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 lapsed 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(gfp_t gfp,
 339                                                 struct request_queue *q,
 340                                                 struct blkcg *blkcg)
 341 {
 342         struct throtl_grp *tg;
 343         int rw;
 344
 345         tg = kzalloc_node(sizeof(*tg), gfp, q->node);
 346         if (!tg)
 347                 return NULL;
 348
 349         if (blkg_rwstat_init(&tg->stat_bytes, gfp))
 350                 goto err_free_tg;
 351
 352         if (blkg_rwstat_init(&tg->stat_ios, gfp))
 353                 goto err_exit_stat_bytes;
 354
 355         throtl_service_queue_init(&tg->service_queue);
 356
 357         for (rw = READ; rw <= WRITE; rw++) {
 358                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 359                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 360         }
 361
 362         RB_CLEAR_NODE(&tg->rb_node);
 363         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 364         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 365         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 366         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 367         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 368         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 369         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 370         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 371         /* LIMIT_LOW will have default value 0 */
 372
 373         tg->latency_target = DFL_LATENCY_TARGET;
 374         tg->latency_target_conf = DFL_LATENCY_TARGET;
 375         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 376         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 377
 378         return &tg->pd;
 379
 380 err_exit_stat_bytes:
 381         blkg_rwstat_exit(&tg->stat_bytes);
 382 err_free_tg:
 383         kfree(tg);
 384         return NULL;
 385 }
 386
 387 static void throtl_pd_init(struct blkg_policy_data *pd)
 388 {
 389         struct throtl_grp *tg = pd_to_tg(pd);
 390         struct blkcg_gq *blkg = tg_to_blkg(tg);
 391         struct throtl_data *td = blkg->q->td;
 392         struct throtl_service_queue *sq = &tg->service_queue;
 393
 394         /*
 395          * If on the default hierarchy, we switch to properly hierarchical
 396          * behavior where limits on a given throtl_grp are applied to the
 397          * whole subtree rather than just the group itself.  e.g. If 16M
 398          * read_bps limit is set on the root group, the whole system can't
 399          * exceed 16M for the 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_eq(start, tg->slice_start[rw]))
 648                 tg->slice_start[rw] = start;
 649
 650         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 651         throtl_log(&tg->service_queue,
 652                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 653                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 654                    tg->slice_end[rw], jiffies);
 655 }
 656
 657 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
 658                                           bool clear_carryover)
 659 {
 660         tg->bytes_disp[rw] = 0;
 661         tg->io_disp[rw] = 0;
 662         tg->slice_start[rw] = jiffies;
 663         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 664         if (clear_carryover) {
 665                 tg->carryover_bytes[rw] = 0;
 666                 tg->carryover_ios[rw] = 0;
 667         }
 668
 669         throtl_log(&tg->service_queue,
 670                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 671                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 672                    tg->slice_end[rw], jiffies);
 673 }
 674
 675 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 676                                         unsigned long jiffy_end)
 677 {
 678         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 679 }
 680
 681 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 682                                        unsigned long jiffy_end)
 683 {
 684         throtl_set_slice_end(tg, rw, jiffy_end);
 685         throtl_log(&tg->service_queue,
 686                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 687                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 688                    tg->slice_end[rw], jiffies);
 689 }
 690
 691 /* Determine if previously allocated or extended slice is complete or not */
 692 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 693 {
 694         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 695                 return false;
 696
 697         return true;
 698 }
 699
 700 /* Trim the used slices and adjust slice start accordingly */
 701 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 702 {
 703         unsigned long nr_slices, time_elapsed, io_trim;
 704         u64 bytes_trim, tmp;
 705
 706         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 707
 708         /*
 709          * If bps are unlimited (-1), then time slice don't get
 710          * renewed. Don't try to trim the slice if slice is used. A new
 711          * slice will start when appropriate.
 712          */
 713         if (throtl_slice_used(tg, rw))
 714                 return;
 715
 716         /*
 717          * A bio has been dispatched. Also adjust slice_end. It might happen
 718          * that initially cgroup limit was very low resulting in high
 719          * slice_end, but later limit was bumped up and bio was dispatched
 720          * sooner, then we need to reduce slice_end. A high bogus slice_end
 721          * is bad because it does not allow new slice to start.
 722          */
 723
 724         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 725
 726         time_elapsed = jiffies - tg->slice_start[rw];
 727
 728         nr_slices = time_elapsed / tg->td->throtl_slice;
 729
 730         if (!nr_slices)
 731                 return;
 732         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 733         do_div(tmp, HZ);
 734         bytes_trim = tmp;
 735
 736         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 737                 HZ;
 738
 739         if (!bytes_trim && !io_trim)
 740                 return;
 741
 742         if (tg->bytes_disp[rw] >= bytes_trim)
 743                 tg->bytes_disp[rw] -= bytes_trim;
 744         else
 745                 tg->bytes_disp[rw] = 0;
 746
 747         if (tg->io_disp[rw] >= io_trim)
 748                 tg->io_disp[rw] -= io_trim;
 749         else
 750                 tg->io_disp[rw] = 0;
 751
 752         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 753
 754         throtl_log(&tg->service_queue,
 755                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 756                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 757                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 758 }
 759
 760 static unsigned int calculate_io_allowed(u32 iops_limit,
 761                                          unsigned long jiffy_elapsed)
 762 {
 763         unsigned int io_allowed;
 764         u64 tmp;
 765
 766         /*
 767          * jiffy_elapsed should not be a big value as minimum iops can be
 768          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 769          * will allow dispatch after 1 second and after that slice should
 770          * have been trimmed.
