time)
2334                 return;
2335
2336         lat = finish_time - start_time;
2337         /* this is only for bio based driver */
2338         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2339                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2340                                      bio_op(bio), lat);
2341
2342         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2343                 int bucket;
2344                 unsigned int threshold;
2345
2346                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2347                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2348                         tg->latency_target;
2349                 if (lat > threshold)
2350                         tg->bad_bio_cnt++;
2351                 /*
2352                  * Not race free, could get wrong count, which means cgroups
2353                  * will be throttled
2354                  */
2355                 tg->bio_cnt++;
2356         }
2357
2358         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2359                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2360                 tg->bio_cnt /= 2;
2361                 tg->bad_bio_cnt /= 2;
2362         }
2363 }
2364 #endif
2365
2366 int blk_throtl_init(struct gendisk *disk)
2367 {
2368         struct request_queue *q = disk->queue;
2369         struct throtl_data *td;
2370         int ret;
2371
2372         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2373         if (!td)
2374                 return -ENOMEM;
2375         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2376                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2377         if (!td->latency_buckets[READ]) {
2378                 kfree(td);
2379                 return -ENOMEM;
2380         }
2381         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2382                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2383         if (!td->latency_buckets[WRITE]) {
2384                 free_percpu(td->latency_buckets[READ]);
2385                 kfree(td);
2386                 return -ENOMEM;
2387         }
2388
2389         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2390         throtl_service_queue_init(&td->service_queue);
2391
2392         q->td = td;
2393         td->queue = q;
2394
2395         td->limit_valid[LIMIT_MAX] = true;
2396         td->limit_index = LIMIT_MAX;
2397         td->low_upgrade_time = jiffies;
2398         td->low_downgrade_time = jiffies;
2399
2400         /* activate policy */
2401         ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
2402         if (ret) {
2403                 free_percpu(td->latency_buckets[READ]);
2404                 free_percpu(td->latency_buckets[WRITE]);
2405                 kfree(td);
2406         }
2407         return ret;
2408 }
2409
2410 void blk_throtl_exit(struct gendisk *disk)
2411 {
2412         struct request_queue *q = disk->queue;
2413
2414         BUG_ON(!q->td);
2415         del_timer_sync(&q->td->service_queue.pending_timer);
2416         throtl_shutdown_wq(q);
2417         blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
2418         free_percpu(q->td->latency_buckets[READ]);
2419         free_percpu(q->td->latency_buckets[WRITE]);
2420         kfree(q->td);
2421 }
2422
2423 void blk_throtl_register(struct gendisk *disk)
2424 {
2425         struct request_queue *q = disk->queue;
2426         struct throtl_data *td;
2427         int i;
2428
2429         td = q->td;
2430         BUG_ON(!td);
2431
2432         if (blk_queue_nonrot(q)) {
2433                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2434                 td->filtered_latency = LATENCY_FILTERED_SSD;
2435         } else {
2436                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2437                 td->filtered_latency = LATENCY_FILTERED_HD;
2438                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2439                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2440                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2441                 }
2442         }
2443 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2444         /* if no low limit, use previous default */
2445         td->throtl_slice = DFL_THROTL_SLICE_HD;
2446
2447 #else
2448         td->track_bio_latency = !queue_is_mq(q);
2449         if (!td->track_bio_latency)
2450                 blk_stat_enable_accounting(q);
2451 #endif
2452 }
2453
2454 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2455 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2456 {
2457         if (!q->td)
2458                 return -EINVAL;
2459         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2460 }
2461
2462 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2463         const char *page, size_t count)
2464 {
2465         unsigned long v;
2466         unsigned long t;
2467
2468         if (!q->td)
2469                 return -EINVAL;
2470         if (kstrtoul(page, 10, &v))
2471                 return -EINVAL;
2472         t = msecs_to_jiffies(v);
2473         if (t == 0 || t > MAX_THROTL_SLICE)
2474                 return -EINVAL;
2475         q->td->throtl_slice = t;
2476         return count;
2477 }
2478 #endif
2479
2480 static int __init throtl_init(void)
2481 {
2482         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2483         if (!kthrotld_workqueue)
2484                 panic("Failed to create kthrotld\n");
2485
2486         return blkcg_policy_register(&blkcg_policy_throtl);
2487 }
2488
2489 module_init(throtl_init);