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 static unsigned int calculate_io_allowed(u32 iops_limit,
 701                                          unsigned long jiffy_elapsed)
 702 {
 703         unsigned int io_allowed;
 704         u64 tmp;
 705
 706         /*
 707          * jiffy_elapsed should not be a big value as minimum iops can be
 708          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 709          * will allow dispatch after 1 second and after that slice should
 710          * have been trimmed.
 711          */
 712
 713         tmp = (u64)iops_limit * jiffy_elapsed;
 714         do_div(tmp, HZ);
 715
 716         if (tmp > UINT_MAX)
 717                 io_allowed = UINT_MAX;
 718         else
 719                 io_allowed = tmp;
 720
 721         return io_allowed;
 722 }
 723
 724 static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
 725 {
 726         return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
 727 }
 728
 729 /* Trim the used slices and adjust slice start accordingly */
 730 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 731 {
 732         unsigned long time_elapsed, io_trim;
 733         u64 bytes_trim;
 734
 735         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 736
 737         /*
 738          * If bps are unlimited (-1), then time slice don't get
 739          * renewed. Don't try to trim the slice if slice is used. A new
 740          * slice will start when appropriate.
 741          */
 742         if (throtl_slice_used(tg, rw))
 743                 return;
 744
 745         /*
 746          * A bio has been dispatched. Also adjust slice_end. It might happen
 747          * that initially cgroup limit was very low resulting in high
 748          * slice_end, but later limit was bumped up and bio was dispatched
 749          * sooner, then we need to reduce slice_end. A high bogus slice_end
 750          * is bad because it does not allow new slice to start.
 751          */
 752
 753         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 754
 755         time_elapsed = rounddown(jiffies - tg->slice_start[rw],
 756                                  tg->td->throtl_slice);
 757         if (!time_elapsed)
 758                 return;
 759
 760         bytes_trim = calculate_bytes_allowed(tg_bps_limit(tg, rw),
 761                                              time_elapsed);
 762         io_trim = calculate_io_allowed(tg_iops_limit(tg, rw), time_elapsed);
 763         if (!bytes_trim && !io_trim)
 764                 return;
 765
 766         if (tg->bytes_disp[rw] >= bytes_trim)
 767                 tg->bytes_disp[rw] -= bytes_trim;
 768         else
 769                 tg->bytes_disp[rw] = 0;
 770
 771         if (tg->io_disp[rw] >= io_trim)
 772                 tg->io_disp[rw] -= io_trim;
 773         else
 774                 tg->io_disp[rw] = 0;
 775
 776         tg->slice_start[rw] += time_elapsed;
 777
 778         throtl_log(&tg->service_queue,
 779                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 780                    rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice,
 781                    bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw],
 782                    jiffies);
 783 }
 784
 785 static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
 786 {
 787         unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
 788         u64 bps_limit = tg_bps_limit(tg, rw);
 789         u32 iops_limit = tg_iops_limit(tg, rw);
 790
 791         /*
 792          * If config is updated while bios are still throttled, calculate and
 793          * accumulate how many bytes/ios are waited across changes. And
 794          * carryover_bytes/ios will be used to calculate new wait time under new
 795          * configuration.
 796          */
 797         if (bps_limit != U64_MAX)
 798                 tg->carryover_bytes[rw] +=
 799                         calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
 800                         tg->bytes_disp[rw];
 801         if (iops_limit != UINT_MAX)
 802                 tg->carryover_ios[rw] +=
 803                         calculate_io_allowed(iops_limit, jiffy_elapsed) -
 804                         tg->io_disp[rw];
 805 }
 806
 807 static void tg_update_carryover(struct throtl_grp *tg)
 808 {
 809         if (tg->service_queue.nr_queued[READ])
 810                 __tg_update_carryover(tg, READ);
 811         if (tg->service_queue.nr_queued[WRITE])
 812                 __tg_update_carryover(tg, WRITE);
 813
 814         /* see comments in struct throtl_grp for meaning of these fields. */
 815         throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
 816                    tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
 817                    tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
 818 }
 819
 820 static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
 821                                  u32 iops_limit)
 822 {
 823         bool rw = bio_data_dir(bio);
 824         int io_allowed;
 825         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 826
 827         if (iops_limit == UINT_MAX) {
 828                 return 0;
 829         }
 830
 831         jiffy_elapsed = jiffies - tg->slice_start[rw];
 832
 833         /* Round up to the next throttle slice, wait time must be nonzero */
 834         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 835         io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
 836                      tg->carryover_ios[rw];
 837         if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
 838                 return 0;
 839
 840         /* Calc approx time to dispatch */
 841         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 842         return jiffy_wait;
 843 }
 844
 845 static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
 846                                 u64 bps_limit)
 847 {
 848         bool rw = bio_data_dir(bio);
 849         long long bytes_allowed;
 850         u64 extra_bytes;
 851         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 852         unsigned int bio_size = throtl_bio_data_size(bio);
 853
 854         /* no need to throttle if this bio's bytes have been accounted */
 855         if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
 856                 return 0;
 857         }
 858
 859         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 860
 861         /* Slice has just started. Consider one slice interval */
 862         if (!jiffy_elapsed)
 863                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 864
 865         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 866         bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
 867                         tg->carryover_bytes[rw];
 868         if (bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
 869                 return 0;
 870
 871         /* Calc approx time to dispatch */
 872         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 873         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 874
 875         if (!jiffy_wait)
 876                 jiffy_wait = 1;
 877
 878         /*
 879          * This wait time is without taking into consideration the rounding
 880          * up we did. Add that time also.
 881          */
 882         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 883         return jiffy_wait;
 884 }
 885
 886 /*
 887  * Returns whether one can dispatch a bio or not. Also returns approx number
 888  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 889  */
 890 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 891                             unsigned long *wait)
 892 {
 893         bool rw = bio_data_dir(bio);
 894         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 895         u64 bps_limit = tg_bps_limit(tg, rw);
 896         u32 iops_limit = tg_iops_limit(tg, rw);
 897
 898         /*
 899          * Currently whole state machine of group depends on first bio
 900          * queued in the group bio list. So one should not be calling
 901          * this function with a different bio if there are other bios
 902          * queued.
 903          */
 904         BUG_ON(tg->service_queue.nr_queued[rw] &&
 905                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 906
 907         /* If tg->bps = -1, then BW is unlimited */
 908         if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
 909             tg->flags & THROTL_TG_CANCELING) {
 910                 if (wait)
 911                         *wait = 0;
 912                 return true;
 913         }
 914
 915         /*
 916          * If previous slice expired, start a new one otherwise renew/extend
 917          * existing slice to make sure it is at least throtl_slice interval
 918          * long since now. New slice is started only for empty throttle group.
 919          * If there is queued bio, that means there should be an active
 920          * slice and it should be extended instead.