 771          */
 772
 773         tmp = (u64)iops_limit * jiffy_elapsed;
 774         do_div(tmp, HZ);
 775
 776         if (tmp > UINT_MAX)
 777                 io_allowed = UINT_MAX;
 778         else
 779                 io_allowed = tmp;
 780
 781         return io_allowed;
 782 }
 783
 784 static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
 785 {
 786         return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
 787 }
 788
 789 static void __tg_update_carryover(struct throtl_grp *tg, bool rw)
 790 {
 791         unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
 792         u64 bps_limit = tg_bps_limit(tg, rw);
 793         u32 iops_limit = tg_iops_limit(tg, rw);
 794
 795         /*
 796          * If config is updated while bios are still throttled, calculate and
 797          * accumulate how many bytes/ios are waited across changes. And
 798          * carryover_bytes/ios will be used to calculate new wait time under new
 799          * configuration.
 800          */
 801         if (bps_limit != U64_MAX)
 802                 tg->carryover_bytes[rw] +=
 803                         calculate_bytes_allowed(bps_limit, jiffy_elapsed) -
 804                         tg->bytes_disp[rw];
 805         if (iops_limit != UINT_MAX)
 806                 tg->carryover_ios[rw] +=
 807                         calculate_io_allowed(iops_limit, jiffy_elapsed) -
 808                         tg->io_disp[rw];
 809 }
 810
 811 static void tg_update_carryover(struct throtl_grp *tg)
 812 {
 813         if (tg->service_queue.nr_queued[READ])
 814                 __tg_update_carryover(tg, READ);
 815         if (tg->service_queue.nr_queued[WRITE])
 816                 __tg_update_carryover(tg, WRITE);
 817
 818         /* see comments in struct throtl_grp for meaning of these fields. */
 819         throtl_log(&tg->service_queue, "%s: %llu %llu %u %u\n", __func__,
 820                    tg->carryover_bytes[READ], tg->carryover_bytes[WRITE],
 821                    tg->carryover_ios[READ], tg->carryover_ios[WRITE]);
 822 }
 823
 824 static bool tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
 825                                  u32 iops_limit, unsigned long *wait)
 826 {
 827         bool rw = bio_data_dir(bio);
 828         unsigned int io_allowed;
 829         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 830
 831         if (iops_limit == UINT_MAX) {
 832                 if (wait)
 833                         *wait = 0;
 834                 return true;
 835         }
 836
 837         jiffy_elapsed = jiffies - tg->slice_start[rw];
 838
 839         /* Round up to the next throttle slice, wait time must be nonzero */
 840         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 841         io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd) +
 842                      tg->carryover_ios[rw];
 843         if (tg->io_disp[rw] + 1 <= io_allowed) {
 844                 if (wait)
 845                         *wait = 0;
 846                 return true;
 847         }
 848
 849         /* Calc approx time to dispatch */
 850         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 851
 852         if (wait)
 853                 *wait = jiffy_wait;
 854         return false;
 855 }
 856
 857 static bool tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
 858                                 u64 bps_limit, unsigned long *wait)
 859 {
 860         bool rw = bio_data_dir(bio);
 861         u64 bytes_allowed, extra_bytes;
 862         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 863         unsigned int bio_size = throtl_bio_data_size(bio);
 864
 865         /* no need to throttle if this bio's bytes have been accounted */
 866         if (bps_limit == U64_MAX || bio_flagged(bio, BIO_BPS_THROTTLED)) {
 867                 if (wait)
 868                         *wait = 0;
 869                 return true;
 870         }
 871
 872         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 873
 874         /* Slice has just started. Consider one slice interval */
 875         if (!jiffy_elapsed)
 876                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 877
 878         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 879         bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd) +
 880                         tg->carryover_bytes[rw];
 881         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 882                 if (wait)
 883                         *wait = 0;
 884                 return true;
 885         }
 886
 887         /* Calc approx time to dispatch */
 888         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 889         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 890
 891         if (!jiffy_wait)
 892                 jiffy_wait = 1;
 893
 894         /*
 895          * This wait time is without taking into consideration the rounding
 896          * up we did. Add that time also.
 897          */
 898         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 899         if (wait)
 900                 *wait = jiffy_wait;
 901         return false;
 902 }
 903
 904 /*
 905  * Returns whether one can dispatch a bio or not. Also returns approx number
 906  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 907  */
 908 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 909                             unsigned long *wait)
 910 {
 911         bool rw = bio_data_dir(bio);
 912         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 913         u64 bps_limit = tg_bps_limit(tg, rw);
 914         u32 iops_limit = tg_iops_limit(tg, rw);
 915
 916         /*
 917          * Currently whole state machine of group depends on first bio
 918          * queued in the group bio list. So one should not be calling
 919          * this function with a different bio if there are other bios
 920          * queued.
 921          */
 922         BUG_ON(tg->service_queue.nr_queued[rw] &&
 923                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 924
 925         /* If tg->bps = -1, then BW is unlimited */
 926         if ((bps_limit == U64_MAX && iops_limit == UINT_MAX) ||
 927             tg->flags & THROTL_TG_CANCELING) {
 928                 if (wait)
 929                         *wait = 0;
 930                 return true;
 931         }
 932
 933         /*
 934          * If previous slice expired, start a new one otherwise renew/extend
 935          * existing slice to make sure it is at least throtl_slice interval
 936          * long since now. New slice is started only for empty throttle group.
 937          * If there is queued bio, that means there should be an active
 938          * slice and it should be extended instead.