 921          */
 922         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 923                 throtl_start_new_slice(tg, rw, true);
 924         else {
 925                 if (time_before(tg->slice_end[rw],
 926                     jiffies + tg->td->throtl_slice))
 927                         throtl_extend_slice(tg, rw,
 928                                 jiffies + tg->td->throtl_slice);
 929         }
 930
 931         bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
 932         iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
 933         if (bps_wait + iops_wait == 0) {
 934                 if (wait)
 935                         *wait = 0;
 936                 return true;
 937         }
 938
 939         max_wait = max(bps_wait, iops_wait);
 940
 941         if (wait)
 942                 *wait = max_wait;
 943
 944         if (time_before(tg->slice_end[rw], jiffies + max_wait))
 945                 throtl_extend_slice(tg, rw, jiffies + max_wait);
 946
 947         return false;
 948 }
 949
 950 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 951 {
 952         bool rw = bio_data_dir(bio);
 953         unsigned int bio_size = throtl_bio_data_size(bio);
 954
 955         /* Charge the bio to the group */
 956         if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
 957                 tg->bytes_disp[rw] += bio_size;
 958                 tg->last_bytes_disp[rw] += bio_size;
 959         }
 960
 961         tg->io_disp[rw]++;
 962         tg->last_io_disp[rw]++;
 963 }
 964
 965 /**
 966  * throtl_add_bio_tg - add a bio to the specified throtl_grp
 967  * @bio: bio to add
 968  * @qn: qnode to use
 969  * @tg: the target throtl_grp
 970  *
 971  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 972  * tg->qnode_on_self[] is used.
 973  */
 974 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 975                               struct throtl_grp *tg)
 976 {
 977         struct throtl_service_queue *sq = &tg->service_queue;
 978         bool rw = bio_data_dir(bio);
 979
 980         if (!qn)
 981                 qn = &tg->qnode_on_self[rw];
 982
 983         /*
 984          * If @tg doesn't currently have any bios queued in the same
 985          * direction, queueing @bio can change when @tg should be
 986          * dispatched.  Mark that @tg was empty.  This is automatically
 987          * cleared on the next tg_update_disptime().
 988          */
 989         if (!sq->nr_queued[rw])
 990                 tg->flags |= THROTL_TG_WAS_EMPTY;
 991
 992         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
 993
 994         sq->nr_queued[rw]++;
 995         throtl_enqueue_tg(tg);
 996 }
 997
 998 static void tg_update_disptime(struct throtl_grp *tg)
 999 {
1000         struct throtl_service_queue *sq = &tg->service_queue;
1001         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1002         struct bio *bio;
1003
1004         bio = throtl_peek_queued(&sq->queued[READ]);
1005         if (bio)
1006                 tg_may_dispatch(tg, bio, &read_wait);
1007
1008         bio = throtl_peek_queued(&sq->queued[WRITE]);
1009         if (bio)
1010                 tg_may_dispatch(tg, bio, &write_wait);
1011
1012         min_wait = min(read_wait, write_wait);
1013         disptime = jiffies + min_wait;
1014
1015         /* Update dispatch time */
1016         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
1017         tg->disptime = disptime;
1018         tg_service_queue_add(tg);
1019
1020         /* see throtl_add_bio_tg() */
1021         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1022 }
1023
1024 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1025                                         struct throtl_grp *parent_tg, bool rw)
1026 {
1027         if (throtl_slice_used(parent_tg, rw)) {
1028                 throtl_start_new_slice_with_credit(parent_tg, rw,
1029                                 child_tg->slice_start[rw]);
1030         }
1031
1032 }
1033
1034 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1035 {
1036         struct throtl_service_queue *sq = &tg->service_queue;
1037         struct throtl_service_queue *parent_sq = sq->parent_sq;
1038         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1039         struct throtl_grp *tg_to_put = NULL;
1040         struct bio *bio;
1041
1042         /*
1043          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1044          * from @tg may put its reference and @parent_sq might end up
1045          * getting released prematurely.  Remember the tg to put and put it
1046          * after @bio is transferred to @parent_sq.
1047          */
1048         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1049         sq->nr_queued[rw]--;
1050
1051         throtl_charge_bio(tg, bio);
1052
1053         /*
1054          * If our parent is another tg, we just need to transfer @bio to
1055          * the parent using throtl_add_bio_tg().  If our parent is
1056          * @td->service_queue, @bio is ready to be issued.  Put it on its
1057          * bio_lists[] and decrease total number queued.  The caller is
1058          * responsible for issuing these bios.
1059          */
1060         if (parent_tg) {
1061                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1062                 start_parent_slice_with_credit(tg, parent_tg, rw);
1063         } else {
1064                 bio_set_flag(bio, BIO_BPS_THROTTLED);
1065                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1066                                      &parent_sq->queued[rw]);
1067                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1068                 tg->td->nr_queued[rw]--;
1069         }
1070
1071         throtl_trim_slice(tg, rw);
1072
1073         if (tg_to_put)
1074                 blkg_put(tg_to_blkg(tg_to_put));
1075 }
1076
1077 static int throtl_dispatch_tg(struct throtl_grp *tg)
1078 {
1079         struct throtl_service_queue *sq = &tg->service_queue;
1080         unsigned int nr_reads = 0, nr_writes = 0;
1081         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1082         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1083         struct bio *bio;
1084
1085         /* Try to dispatch 75% READS and 25% WRITES */
1086
1087         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1088                tg_may_dispatch(tg, bio, NULL)) {
1089
1090                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1091                 nr_reads++;
1092
1093                 if (nr_reads >= max_nr_reads)
1094                         break;
1095         }
1096
1097         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1098                tg_may_dispatch(tg, bio, NULL)) {
1099
1100                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1101                 nr_writes++;
1102
1103                 if (nr_writes >= max_nr_writes)
1104                         break;
1105         }
1106
1107         return nr_reads + nr_writes;
1108 }
1109
1110 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1111 {
1112         unsigned int nr_disp = 0;
1113
1114         while (1) {
1115                 struct throtl_grp *tg;
1116                 struct throtl_service_queue *sq;
1117
1118                 if (!parent_sq->nr_pending)
1119                         break;
1120
1121                 tg = throtl_rb_first(parent_sq);
1122                 if (!tg)
1123                         break;
1124
1125                 if (time_before(jiffies, tg->disptime))
1126                         break;
1127
1128                 nr_disp += throtl_dispatch_tg(tg);
1129
1130                 sq = &tg->service_queue;
1131                 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
1132                         tg_update_disptime(tg);
1133                 else
1134                         throtl_dequeue_tg(tg);
1135
1136                 if (nr_disp >= THROTL_QUANTUM)
1137                         break;
1138         }
1139
1140         return nr_disp;
1141 }
1142
1143 static bool throtl_can_upgrade(struct throtl_data *td,
1144         struct throtl_grp *this_tg);
1145 /**
1146  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1147  * @t: the pending_timer member of the throtl_service_queue being serviced
1148  *
1149  * This timer is armed when a child throtl_grp with active bio's become
1150  * pending and queued on the service_queue's pending_tree and expires when
1151  * the first child throtl_grp should be dispatched.  This function
1152  * dispatches bio's from the children throtl_grps to the parent
1153  * service_queue.