 939          */
 940         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 941                 throtl_start_new_slice(tg, rw, true);
 942         else {
 943                 if (time_before(tg->slice_end[rw],
 944                     jiffies + tg->td->throtl_slice))
 945                         throtl_extend_slice(tg, rw,
 946                                 jiffies + tg->td->throtl_slice);
 947         }
 948
 949         if (tg_within_bps_limit(tg, bio, bps_limit, &bps_wait) &&
 950             tg_within_iops_limit(tg, bio, iops_limit, &iops_wait)) {
 951                 if (wait)
 952                         *wait = 0;
 953                 return true;
 954         }
 955
 956         max_wait = max(bps_wait, iops_wait);
 957
 958         if (wait)
 959                 *wait = max_wait;
 960
 961         if (time_before(tg->slice_end[rw], jiffies + max_wait))
 962                 throtl_extend_slice(tg, rw, jiffies + max_wait);
 963
 964         return false;
 965 }
 966
 967 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 968 {
 969         bool rw = bio_data_dir(bio);
 970         unsigned int bio_size = throtl_bio_data_size(bio);
 971
 972         /* Charge the bio to the group */
 973         if (!bio_flagged(bio, BIO_BPS_THROTTLED)) {
 974                 tg->bytes_disp[rw] += bio_size;
 975                 tg->last_bytes_disp[rw] += bio_size;
 976         }
 977
 978         tg->io_disp[rw]++;
 979         tg->last_io_disp[rw]++;
 980 }
 981
 982 /**
 983  * throtl_add_bio_tg - add a bio to the specified throtl_grp
 984  * @bio: bio to add
 985  * @qn: qnode to use
 986  * @tg: the target throtl_grp
 987  *
 988  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 989  * tg->qnode_on_self[] is used.
 990  */
 991 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 992                               struct throtl_grp *tg)
 993 {
 994         struct throtl_service_queue *sq = &tg->service_queue;
 995         bool rw = bio_data_dir(bio);
 996
 997         if (!qn)
 998                 qn = &tg->qnode_on_self[rw];
 999
1000         /*
1001          * If @tg doesn't currently have any bios queued in the same
1002          * direction, queueing @bio can change when @tg should be
1003          * dispatched.  Mark that @tg was empty.  This is automatically
1004          * cleared on the next tg_update_disptime().
1005          */
1006         if (!sq->nr_queued[rw])
1007                 tg->flags |= THROTL_TG_WAS_EMPTY;
1008
1009         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1010
1011         sq->nr_queued[rw]++;
1012         throtl_enqueue_tg(tg);
1013 }
1014
1015 static void tg_update_disptime(struct throtl_grp *tg)
1016 {
1017         struct throtl_service_queue *sq = &tg->service_queue;
1018         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1019         struct bio *bio;
1020
1021         bio = throtl_peek_queued(&sq->queued[READ]);
1022         if (bio)
1023                 tg_may_dispatch(tg, bio, &read_wait);
1024
1025         bio = throtl_peek_queued(&sq->queued[WRITE]);
1026         if (bio)
1027                 tg_may_dispatch(tg, bio, &write_wait);
1028
1029         min_wait = min(read_wait, write_wait);
1030         disptime = jiffies + min_wait;
1031
1032         /* Update dispatch time */
1033         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
1034         tg->disptime = disptime;
1035         tg_service_queue_add(tg);
1036
1037         /* see throtl_add_bio_tg() */
1038         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1039 }
1040
1041 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1042                                         struct throtl_grp *parent_tg, bool rw)
1043 {
1044         if (throtl_slice_used(parent_tg, rw)) {
1045                 throtl_start_new_slice_with_credit(parent_tg, rw,
1046                                 child_tg->slice_start[rw]);
1047         }
1048
1049 }
1050
1051 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1052 {
1053         struct throtl_service_queue *sq = &tg->service_queue;
1054         struct throtl_service_queue *parent_sq = sq->parent_sq;
1055         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1056         struct throtl_grp *tg_to_put = NULL;
1057         struct bio *bio;
1058
1059         /*
1060          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1061          * from @tg may put its reference and @parent_sq might end up
1062          * getting released prematurely.  Remember the tg to put and put it
1063          * after @bio is transferred to @parent_sq.
1064          */
1065         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1066         sq->nr_queued[rw]--;
1067
1068         throtl_charge_bio(tg, bio);
1069         bio_set_flag(bio, BIO_BPS_THROTTLED);
1070
1071         /*
1072          * If our parent is another tg, we just need to transfer @bio to
1073          * the parent using throtl_add_bio_tg().  If our parent is
1074          * @td->service_queue, @bio is ready to be issued.  Put it on its
1075          * bio_lists[] and decrease total number queued.  The caller is
1076          * responsible for issuing these bios.
1077          */
1078         if (parent_tg) {
1079                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1080                 start_parent_slice_with_credit(tg, parent_tg, rw);
1081         } else {
1082                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1083                                      &parent_sq->queued[rw]);
1084                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1085                 tg->td->nr_queued[rw]--;
1086         }
1087
1088         throtl_trim_slice(tg, rw);
1089
1090         if (tg_to_put)
1091                 blkg_put(tg_to_blkg(tg_to_put));
1092 }
1093
1094 static int throtl_dispatch_tg(struct throtl_grp *tg)
1095 {
1096         struct throtl_service_queue *sq = &tg->service_queue;
1097         unsigned int nr_reads = 0, nr_writes = 0;
1098         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1099         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1100         struct bio *bio;
1101
1102         /* Try to dispatch 75% READS and 25% WRITES */
1103
1104         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1105                tg_may_dispatch(tg, bio, NULL)) {
1106
1107                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1108                 nr_reads++;
1109
1110                 if (nr_reads >= max_nr_reads)
1111                         break;
1112         }
1113
1114         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1115                tg_may_dispatch(tg, bio, NULL)) {
1116
1117                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1118                 nr_writes++;
1119
1120                 if (nr_writes >= max_nr_writes)
1121                         break;
1122         }
1123
1124         return nr_reads + nr_writes;
1125 }
1126
1127 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1128 {
1129         unsigned int nr_disp = 0;
1130
1131         while (1) {
1132                 struct throtl_grp *tg;
1133                 struct throtl_service_queue *sq;
1134
1135                 if (!parent_sq->nr_pending)
1136                         break;
1137
1138                 tg = throtl_rb_first(parent_sq);
1139                 if (!tg)
1140                         break;
1141
1142                 if (time_before(jiffies, tg->disptime))
1143                         break;
1144
1145                 nr_disp += throtl_dispatch_tg(tg);
1146
1147                 sq = &tg->service_queue;
1148                 if (sq->nr_queued[READ] || sq->nr_queued[WRITE])
1149                         tg_update_disptime(tg);
1150                 else
1151                         throtl_dequeue_tg(tg);
1152
1153                 if (nr_disp >= THROTL_QUANTUM)
1154                         break;
1155         }
1156
1157         return nr_disp;
1158 }
1159
1160 static bool throtl_can_upgrade(struct throtl_data *td,
1161         struct throtl_grp *this_tg);
1162 /**
1163  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1164  * @t: the pending_timer member of the throtl_service_queue being serviced
1165  *
1166  * This timer is armed when a child throtl_grp with active bio's become
1167  * pending and queued on the service_queue's pending_tree and expires when
1168  * the first child throtl_grp should be dispatched.  This function
1169  * dispatches bio's from the children throtl_grps to the parent
1170  * service_queue.
1171  *
1172  * If the parent's parent is another throtl_grp, dispatching is propagated
1173  * by either arming its pending_timer or repeating dispatch directly.  If
1174  * the top-level service_tree is reached, throtl_data->dispatch_work is
1175  * kicked so that the ready bio's are issued.
1176  */
1177 static void throtl_pending_timer_fn(struct timer_list *t)
1178 {
1179         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1180         struct throtl_grp *tg = sq_to_tg(sq);
1181         struct throtl_data *td = sq_to_td(sq);
1182         struct throtl_service_queue *parent_sq;
1183         struct request_queue *q;
1184         bool dispatched;
1185         int ret;
1186
1187         /* throtl_data may be gone, so figure out request queue by blkg */
1188         if (tg)
1189                 q = tg->pd.blkg->q;
1190         else
1191                 q = td->queue;
1192
1193         spin_lock_irq(&q->queue_lock);
1194
1195         if (!q->root_blkg)
1196                 goto out_unlock;
1197
1198         if (throtl_can_upgrade(td, NULL))
1199                 throtl_upgrade_state(td);
1200
1201 again:
1202         parent_sq = sq->parent_sq;
1203         dispatched = false;
1204
1205         while (true) {
1206                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1207                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1208                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1209
1210                 ret = throtl_select_dispatch(sq);
1211                 if (ret) {
1212                         throtl_log(sq, "bios disp=%u", ret);
1213                         dispatched = true;
1214                 }
1215
1216                 if (throtl_schedule_next_dispatch(sq, false))
1217                         break;
1218
1219                 /* this dispatch windows is still open, relax and repeat */
1220                 spin_unlock_irq(&q->queue_lock);
1221                 cpu_relax();
1222                 spin_lock_irq(&q->queue_lock);
1223         }
1224
1225         if (!dispatched)
1226                 goto out_unlock;
1227
1228         if (parent_sq) {
1229                 /* @parent_sq is another throl_grp, propagate dispatch */
1230                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1231                         tg_update_disptime(tg);
1232                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1233                                 /* window is already open, repeat dispatching */
1234                                 sq = parent_sq;
1235                                 tg = sq_to_tg(sq);
1236                                 goto again;
1237                         }
1238                 }
1239         } else {
1240                 /* reached the top-level, queue issuing */
1241                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1242         }
1243 out_unlock:
1244         spin_unlock_irq(&q->queue_lock);
1245 }
1246
1247 /**
1248  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1249  * @work: work item being executed
1250  *
1251  * This function is queued for execution when bios reach the bio_lists[]
1252  * of throtl_data->service_queue.  Those bios are ready and issued by this
1253  * function.