1154  *
1155  * If the parent's parent is another throtl_grp, dispatching is propagated
1156  * by either arming its pending_timer or repeating dispatch directly.  If
1157  * the top-level service_tree is reached, throtl_data->dispatch_work is
1158  * kicked so that the ready bio's are issued.
1159  */
1160 static void throtl_pending_timer_fn(struct timer_list *t)
1161 {
1162         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1163         struct throtl_grp *tg = sq_to_tg(sq);
1164         struct throtl_data *td = sq_to_td(sq);
1165         struct throtl_service_queue *parent_sq;
1166         struct request_queue *q;
1167         bool dispatched;
1168         int ret;
1169
1170         /* throtl_data may be gone, so figure out request queue by blkg */
1171         if (tg)
1172                 q = tg->pd.blkg->q;
1173         else
1174                 q = td->queue;
1175
1176         spin_lock_irq(&q->queue_lock);
1177
1178         if (!q->root_blkg)
1179                 goto out_unlock;
1180
1181         if (throtl_can_upgrade(td, NULL))
1182                 throtl_upgrade_state(td);
1183
1184 again:
1185         parent_sq = sq->parent_sq;
1186         dispatched = false;
1187
1188         while (true) {
1189                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1190                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1191                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1192
1193                 ret = throtl_select_dispatch(sq);
1194                 if (ret) {
1195                         throtl_log(sq, "bios disp=%u", ret);
1196                         dispatched = true;
1197                 }
1198
1199                 if (throtl_schedule_next_dispatch(sq, false))
1200                         break;
1201
1202                 /* this dispatch windows is still open, relax and repeat */
1203                 spin_unlock_irq(&q->queue_lock);
1204                 cpu_relax();
1205                 spin_lock_irq(&q->queue_lock);
1206         }
1207
1208         if (!dispatched)
1209                 goto out_unlock;
1210
1211         if (parent_sq) {
1212                 /* @parent_sq is another throl_grp, propagate dispatch */
1213                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1214                         tg_update_disptime(tg);
1215                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1216                                 /* window is already open, repeat dispatching */
1217                                 sq = parent_sq;
1218                                 tg = sq_to_tg(sq);
1219                                 goto again;
1220                         }
1221                 }
1222         } else {
1223                 /* reached the top-level, queue issuing */
1224                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1225         }
1226 out_unlock:
1227         spin_unlock_irq(&q->queue_lock);
1228 }
1229
1230 /**
1231  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1232  * @work: work item being executed
1233  *
1234  * This function is queued for execution when bios reach the bio_lists[]
1235  * of throtl_data->service_queue.  Those bios are ready and issued by this
1236  * function.
1237  */
1238 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1239 {
1240         struct throtl_data *td = container_of(work, struct throtl_data,
1241                                               dispatch_work);
1242         struct throtl_service_queue *td_sq = &td->service_queue;
1243         struct request_queue *q = td->queue;
1244         struct bio_list bio_list_on_stack;
1245         struct bio *bio;
1246         struct blk_plug plug;
1247         int rw;
1248
1249         bio_list_init(&bio_list_on_stack);
1250
1251         spin_lock_irq(&q->queue_lock);
1252         for (rw = READ; rw <= WRITE; rw++)
1253                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1254                         bio_list_add(&bio_list_on_stack, bio);
1255         spin_unlock_irq(&q->queue_lock);
1256
1257         if (!bio_list_empty(&bio_list_on_stack)) {
1258                 blk_start_plug(&plug);
1259                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1260                         submit_bio_noacct_nocheck(bio);
1261                 blk_finish_plug(&plug);
1262         }
1263 }
1264
1265 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1266                               int off)
1267 {
1268         struct throtl_grp *tg = pd_to_tg(pd);
1269         u64 v = *(u64 *)((void *)tg + off);
1270
1271         if (v == U64_MAX)
1272                 return 0;
1273         return __blkg_prfill_u64(sf, pd, v);
1274 }
1275
1276 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1277                                int off)
1278 {
1279         struct throtl_grp *tg = pd_to_tg(pd);
1280         unsigned int v = *(unsigned int *)((void *)tg + off);
1281
1282         if (v == UINT_MAX)
1283                 return 0;
1284         return __blkg_prfill_u64(sf, pd, v);
1285 }
1286
1287 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1288 {
1289         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1290                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1291         return 0;
1292 }
1293
1294 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1295 {
1296         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1297                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1298         return 0;
1299 }
1300
1301 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1302 {
1303         struct throtl_service_queue *sq = &tg->service_queue;
1304         struct cgroup_subsys_state *pos_css;
1305         struct blkcg_gq *blkg;
1306
1307         throtl_log(&tg->service_queue,
1308                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1309                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1310                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1311
1312         /*
1313          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1314          * considered to have rules if either the tg itself or any of its
1315          * ancestors has rules.  This identifies groups without any
1316          * restrictions in the whole hierarchy and allows them to bypass
1317          * blk-throttle.
1318          */
1319         blkg_for_each_descendant_pre(blkg, pos_css,
1320                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1321                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1322                 struct throtl_grp *parent_tg;
1323
1324                 tg_update_has_rules(this_tg);
1325                 /* ignore root/second level */
1326                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1327                     !blkg->parent->parent)
1328                         continue;
1329                 parent_tg = blkg_to_tg(blkg->parent);
1330                 /*
1331                  * make sure all children has lower idle time threshold and
1332                  * higher latency target
1333                  */
1334                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1335                                 parent_tg->idletime_threshold);
1336                 this_tg->latency_target = max(this_tg->latency_target,
1337                                 parent_tg->latency_target);
1338         }
1339
1340         /*
1341          * We're already holding queue_lock and know @tg is valid.  Let's
1342          * apply the new config directly.
1343          *
1344          * Restart the slices for both READ and WRITES. It might happen
1345          * that a group's limit are dropped suddenly and we don't want to
1346          * account recently dispatched IO with new low rate.