1254  */
1255 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1256 {
1257         struct throtl_data *td = container_of(work, struct throtl_data,
1258                                               dispatch_work);
1259         struct throtl_service_queue *td_sq = &td->service_queue;
1260         struct request_queue *q = td->queue;
1261         struct bio_list bio_list_on_stack;
1262         struct bio *bio;
1263         struct blk_plug plug;
1264         int rw;
1265
1266         bio_list_init(&bio_list_on_stack);
1267
1268         spin_lock_irq(&q->queue_lock);
1269         for (rw = READ; rw <= WRITE; rw++)
1270                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1271                         bio_list_add(&bio_list_on_stack, bio);
1272         spin_unlock_irq(&q->queue_lock);
1273
1274         if (!bio_list_empty(&bio_list_on_stack)) {
1275                 blk_start_plug(&plug);
1276                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1277                         submit_bio_noacct_nocheck(bio);
1278                 blk_finish_plug(&plug);
1279         }
1280 }
1281
1282 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1283                               int off)
1284 {
1285         struct throtl_grp *tg = pd_to_tg(pd);
1286         u64 v = *(u64 *)((void *)tg + off);
1287
1288         if (v == U64_MAX)
1289                 return 0;
1290         return __blkg_prfill_u64(sf, pd, v);
1291 }
1292
1293 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1294                                int off)
1295 {
1296         struct throtl_grp *tg = pd_to_tg(pd);
1297         unsigned int v = *(unsigned int *)((void *)tg + off);
1298
1299         if (v == UINT_MAX)
1300                 return 0;
1301         return __blkg_prfill_u64(sf, pd, v);
1302 }
1303
1304 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1305 {
1306         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1307                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1308         return 0;
1309 }
1310
1311 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1312 {
1313         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1314                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1315         return 0;
1316 }
1317
1318 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1319 {
1320         struct throtl_service_queue *sq = &tg->service_queue;
1321         struct cgroup_subsys_state *pos_css;
1322         struct blkcg_gq *blkg;
1323
1324         throtl_log(&tg->service_queue,
1325                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1326                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1327                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1328
1329         /*
1330          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1331          * considered to have rules if either the tg itself or any of its
1332          * ancestors has rules.  This identifies groups without any
1333          * restrictions in the whole hierarchy and allows them to bypass
1334          * blk-throttle.
1335          */
1336         blkg_for_each_descendant_pre(blkg, pos_css,
1337                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1338                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1339                 struct throtl_grp *parent_tg;
1340
1341                 tg_update_has_rules(this_tg);
1342                 /* ignore root/second level */
1343                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1344                     !blkg->parent->parent)
1345                         continue;
1346                 parent_tg = blkg_to_tg(blkg->parent);
1347                 /*
1348                  * make sure all children has lower idle time threshold and
1349                  * higher latency target
1350                  */
1351                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1352                                 parent_tg->idletime_threshold);
1353                 this_tg->latency_target = max(this_tg->latency_target,
1354                                 parent_tg->latency_target);
1355         }
1356
1357         /*
1358          * We're already holding queue_lock and know @tg is valid.  Let's
1359          * apply the new config directly.
1360          *
1361          * Restart the slices for both READ and WRITES. It might happen
1362          * that a group's limit are dropped suddenly and we don't want to
1363          * account recently dispatched IO with new low rate.
1364          */
1365         throtl_start_new_slice(tg, READ, false);
1366         throtl_start_new_slice(tg, WRITE, false);
1367
1368         if (tg->flags & THROTL_TG_PENDING) {
1369                 tg_update_disptime(tg);
1370                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1371         }
1372 }
1373
1374 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1375                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1376 {
1377         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1378         struct blkg_conf_ctx ctx;
1379         struct throtl_grp *tg;
1380         int ret;
1381         u64 v;
1382
1383         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1384         if (ret)
1385                 return ret;
1386
1387         ret = -EINVAL;
1388         if (sscanf(ctx.body, "%llu", &v) != 1)
1389                 goto out_finish;
1390         if (!v)
1391                 v = U64_MAX;
1392
1393         tg = blkg_to_tg(ctx.blkg);
1394         tg_update_carryover(tg);
1395
1396         if (is_u64)
1397                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1398         else
1399                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1400
1401         tg_conf_updated(tg, false);
1402         ret = 0;
1403 out_finish:
1404         blkg_conf_finish(&ctx);
1405         return ret ?: nbytes;
1406 }
1407
1408 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1409                                char *buf, size_t nbytes, loff_t off)
1410 {
1411         return tg_set_conf(of, buf, nbytes, off, true);
1412 }
1413
1414 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1415                                 char *buf, size_t nbytes, loff_t off)
1416 {
1417         return tg_set_conf(of, buf, nbytes, off, false);
1418 }
1419
1420 static int tg_print_rwstat(struct seq_file *sf, void *v)
1421 {
1422         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1423                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1424                           seq_cft(sf)->private, true);
1425         return 0;
1426 }
1427
1428 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1429                                       struct blkg_policy_data *pd, int off)
1430 {
1431         struct blkg_rwstat_sample sum;
1432
1433         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1434                                   &sum);
1435         return __blkg_prfill_rwstat(sf, pd, &sum);
1436 }
1437
1438 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1439 {
1440         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1441                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1442                           seq_cft(sf)->private, true);
1443         return 0;
1444 }
1445
1446 static struct cftype throtl_legacy_files[] = {
1447         {
1448                 .name = "throttle.read_bps_device",
1449                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1450                 .seq_show = tg_print_conf_u64,
1451                 .write = tg_set_conf_u64,
1452         },
1453         {
1454                 .name = "throttle.write_bps_device",
1455                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1456                 .seq_show = tg_print_conf_u64,
1457                 .write = tg_set_conf_u64,
1458         },
1459         {
1460                 .name = "throttle.read_iops_device",
1461                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1462                 .seq_show = tg_print_conf_uint,
1463                 .write = tg_set_conf_uint,
1464         },
1465         {
1466                 .name = "throttle.write_iops_device",
1467                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1468                 .seq_show = tg_print_conf_uint,
1469                 .write = tg_set_conf_uint,
1470         },
1471         {
1472                 .name = "throttle.io_service_bytes",
1473                 .private = offsetof(struct throtl_grp, stat_bytes),
1474                 .seq_show = tg_print_rwstat,
1475         },
1476         {
1477                 .name = "throttle.io_service_bytes_recursive",
1478                 .private = offsetof(struct throtl_grp, stat_bytes),
1479                 .seq_show = tg_print_rwstat_recursive,
1480         },
1481         {
1482                 .name = "throttle.io_serviced",
1483                 .private = offsetof(struct throtl_grp, stat_ios),
1484                 .seq_show = tg_print_rwstat,
1485         },
1486         {
1487                 .name = "throttle.io_serviced_recursive",
1488                 .private = offsetof(struct throtl_grp, stat_ios),
1489                 .seq_show = tg_print_rwstat_recursive,
1490         },
1491         { }     /* terminate */
1492 };
1493
1494 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1495                          int off)
1496 {
1497         struct throtl_grp *tg = pd_to_tg(pd);
1498         const char *dname = blkg_dev_name(pd->blkg);
1499         char bufs[4][21] = { "max", "max", "max", "max" };
1500         u64 bps_dft;
1501         unsigned int iops_dft;
1502         char idle_time[26] = "";
1503         char latency_time[26] = "";
1504
1505         if (!dname)
1506                 return 0;
1507
1508         if (off == LIMIT_LOW) {
1509                 bps_dft = 0;
1510                 iops_dft = 0;
1511         } else {
1512                 bps_dft = U64_MAX;
1513                 iops_dft = UINT_MAX;
1514         }
1515
1516         if (tg->bps_conf[READ][off] == bps_dft &&
1517             tg->bps_conf[WRITE][off] == bps_dft &&
1518             tg->iops_conf[READ][off] == iops_dft &&
1519             tg->iops_conf[WRITE][off] == iops_dft &&
1520             (off != LIMIT_LOW ||
1521              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1522               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1523                 return 0;
1524
1525         if (tg->bps_conf[READ][off] != U64_MAX)
1526                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1527                         tg->bps_conf[READ][off]);
1528         if (tg->bps_conf[WRITE][off] != U64_MAX)
1529                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1530                         tg->bps_conf[WRITE][off]);
1531         if (tg->iops_conf[READ][off] != UINT_MAX)
1532                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1533                         tg->iops_conf[READ][off]);
1534         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1535                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1536                         tg->iops_conf[WRITE][off]);
1537         if (off == LIMIT_LOW) {
1538                 if (tg->idletime_threshold_conf == ULONG_MAX)
1539                         strcpy(idle_time, " idle=max");
1540                 else
1541                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1542                                 tg->idletime_threshold_conf);
1543
1544                 if (tg->latency_target_conf == ULONG_MAX)
1545                         strcpy(latency_time, " latency=max");
1546                 else
1547                         snprintf(latency_time, sizeof(latency_time),
1548                                 " latency=%lu", tg->latency_target_conf);
1549         }
1550
1551         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1552                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1553                    latency_time);
1554         return 0;
1555 }
1556
1557 static int tg_print_limit(struct seq_file *sf, void *v)
1558 {
1559         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1560                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1561         return 0;
1562 }
1563
1564 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1565                           char *buf, size_t nbytes, loff_t off)
1566 {
1567         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1568         struct blkg_conf_ctx ctx;
1569         struct throtl_grp *tg;
1570         u64 v[4];
1571         unsigned long idle_time;
1572         unsigned long latency_time;
1573         int ret;
1574         int index = of_cft(of)->private;
1575
1576         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1577         if (ret)
1578                 return ret;
1579
1580         tg = blkg_to_tg(ctx.blkg);
1581         tg_update_carryover(tg);
1582
1583         v[0] = tg->bps_conf[READ][index];
1584         v[1] = tg->bps_conf[WRITE][index];
1585         v[2] = tg->iops_conf[READ][index];
1586         v[3] = tg->iops_conf[WRITE][index];
1587
1588         idle_time = tg->idletime_threshold_conf;