1347          */
1348         throtl_start_new_slice(tg, READ, false);
1349         throtl_start_new_slice(tg, WRITE, false);
1350
1351         if (tg->flags & THROTL_TG_PENDING) {
1352                 tg_update_disptime(tg);
1353                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1354         }
1355 }
1356
1357 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1358                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1359 {
1360         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1361         struct blkg_conf_ctx ctx;
1362         struct throtl_grp *tg;
1363         int ret;
1364         u64 v;
1365
1366         blkg_conf_init(&ctx, buf);
1367
1368         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1369         if (ret)
1370                 goto out_finish;
1371
1372         ret = -EINVAL;
1373         if (sscanf(ctx.body, "%llu", &v) != 1)
1374                 goto out_finish;
1375         if (!v)
1376                 v = U64_MAX;
1377
1378         tg = blkg_to_tg(ctx.blkg);
1379         tg_update_carryover(tg);
1380
1381         if (is_u64)
1382                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1383         else
1384                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1385
1386         tg_conf_updated(tg, false);
1387         ret = 0;
1388 out_finish:
1389         blkg_conf_exit(&ctx);
1390         return ret ?: nbytes;
1391 }
1392
1393 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1394                                char *buf, size_t nbytes, loff_t off)
1395 {
1396         return tg_set_conf(of, buf, nbytes, off, true);
1397 }
1398
1399 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1400                                 char *buf, size_t nbytes, loff_t off)
1401 {
1402         return tg_set_conf(of, buf, nbytes, off, false);
1403 }
1404
1405 static int tg_print_rwstat(struct seq_file *sf, void *v)
1406 {
1407         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1408                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1409                           seq_cft(sf)->private, true);
1410         return 0;
1411 }
1412
1413 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1414                                       struct blkg_policy_data *pd, int off)
1415 {
1416         struct blkg_rwstat_sample sum;
1417
1418         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1419                                   &sum);
1420         return __blkg_prfill_rwstat(sf, pd, &sum);
1421 }
1422
1423 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1424 {
1425         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1426                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1427                           seq_cft(sf)->private, true);
1428         return 0;
1429 }
1430
1431 static struct cftype throtl_legacy_files[] = {
1432         {
1433                 .name = "throttle.read_bps_device",
1434                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1435                 .seq_show = tg_print_conf_u64,
1436                 .write = tg_set_conf_u64,
1437         },
1438         {
1439                 .name = "throttle.write_bps_device",
1440                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1441                 .seq_show = tg_print_conf_u64,
1442                 .write = tg_set_conf_u64,
1443         },
1444         {
1445                 .name = "throttle.read_iops_device",
1446                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1447                 .seq_show = tg_print_conf_uint,
1448                 .write = tg_set_conf_uint,
1449         },
1450         {
1451                 .name = "throttle.write_iops_device",
1452                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1453                 .seq_show = tg_print_conf_uint,
1454                 .write = tg_set_conf_uint,
1455         },
1456         {
1457                 .name = "throttle.io_service_bytes",
1458                 .private = offsetof(struct throtl_grp, stat_bytes),
1459                 .seq_show = tg_print_rwstat,
1460         },
1461         {
1462                 .name = "throttle.io_service_bytes_recursive",
1463                 .private = offsetof(struct throtl_grp, stat_bytes),
1464                 .seq_show = tg_print_rwstat_recursive,
1465         },
1466         {
1467                 .name = "throttle.io_serviced",
1468                 .private = offsetof(struct throtl_grp, stat_ios),
1469                 .seq_show = tg_print_rwstat,
1470         },
1471         {
1472                 .name = "throttle.io_serviced_recursive",
1473                 .private = offsetof(struct throtl_grp, stat_ios),
1474                 .seq_show = tg_print_rwstat_recursive,
1475         },
1476         { }     /* terminate */
1477 };
1478
1479 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1480                          int off)
1481 {
1482         struct throtl_grp *tg = pd_to_tg(pd);
1483         const char *dname = blkg_dev_name(pd->blkg);
1484         char bufs[4][21] = { "max", "max", "max", "max" };
1485         u64 bps_dft;
1486         unsigned int iops_dft;
1487         char idle_time[26] = "";
1488         char latency_time[26] = "";
1489
1490         if (!dname)
1491                 return 0;
1492
1493         if (off == LIMIT_LOW) {
1494                 bps_dft = 0;
1495                 iops_dft = 0;
1496         } else {
1497                 bps_dft = U64_MAX;
1498                 iops_dft = UINT_MAX;
1499         }
1500
1501         if (tg->bps_conf[READ][off] == bps_dft &&
1502             tg->bps_conf[WRITE][off] == bps_dft &&
1503             tg->iops_conf[READ][off] == iops_dft &&
1504             tg->iops_conf[WRITE][off] == iops_dft &&
1505             (off != LIMIT_LOW ||
1506              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1507               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1508                 return 0;
1509
1510         if (tg->bps_conf[READ][off] != U64_MAX)
1511                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1512                         tg->bps_conf[READ][off]);
1513         if (tg->bps_conf[WRITE][off] != U64_MAX)
1514                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1515                         tg->bps_conf[WRITE][off]);
1516         if (tg->iops_conf[READ][off] != UINT_MAX)
1517                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1518                         tg->iops_conf[READ][off]);
1519         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1520                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1521                         tg->iops_conf[WRITE][off]);
1522         if (off == LIMIT_LOW) {
1523                 if (tg->idletime_threshold_conf == ULONG_MAX)
1524                         strcpy(idle_time, " idle=max");
1525                 else
1526                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1527                                 tg->idletime_threshold_conf);
1528
1529                 if (tg->latency_target_conf == ULONG_MAX)
1530                         strcpy(latency_time, " latency=max");
1531                 else
1532                         snprintf(latency_time, sizeof(latency_time),
1533                                 " latency=%lu", tg->latency_target_conf);
1534         }
1535
1536         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1537                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1538                    latency_time);
1539         return 0;
1540 }
1541
1542 static int tg_print_limit(struct seq_file *sf, void *v)
1543 {
1544         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1545                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1546         return 0;
1547 }
1548
1549 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1550                           char *buf, size_t nbytes, loff_t off)
1551 {
1552         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1553         struct blkg_conf_ctx ctx;
1554         struct throtl_grp *tg;
1555         u64 v[4];
1556         unsigned long idle_time;
1557         unsigned long latency_time;
1558         int ret;
1559         int index = of_cft(of)->private;
1560
1561         blkg_conf_init(&ctx, buf);
1562
1563         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1564         if (ret)
1565                 goto out_finish;
1566
1567         tg = blkg_to_tg(ctx.blkg);
1568         tg_update_carryover(tg);
1569
1570         v[0] = tg->bps_conf[READ][index];
1571         v[1] = tg->bps_conf[WRITE][index];
1572         v[2] = tg->iops_conf[READ][index];
1573         v[3] = tg->iops_conf[WRITE][index];
1574
1575         idle_time = tg->idletime_threshold_conf;
1576         latency_time = tg->latency_target_conf;
1577         while (true) {
1578                 char tok[27];   /* wiops=18446744073709551616 */
1579                 char *p;
1580                 u64 val = U64_MAX;
1581                 int len;
1582
1583                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1584                         break;
1585                 if (tok[0] == '\0')
1586                         break;
1587                 ctx.body += len;
1588
1589                 ret = -EINVAL;
1590                 p = tok;
1591                 strsep(&p, "=");
1592                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1593                         goto out_finish;
1594
1595                 ret = -ERANGE;
1596                 if (!val)
1597                         goto out_finish;
1598
1599                 ret = -EINVAL;
1600                 if (!strcmp(tok, "rbps") && val > 1)
1601                         v[0] = val;
1602                 else if (!strcmp(tok, "wbps") && val > 1)
1603                         v[1] = val;
1604                 else if (!strcmp(tok, "riops") && val > 1)
1605                         v[2] = min_t(u64, val, UINT_MAX);
1606                 else if (!strcmp(tok, "wiops") && val > 1)
1607                         v[3] = min_t(u64, val, UINT_MAX);
1608                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1609                         idle_time = val;
1610                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1611                         latency_time = val;
1612                 else
1613                         goto out_finish;
1614         }
1615
1616         tg->bps_conf[READ][index] = v[0];
1617         tg->bps_conf[WRITE][index] = v[1];
1618         tg->iops_conf[READ][index] = v[2];
1619         tg->iops_conf[WRITE][index] = v[3];
1620
1621         if (index == LIMIT_MAX) {
1622                 tg->bps[READ][index] = v[0];
1623                 tg->bps[WRITE][index] = v[1];
1624                 tg->iops[READ][index] = v[2];
1625                 tg->iops[WRITE][index] = v[3];
1626         }
1627         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1628                 tg->bps_conf[READ][LIMIT_MAX]);
1629         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1630                 tg->bps_conf[WRITE][LIMIT_MAX]);
1631         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1632                 tg->iops_conf[READ][LIMIT_MAX]);
1633         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1634                 tg->iops_conf[WRITE][LIMIT_MAX]);
1635         tg->idletime_threshold_conf = idle_time;
1636         tg->latency_target_conf = latency_time;
1637
1638         /* force user to configure all settings for low limit  */
1639         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1640               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1641             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1642             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1643                 tg->bps[READ][LIMIT_LOW] = 0;
1644                 tg->bps[WRITE][LIMIT_LOW] = 0;
1645                 tg->iops[READ][LIMIT_LOW] = 0;
1646                 tg->iops[WRITE][LIMIT_LOW] = 0;
1647                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1648                 tg->latency_target = DFL_LATENCY_TARGET;
1649         } else if (index == LIMIT_LOW) {
1650                 tg->idletime_threshold = tg->idletime_threshold_conf;
1651                 tg->latency_target = tg->latency_target_conf;
1652         }
1653
1654         blk_throtl_update_limit_valid(tg->td);
1655         if (tg->td->limit_valid[LIMIT_LOW]) {
1656                 if (index == LIMIT_LOW)
1657                         tg->td->limit_index = LIMIT_LOW;
1658         } else
1659                 tg->td->limit_index = LIMIT_MAX;
1660         tg_conf_updated(tg, index == LIMIT_LOW &&
1661                 tg->td->limit_valid[LIMIT_LOW]);
1662         ret = 0;
1663 out_finish:
1664         blkg_conf_exit(&ctx);
1665         return ret ?: nbytes;
1666 }
1667
1668 static struct cftype throtl_files[] = {
1669 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1670         {
1671                 .name = "low",
1672                 .flags = CFTYPE_NOT_ON_ROOT,
1673                 .seq_show = tg_print_limit,
1674                 .write = tg_set_limit,
1675                 .private = LIMIT_LOW,
1676         },
1677 #endif
1678         {
1679                 .name = "max",
1680                 .flags = CFTYPE_NOT_ON_ROOT,
1681                 .seq_show = tg_print_limit,
1682                 .write = tg_set_limit,
1683                 .private = LIMIT_MAX,
1684         },
1685         { }     /* terminate */
1686 };
1687
1688 static void throtl_shutdown_wq(struct request_queue *q)
1689 {
1690         struct throtl_data *td = q->td;
1691
1692         cancel_work_sync(&td->dispatch_work);
1693 }
1694
1695 struct blkcg_policy blkcg_policy_throtl = {
1696         .dfl_cftypes            = throtl_files,
1697         .legacy_cftypes         = throtl_legacy_files,
1698
1699         .pd_alloc_fn            = throtl_pd_alloc,
1700         .pd_init_fn             = throtl_pd_init,
1701         .pd_online_fn           = throtl_pd_online,
1702         .pd_offline_fn          = throtl_pd_offline,
1703         .pd_free_fn             = throtl_pd_free,
1704 };
1705
1706 void blk_throtl_cancel_bios(struct gendisk *disk)
1707 {
1708         struct request_queue *q = disk->queue;
1709         struct cgroup_subsys_state *pos_css;
1710         struct blkcg_gq *blkg;
1711
1712         spin_lock_irq(&q->queue_lock);
1713         /*
1714          * queue_lock is held, rcu lock is not needed here technically.
1715          * However, rcu lock is still held to emphasize that following
1716          * path need RCU protection and to prevent warning from lockdep.
1717          */
1718         rcu_read_lock();
1719         blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1720                 struct throtl_grp *tg = blkg_to_tg(blkg);
1721                 struct throtl_service_queue *sq = &tg->service_queue;
1722
1723                 /*
1724                  * Set the flag to make sure throtl_pending_timer_fn() won't
1725                  * stop until all throttled bios are dispatched.
1726                  */
1727                 tg->flags |= THROTL_TG_CANCELING;
1728
1729                 /*
1730                  * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1731                  * will be inserted to service queue without THROTL_TG_PENDING
1732                  * set in tg_update_disptime below. Then IO dispatched from
1733                  * child in tg_dispatch_one_bio will trigger double insertion
1734                  * and corrupt the tree.
1735                  */
1736                 if (!(tg->flags & THROTL_TG_PENDING))
1737                         continue;
1738
1739                 /*
1740                  * Update disptime after setting the above flag to make sure
1741                  * throtl_select_dispatch() won't exit without dispatching.
1742                  */
1743                 tg_update_disptime(tg);
1744
1745                 throtl_schedule_pending_timer(sq, jiffies + 1);
1746         }
1747         rcu_read_unlock();
1748         spin_unlock_irq(&q->queue_lock);
1749 }
1750
1751 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1752 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1753 {
1754         unsigned long rtime = jiffies, wtime = jiffies;
1755
1756         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1757                 rtime = tg->last_low_overflow_time[READ];
1758         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1759                 wtime = tg->last_low_overflow_time[WRITE];
1760         return min(rtime, wtime);
1761 }
1762
1763 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1764 {
1765         struct throtl_service_queue *parent_sq;
1766         struct throtl_grp *parent = tg;
1767         unsigned long ret = __tg_last_low_overflow_time(tg);
1768
1769         while (true) {
1770                 parent_sq = parent->service_queue.parent_sq;
1771                 parent = sq_to_tg(parent_sq);
1772                 if (!parent)
1773                         break;
1774
1775                 /*
1776                  * The parent doesn't have low limit, it always reaches low
1777                  * limit. Its overflow time is useless for children
1778                  */
1779                 if (!parent->bps[READ][LIMIT_LOW] &&
1780                     !parent->iops[READ][LIMIT_LOW] &&
1781                     !parent->bps[WRITE][LIMIT_LOW] &&
1782                     !parent->iops[WRITE][LIMIT_LOW])
1783                         continue;
1784                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1785                         ret = __tg_last_low_overflow_time(parent);
1786         }
1787         return ret;
1788 }
1789
1790 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1791 {
1792         /*
1793          * cgroup is idle if:
1794          * - single idle is too long, longer than a fixed value (in case user
1795          *   configure a too big threshold) or 4 times of idletime threshold
1796          * - average think time is more than threshold
1797          * - IO latency is largely below threshold
1798          */
1799         unsigned long time;
1800         bool ret;
1801
1802         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1803         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1804               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1805               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1806               tg->avg_idletime > tg->idletime_threshold ||
1807               (tg->latency_target && tg->bio_cnt &&
1808                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1809         throtl_log(&tg->service_queue,
1810                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1811                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1812                 tg->bio_cnt, ret, tg->td->scale);
1813         return ret;
1814 }
1815
1816 static bool throtl_low_limit_reached(struct throtl_grp *tg, int rw)
1817 {
1818         struct throtl_service_queue *sq = &tg->service_queue;
1819         bool limit = tg->bps[rw][LIMIT_LOW] || tg->iops[rw][LIMIT_LOW];
1820
1821         /*
1822          * if low limit is zero, low limit is always reached.