1589         latency_time = tg->latency_target_conf;
1590         while (true) {
1591                 char tok[27];   /* wiops=18446744073709551616 */
1592                 char *p;
1593                 u64 val = U64_MAX;
1594                 int len;
1595
1596                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1597                         break;
1598                 if (tok[0] == '\0')
1599                         break;
1600                 ctx.body += len;
1601
1602                 ret = -EINVAL;
1603                 p = tok;
1604                 strsep(&p, "=");
1605                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1606                         goto out_finish;
1607
1608                 ret = -ERANGE;
1609                 if (!val)
1610                         goto out_finish;
1611
1612                 ret = -EINVAL;
1613                 if (!strcmp(tok, "rbps") && val > 1)
1614                         v[0] = val;
1615                 else if (!strcmp(tok, "wbps") && val > 1)
1616                         v[1] = val;
1617                 else if (!strcmp(tok, "riops") && val > 1)
1618                         v[2] = min_t(u64, val, UINT_MAX);
1619                 else if (!strcmp(tok, "wiops") && val > 1)
1620                         v[3] = min_t(u64, val, UINT_MAX);
1621                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1622                         idle_time = val;
1623                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1624                         latency_time = val;
1625                 else
1626                         goto out_finish;
1627         }
1628
1629         tg->bps_conf[READ][index] = v[0];
1630         tg->bps_conf[WRITE][index] = v[1];
1631         tg->iops_conf[READ][index] = v[2];
1632         tg->iops_conf[WRITE][index] = v[3];
1633
1634         if (index == LIMIT_MAX) {
1635                 tg->bps[READ][index] = v[0];
1636                 tg->bps[WRITE][index] = v[1];
1637                 tg->iops[READ][index] = v[2];
1638                 tg->iops[WRITE][index] = v[3];
1639         }
1640         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1641                 tg->bps_conf[READ][LIMIT_MAX]);
1642         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1643                 tg->bps_conf[WRITE][LIMIT_MAX]);
1644         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1645                 tg->iops_conf[READ][LIMIT_MAX]);
1646         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1647                 tg->iops_conf[WRITE][LIMIT_MAX]);
1648         tg->idletime_threshold_conf = idle_time;
1649         tg->latency_target_conf = latency_time;
1650
1651         /* force user to configure all settings for low limit  */
1652         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1653               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1654             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1655             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1656                 tg->bps[READ][LIMIT_LOW] = 0;
1657                 tg->bps[WRITE][LIMIT_LOW] = 0;
1658                 tg->iops[READ][LIMIT_LOW] = 0;
1659                 tg->iops[WRITE][LIMIT_LOW] = 0;
1660                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1661                 tg->latency_target = DFL_LATENCY_TARGET;
1662         } else if (index == LIMIT_LOW) {
1663                 tg->idletime_threshold = tg->idletime_threshold_conf;
1664                 tg->latency_target = tg->latency_target_conf;
1665         }
1666
1667         blk_throtl_update_limit_valid(tg->td);
1668         if (tg->td->limit_valid[LIMIT_LOW]) {
1669                 if (index == LIMIT_LOW)
1670                         tg->td->limit_index = LIMIT_LOW;
1671         } else
1672                 tg->td->limit_index = LIMIT_MAX;
1673         tg_conf_updated(tg, index == LIMIT_LOW &&
1674                 tg->td->limit_valid[LIMIT_LOW]);
1675         ret = 0;
1676 out_finish:
1677         blkg_conf_finish(&ctx);
1678         return ret ?: nbytes;
1679 }
1680
1681 static struct cftype throtl_files[] = {
1682 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1683         {
1684                 .name = "low",
1685                 .flags = CFTYPE_NOT_ON_ROOT,
1686                 .seq_show = tg_print_limit,
1687                 .write = tg_set_limit,
1688                 .private = LIMIT_LOW,
1689         },
1690 #endif
1691         {
1692                 .name = "max",
1693                 .flags = CFTYPE_NOT_ON_ROOT,
1694                 .seq_show = tg_print_limit,
1695                 .write = tg_set_limit,
1696                 .private = LIMIT_MAX,
1697         },
1698         { }     /* terminate */
1699 };
1700
1701 static void throtl_shutdown_wq(struct request_queue *q)
1702 {
1703         struct throtl_data *td = q->td;
1704
1705         cancel_work_sync(&td->dispatch_work);
1706 }
1707
1708 struct blkcg_policy blkcg_policy_throtl = {
1709         .dfl_cftypes            = throtl_files,
1710         .legacy_cftypes         = throtl_legacy_files,
1711
1712         .pd_alloc_fn            = throtl_pd_alloc,
1713         .pd_init_fn             = throtl_pd_init,
1714         .pd_online_fn           = throtl_pd_online,
1715         .pd_offline_fn          = throtl_pd_offline,
1716         .pd_free_fn             = throtl_pd_free,
1717 };
1718
1719 void blk_throtl_cancel_bios(struct gendisk *disk)
1720 {
1721         struct request_queue *q = disk->queue;
1722         struct cgroup_subsys_state *pos_css;
1723         struct blkcg_gq *blkg;
1724
1725         spin_lock_irq(&q->queue_lock);
1726         /*
1727          * queue_lock is held, rcu lock is not needed here technically.
1728          * However, rcu lock is still held to emphasize that following
1729          * path need RCU protection and to prevent warning from lockdep.
1730          */
1731         rcu_read_lock();
1732         blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1733                 struct throtl_grp *tg = blkg_to_tg(blkg);
1734                 struct throtl_service_queue *sq = &tg->service_queue;
1735
1736                 /*
1737                  * Set the flag to make sure throtl_pending_timer_fn() won't
1738                  * stop until all throttled bios are dispatched.
1739                  */
1740                 blkg_to_tg(blkg)->flags |= THROTL_TG_CANCELING;
1741                 /*
1742                  * Update disptime after setting the above flag to make sure
1743                  * throtl_select_dispatch() won't exit without dispatching.