1823          * if low limit is non-zero, we can check if there is any request
1824          * is queued to determine if low limit is reached as we throttle
1825          * request according to limit.
1826          */
1827         return !limit || sq->nr_queued[rw];
1828 }
1829
1830 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1831 {
1832         /*
1833          * cgroup reaches low limit when low limit of READ and WRITE are
1834          * both reached, it's ok to upgrade to next limit if cgroup reaches
1835          * low limit
1836          */
1837         if (throtl_low_limit_reached(tg, READ) &&
1838             throtl_low_limit_reached(tg, WRITE))
1839                 return true;
1840
1841         if (time_after_eq(jiffies,
1842                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1843             throtl_tg_is_idle(tg))
1844                 return true;
1845         return false;
1846 }
1847
1848 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1849 {
1850         while (true) {
1851                 if (throtl_tg_can_upgrade(tg))
1852                         return true;
1853                 tg = sq_to_tg(tg->service_queue.parent_sq);
1854                 if (!tg || !tg_to_blkg(tg)->parent)
1855                         return false;
1856         }
1857         return false;
1858 }
1859
1860 static bool throtl_can_upgrade(struct throtl_data *td,
1861         struct throtl_grp *this_tg)
1862 {
1863         struct cgroup_subsys_state *pos_css;
1864         struct blkcg_gq *blkg;
1865
1866         if (td->limit_index != LIMIT_LOW)
1867                 return false;
1868
1869         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1870                 return false;
1871
1872         rcu_read_lock();
1873         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1874                 struct throtl_grp *tg = blkg_to_tg(blkg);
1875
1876                 if (tg == this_tg)
1877                         continue;
1878                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1879                         continue;
1880                 if (!throtl_hierarchy_can_upgrade(tg)) {
1881                         rcu_read_unlock();
1882                         return false;
1883                 }
1884         }
1885         rcu_read_unlock();
1886         return true;
1887 }
1888
1889 static void throtl_upgrade_check(struct throtl_grp *tg)
1890 {
1891         unsigned long now = jiffies;
1892
1893         if (tg->td->limit_index != LIMIT_LOW)
1894                 return;
1895
1896         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1897                 return;
1898
1899         tg->last_check_time = now;
1900
1901         if (!time_after_eq(now,
1902              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1903                 return;
1904
1905         if (throtl_can_upgrade(tg->td, NULL))
1906                 throtl_upgrade_state(tg->td);
1907 }
1908
1909 static void throtl_upgrade_state(struct throtl_data *td)
1910 {
1911         struct cgroup_subsys_state *pos_css;
1912         struct blkcg_gq *blkg;
1913
1914         throtl_log(&td->service_queue, "upgrade to max");
1915         td->limit_index = LIMIT_MAX;
1916         td->low_upgrade_time = jiffies;
1917         td->scale = 0;
1918         rcu_read_lock();
1919         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1920                 struct throtl_grp *tg = blkg_to_tg(blkg);
1921                 struct throtl_service_queue *sq = &tg->service_queue;
1922
1923                 tg->disptime = jiffies - 1;
1924                 throtl_select_dispatch(sq);
1925                 throtl_schedule_next_dispatch(sq, true);
1926         }
1927         rcu_read_unlock();
1928         throtl_select_dispatch(&td->service_queue);
1929         throtl_schedule_next_dispatch(&td->service_queue, true);
1930         queue_work(kthrotld_workqueue, &td->dispatch_work);
1931 }
1932
1933 static void throtl_downgrade_state(struct throtl_data *td)
1934 {
1935         td->scale /= 2;
1936
1937         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1938         if (td->scale) {
1939                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1940                 return;
1941         }
1942
1943         td->limit_index = LIMIT_LOW;
1944         td->low_downgrade_time = jiffies;
1945 }
1946
1947 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1948 {
1949         struct throtl_data *td = tg->td;
1950         unsigned long now = jiffies;
1951
1952         /*
1953          * If cgroup is below low limit, consider downgrade and throttle other
1954          * cgroups
1955          */
1956         if (time_after_eq(now, tg_last_low_overflow_time(tg) +
1957                                         td->throtl_slice) &&
1958             (!throtl_tg_is_idle(tg) ||
1959              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1960                 return true;
1961         return false;
1962 }
1963
1964 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1965 {
1966         struct throtl_data *td = tg->td;
1967
1968         if (time_before(jiffies, td->low_upgrade_time + td->throtl_slice))
1969                 return false;
1970
1971         while (true) {
1972                 if (!throtl_tg_can_downgrade(tg))
1973                         return false;
1974                 tg = sq_to_tg(tg->service_queue.parent_sq);
1975                 if (!tg || !tg_to_blkg(tg)->parent)
1976                         break;
1977         }
1978         return true;
1979 }
1980
1981 static void throtl_downgrade_check(struct throtl_grp *tg)
1982 {
1983         uint64_t bps;
1984         unsigned int iops;
1985         unsigned long elapsed_time;
1986         unsigned long now = jiffies;
1987
1988         if (tg->td->limit_index != LIMIT_MAX ||
1989             !tg->td->limit_valid[LIMIT_LOW])
1990                 return;
1991         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1992                 return;
1993         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1994                 return;
1995
1996         elapsed_time = now - tg->last_check_time;
1997         tg->last_check_time = now;
1998
1999         if (time_before(now, tg_last_low_overflow_time(tg) +
2000                         tg->td->throtl_slice))
2001                 return;
2002
2003         if (tg->bps[READ][LIMIT_LOW]) {
2004                 bps = tg->last_bytes_disp[READ] * HZ;
2005                 do_div(bps, elapsed_time);
2006                 if (bps >= tg->bps[READ][LIMIT_LOW])
2007                         tg->last_low_overflow_time[READ] = now;
2008         }
2009
2010         if (tg->bps[WRITE][LIMIT_LOW]) {
2011                 bps = tg->last_bytes_disp[WRITE] * HZ;
2012                 do_div(bps, elapsed_time);
2013                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2014                         tg->last_low_overflow_time[WRITE] = now;
2015         }
2016
2017         if (tg->iops[READ][LIMIT_LOW]) {
2018                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2019                 if (iops >= tg->iops[READ][LIMIT_LOW])
2020                         tg->last_low_overflow_time[READ] = now;
2021         }
2022
2023         if (tg->iops[WRITE][LIMIT_LOW]) {
2024                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2025                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2026                         tg->last_low_overflow_time[WRITE] = now;
2027         }
2028
2029         /*
2030          * If cgroup is below low limit, consider downgrade and throttle other
2031          * cgroups
2032          */
2033         if (throtl_hierarchy_can_downgrade(tg))
2034                 throtl_downgrade_state(tg->td);
2035
2036         tg->last_bytes_disp[READ] = 0;
2037         tg->last_bytes_disp[WRITE] = 0;
2038         tg->last_io_disp[READ] = 0;
2039         tg->last_io_disp[WRITE] = 0;
2040 }
2041
2042 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2043 {
2044         unsigned long now;
2045         unsigned long last_finish_time = tg->last_finish_time;
2046
2047         if (last_finish_time == 0)
2048                 return;
2049
2050         now = ktime_get_ns() >> 10;
2051         if (now <= last_finish_time ||
2052             last_finish_time == tg->checked_last_finish_time)
2053                 return;
2054
2055         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2056         tg->checked_last_finish_time = last_finish_time;
2057 }
2058
2059 static void throtl_update_latency_buckets(struct throtl_data *td)
2060 {
2061         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2062         int i, cpu, rw;
2063         unsigned long last_latency[2] = { 0 };
2064         unsigned long latency[2];
2065
2066         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2067                 return;
2068         if (time_before(jiffies, td->last_calculate_time + HZ))
2069                 return;
2070         td->last_calculate_time = jiffies;
2071
2072         memset(avg_latency, 0, sizeof(avg_latency));
2073         for (rw = READ; rw <= WRITE; rw++) {
2074                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2075                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2076
2077                         for_each_possible_cpu(cpu) {
2078                                 struct latency_bucket *bucket;
2079
2080                                 /* this isn't race free, but ok in practice */
2081                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2082                                         cpu);
2083                                 tmp->total_latency += bucket[i].total_latency;
2084                                 tmp->samples += bucket[i].samples;
2085                                 bucket[i].total_latency = 0;
2086                                 bucket[i].samples = 0;
2087                         }
2088
2089                         if (tmp->samples >= 32) {
2090                                 int samples = tmp->samples;
2091
2092                                 latency[rw] = tmp->total_latency;
2093
2094                                 tmp->total_latency = 0;
2095                                 tmp->samples = 0;
2096                                 latency[rw] /= samples;
2097                                 if (latency[rw] == 0)
2098                                         continue;
2099                                 avg_latency[rw][i].latency = latency[rw];
2100                         }
2101                 }
2102         }
2103
2104         for (rw = READ; rw <= WRITE; rw++) {
2105                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2106                         if (!avg_latency[rw][i].latency) {
2107                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2108                                         td->avg_buckets[rw][i].latency =
2109                                                 last_latency[rw];
2110                                 continue;
2111                         }
2112
2113                         if (!td->avg_buckets[rw][i].valid)
2114                                 latency[rw] = avg_latency[rw][i].latency;
2115                         else
2116                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2117                                         avg_latency[rw][i].latency) >> 3;
2118
2119                         td->avg_buckets[rw][i].latency = max(latency[rw],
2120                                 last_latency[rw]);
2121                         td->avg_buckets[rw][i].valid = true;
2122                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2123                 }
2124         }
2125
2126         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2127                 throtl_log(&td->service_queue,
2128                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2129                         "write latency=%ld, write valid=%d", i,
2130                         td->avg_buckets[READ][i].latency,
2131                         td->avg_buckets[READ][i].valid,
2132                         td->avg_buckets[WRITE][i].latency,
2133                         td->avg_buckets[WRITE][i].valid);
2134 }
2135 #else
2136 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2137 {
2138 }
2139
2140 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2141 {
2142 }
2143
2144 static void throtl_downgrade_check(struct throtl_grp *tg)
2145 {
2146 }
2147
2148 static void throtl_upgrade_check(struct throtl_grp *tg)
2149 {
2150 }
2151
2152 static bool throtl_can_upgrade(struct throtl_data *td,
2153         struct throtl_grp *this_tg)
2154 {
2155         return false;
2156 }
2157
2158 static void throtl_upgrade_state(struct throtl_data *td)
2159 {
2160 }
2161 #endif
2162
2163 bool __blk_throtl_bio(struct bio *bio)
2164 {
2165         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2166         struct blkcg_gq *blkg = bio->bi_blkg;
2167         struct throtl_qnode *qn = NULL;
2168         struct throtl_grp *tg = blkg_to_tg(blkg);
2169         struct throtl_service_queue *sq;
2170         bool rw = bio_data_dir(bio);
2171         bool throttled = false;
2172         struct throtl_data *td = tg->td;
2173
2174         rcu_read_lock();
2175
2176         spin_lock_irq(&q->queue_lock);
2177
2178         throtl_update_latency_buckets(td);
2179
2180         blk_throtl_update_idletime(tg);
2181
2182         sq = &tg->service_queue;
2183
2184 again:
2185         while (true) {
2186                 if (tg->last_low_overflow_time[rw] == 0)
2187                         tg->last_low_overflow_time[rw] = jiffies;
2188                 throtl_downgrade_check(tg);
2189                 throtl_upgrade_check(tg);
2190                 /* throtl is FIFO - if bios are already queued, should queue */
2191                 if (sq->nr_queued[rw])
2192                         break;
2193
2194                 /* if above limits, break to queue */
2195                 if (!tg_may_dispatch(tg, bio, NULL)) {
2196                         tg->last_low_overflow_time[rw] = jiffies;
2197                         if (throtl_can_upgrade(td, tg)) {
2198                                 throtl_upgrade_state(td);
2199                                 goto again;
2200                         }
2201                         break;
2202                 }
2203
2204                 /* within limits, let's charge and dispatch directly */
2205                 throtl_charge_bio(tg, bio);
2206
2207                 /*
2208                  * We need to trim slice even when bios are not being queued
2209                  * otherwise it might happen that a bio is not queued for
2210                  * a long time and slice keeps on extending and trim is not
2211                  * called for a long time. Now if limits are reduced suddenly
2212                  * we take into account all the IO dispatched so far at new
2213                  * low rate and * newly queued IO gets a really long dispatch
2214                  * time.
2215                  *
2216                  * So keep on trimming slice even if bio is not queued.
2217                  */
2218                 throtl_trim_slice(tg, rw);
2219
2220                 /*
2221                  * @bio passed through this layer without being throttled.
2222                  * Climb up the ladder.  If we're already at the top, it
2223                  * can be executed directly.
2224                  */
2225                 qn = &tg->qnode_on_parent[rw];
2226                 sq = sq->parent_sq;
2227                 tg = sq_to_tg(sq);
2228                 if (!tg) {
2229                         bio_set_flag(bio, BIO_BPS_THROTTLED);
2230                         goto out_unlock;
2231                 }
2232         }
2233
2234         /* out-of-limit, queue to @tg */
2235         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2236                    rw == READ ? 'R' : 'W',
2237                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2238                    tg_bps_limit(tg, rw),
2239                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2240                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2241
2242         tg->last_low_overflow_time[rw] = jiffies;
2243
2244         td->nr_queued[rw]++;
2245         throtl_add_bio_tg(bio, qn, tg);
2246         throttled = true;
2247
2248         /*
2249          * Update @tg's dispatch time and force schedule dispatch if @tg
2250          * was empty before @bio.  The forced scheduling isn't likely to
2251          * cause undue delay as @bio is likely to be dispatched directly if
2252          * its @tg's disptime is not in the future.
2253          */
2254         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2255                 tg_update_disptime(tg);
2256                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2257         }
2258
2259 out_unlock:
2260 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2261         if (throttled || !td->track_bio_latency)
2262                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2263 #endif
2264         spin_unlock_irq(&q->queue_lock);
2265
2266         rcu_read_unlock();
2267         return throttled;
2268 }
2269
2270 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2271 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2272                                  enum req_op op, unsigned long time)
2273 {
2274         const bool rw = op_is_write(op);
2275         struct latency_bucket *latency;
2276         int index;
2277
2278         if (!td || td->limit_index != LIMIT_LOW ||
2279             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2280             !blk_queue_nonrot(td->queue))
2281                 return;
2282
2283         index = request_bucket_index(size);
2284
2285         latency = get_cpu_ptr(td->latency_buckets[rw]);
2286         latency[index].total_latency += time;
2287         latency[index].samples++;
2288         put_cpu_ptr(td->latency_buckets[rw]);
2289 }
2290
2291 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2292 {
2293         struct request_queue *q = rq->q;
2294         struct throtl_data *td = q->td;
2295
2296         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2297                              time_ns >> 10);
2298 }
2299
2300 void blk_throtl_bio_endio(struct bio *bio)
2301 {
2302         struct blkcg_gq *blkg;
2303         struct throtl_grp *tg;
2304         u64 finish_time_ns;
2305         unsigned long finish_time;
2306         unsigned long start_time;
2307         unsigned long lat;
2308         int rw = bio_data_dir(bio);
2309
2310         blkg = bio->bi_blkg;
2311         if (!blkg)
2312                 return;
2313         tg = blkg_to_tg(blkg);
2314         if (!tg->td->limit_valid[LIMIT_LOW])
2315                 return;
2316
2317         finish_time_ns = ktime_get_ns();
2318         tg->last_finish_time = finish_time_ns >> 10;
2319
2320         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2321         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2322         if (!start_time || finish_time <= start_time)
2323                 return;
2324
2325         lat = finish_time - start_time;
2326         /* this is only for bio based driver */
2327         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2328                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2329                                      bio_op(bio), lat);
2330
2331         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2332                 int bucket;
2333                 unsigned int threshold;
2334
2335                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2336                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2337                         tg->latency_target;
2338                 if (lat > threshold)
2339                         tg->bad_bio_cnt++;
2340                 /*
2341                  * Not race free, could get wrong count, which means cgroups
2342                  * will be throttled
2343                  */
2344                 tg->bio_cnt++;
2345         }
2346
2347         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2348                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2349                 tg->bio_cnt /= 2;
2350                 tg->bad_bio_cnt /= 2;
2351         }
2352 }
2353 #endif
2354
2355 int blk_throtl_init(struct gendisk *disk)
2356 {
2357         struct request_queue *q = disk->queue;
2358         struct throtl_data *td;
2359         int ret;
2360
2361         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2362         if (!td)
2363                 return -ENOMEM;
2364         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2365                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2366         if (!td->latency_buckets[READ]) {
2367                 kfree(td);
2368                 return -ENOMEM;
2369         }
2370         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2371                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2372         if (!td->latency_buckets[WRITE]) {
2373                 free_percpu(td->latency_buckets[READ]);
2374                 kfree(td);
2375                 return -ENOMEM;
2376         }
2377
2378         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2379         throtl_service_queue_init(&td->service_queue);
2380
2381         q->td = td;
2382         td->queue = q;
2383
2384         td->limit_valid[LIMIT_MAX] = true;
2385         td->limit_index = LIMIT_MAX;
2386         td->low_upgrade_time = jiffies;
2387         td->low_downgrade_time = jiffies;
2388
2389         /* activate policy */
2390         ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
2391         if (ret) {
2392                 free_percpu(td->latency_buckets[READ]);
2393                 free_percpu(td->latency_buckets[WRITE]);
2394                 kfree(td);
2395         }
2396         return ret;
2397 }
2398
2399 void blk_throtl_exit(struct gendisk *disk)
2400 {
2401         struct request_queue *q = disk->queue;
2402
2403         BUG_ON(!q->td);
2404         del_timer_sync(&q->td->service_queue.pending_timer);
2405         throtl_shutdown_wq(q);
2406         blkcg_deactivate_policy(disk, &blkcg_policy_throtl);
2407         free_percpu(q->td->latency_buckets[READ]);
2408         free_percpu(q->td->latency_buckets[WRITE]);
2409         kfree(q->td);
2410 }
2411
2412 void blk_throtl_register(struct gendisk *disk)
2413 {
2414         struct request_queue *q = disk->queue;
2415         struct throtl_data *td;
2416         int i;
2417
2418         td = q->td;
2419         BUG_ON(!td);
2420
2421         if (blk_queue_nonrot(q)) {
2422                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2423                 td->filtered_latency = LATENCY_FILTERED_SSD;
2424         } else {
2425                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2426                 td->filtered_latency = LATENCY_FILTERED_HD;
2427                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2428                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2429                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2430                 }
2431         }
2432 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2433         /* if no low limit, use previous default */
2434         td->throtl_slice = DFL_THROTL_SLICE_HD;
2435
2436 #else
2437         td->track_bio_latency = !queue_is_mq(q);
2438         if (!td->track_bio_latency)
2439                 blk_stat_enable_accounting(q);
2440 #endif
2441 }
2442
2443 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2444 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2445 {
2446         if (!q->td)
2447                 return -EINVAL;
2448         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2449 }
2450
2451 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2452         const char *page, size_t count)
2453 {
2454         unsigned long v;
2455         unsigned long t;
2456
2457         if (!q->td)
2458                 return -EINVAL;
2459         if (kstrtoul(page, 10, &v))
2460                 return -EINVAL;
2461         t = msecs_to_jiffies(v);
2462         if (t == 0 || t > MAX_THROTL_SLICE)
2463                 return -EINVAL;
2464         q->td->throtl_slice = t;
2465         return count;
2466 }
2467 #endif
2468
2469 static int __init throtl_init(void)
2470 {
2471         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2472         if (!kthrotld_workqueue)
2473                 panic("Failed to create kthrotld\n");
2474
2475         return blkcg_policy_register(&blkcg_policy_throtl);
2476 }
2477
2478 module_init(throtl_init);