1744                  */
1745                 tg_update_disptime(tg);
1746
1747                 throtl_schedule_pending_timer(sq, jiffies + 1);
1748         }
1749         rcu_read_unlock();
1750         spin_unlock_irq(&q->queue_lock);
1751 }
1752
1753 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1754 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1755 {
1756         unsigned long rtime = jiffies, wtime = jiffies;
1757
1758         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1759                 rtime = tg->last_low_overflow_time[READ];
1760         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1761                 wtime = tg->last_low_overflow_time[WRITE];
1762         return min(rtime, wtime);
1763 }
1764
1765 /* tg should not be an intermediate node */
1766 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1767 {
1768         struct throtl_service_queue *parent_sq;
1769         struct throtl_grp *parent = tg;
1770         unsigned long ret = __tg_last_low_overflow_time(tg);
1771
1772         while (true) {
1773                 parent_sq = parent->service_queue.parent_sq;
1774                 parent = sq_to_tg(parent_sq);
1775                 if (!parent)
1776                         break;
1777
1778                 /*
1779                  * The parent doesn't have low limit, it always reaches low
1780                  * limit. Its overflow time is useless for children
1781                  */
1782                 if (!parent->bps[READ][LIMIT_LOW] &&
1783                     !parent->iops[READ][LIMIT_LOW] &&
1784                     !parent->bps[WRITE][LIMIT_LOW] &&
1785                     !parent->iops[WRITE][LIMIT_LOW])
1786                         continue;
1787                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1788                         ret = __tg_last_low_overflow_time(parent);
1789         }
1790         return ret;
1791 }
1792
1793 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1794 {
1795         /*
1796          * cgroup is idle if:
1797          * - single idle is too long, longer than a fixed value (in case user
1798          *   configure a too big threshold) or 4 times of idletime threshold
1799          * - average think time is more than threshold
1800          * - IO latency is largely below threshold
1801          */
1802         unsigned long time;
1803         bool ret;
1804
1805         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1806         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1807               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1808               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1809               tg->avg_idletime > tg->idletime_threshold ||
1810               (tg->latency_target && tg->bio_cnt &&
1811                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1812         throtl_log(&tg->service_queue,
1813                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1814                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1815                 tg->bio_cnt, ret, tg->td->scale);
1816         return ret;
1817 }
1818
1819 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1820 {
1821         struct throtl_service_queue *sq = &tg->service_queue;
1822         bool read_limit, write_limit;
1823
1824         /*
1825          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1826          * reaches), it's ok to upgrade to next limit
1827          */
1828         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1829         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1830         if (!read_limit && !write_limit)
1831                 return true;
1832         if (read_limit && sq->nr_queued[READ] &&
1833             (!write_limit || sq->nr_queued[WRITE]))
1834                 return true;
1835         if (write_limit && sq->nr_queued[WRITE] &&
1836             (!read_limit || sq->nr_queued[READ]))
1837                 return true;
1838
1839         if (time_after_eq(jiffies,
1840                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1841             throtl_tg_is_idle(tg))
1842                 return true;
1843         return false;
1844 }
1845
1846 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1847 {
1848         while (true) {
1849                 if (throtl_tg_can_upgrade(tg))
1850                         return true;
1851                 tg = sq_to_tg(tg->service_queue.parent_sq);
1852                 if (!tg || !tg_to_blkg(tg)->parent)
1853                         return false;
1854         }
1855         return false;
1856 }
1857
1858 static bool throtl_can_upgrade(struct throtl_data *td,
1859         struct throtl_grp *this_tg)
1860 {
1861         struct cgroup_subsys_state *pos_css;
1862         struct blkcg_gq *blkg;
1863
1864         if (td->limit_index != LIMIT_LOW)
1865                 return false;
1866
1867         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1868                 return false;
1869
1870         rcu_read_lock();
1871         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1872                 struct throtl_grp *tg = blkg_to_tg(blkg);
1873
1874                 if (tg == this_tg)
1875                         continue;
1876                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1877                         continue;
1878                 if (!throtl_hierarchy_can_upgrade(tg)) {
1879                         rcu_read_unlock();
1880                         return false;
1881                 }
1882         }
1883         rcu_read_unlock();
1884         return true;
1885 }
1886
1887 static void throtl_upgrade_check(struct throtl_grp *tg)
1888 {
1889         unsigned long now = jiffies;
1890
1891         if (tg->td->limit_index != LIMIT_LOW)
1892                 return;
1893
1894         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1895                 return;
1896
1897         tg->last_check_time = now;
1898
1899         if (!time_after_eq(now,
1900              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1901                 return;
1902
1903         if (throtl_can_upgrade(tg->td, NULL))
1904                 throtl_upgrade_state(tg->td);
1905 }
1906
1907 static void throtl_upgrade_state(struct throtl_data *td)
1908 {
1909         struct cgroup_subsys_state *pos_css;
1910         struct blkcg_gq *blkg;
1911
1912         throtl_log(&td->service_queue, "upgrade to max");
1913         td->limit_index = LIMIT_MAX;
1914         td->low_upgrade_time = jiffies;
1915         td->scale = 0;
1916         rcu_read_lock();
1917         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1918                 struct throtl_grp *tg = blkg_to_tg(blkg);
1919                 struct throtl_service_queue *sq = &tg->service_queue;
1920
1921                 tg->disptime = jiffies - 1;
1922                 throtl_select_dispatch(sq);
1923                 throtl_schedule_next_dispatch(sq, true);
1924         }
1925         rcu_read_unlock();
1926         throtl_select_dispatch(&td->service_queue);
1927         throtl_schedule_next_dispatch(&td->service_queue, true);
1928         queue_work(kthrotld_workqueue, &td->dispatch_work);
1929 }
1930
1931 static void throtl_downgrade_state(struct throtl_data *td)
1932 {
1933         td->scale /= 2;
1934
1935         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1936         if (td->scale) {
1937                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1938                 return;
1939         }
1940
1941         td->limit_index = LIMIT_LOW;
1942         td->low_downgrade_time = jiffies;
1943 }
1944
1945 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1946 {
1947         struct throtl_data *td = tg->td;
1948         unsigned long now = jiffies;
1949
1950         /*
1951          * If cgroup is below low limit, consider downgrade and throttle other
1952          * cgroups
1953          */
1954         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1955             time_after_eq(now, tg_last_low_overflow_time(tg) +
1956                                         td->throtl_slice) &&
1957             (!throtl_tg_is_idle(tg) ||
1958              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1959                 return true;
1960         return false;
1961 }
1962
1963 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1964 {
1965         while (true) {
1966                 if (!throtl_tg_can_downgrade(tg))
1967                         return false;
1968                 tg = sq_to_tg(tg->service_queue.parent_sq);
1969                 if (!tg || !tg_to_blkg(tg)->parent)
1970                         break;
1971         }
1972         return true;
1973 }
1974
1975 static void throtl_downgrade_check(struct throtl_grp *tg)
1976 {
1977         uint64_t bps;
1978         unsigned int iops;
1979         unsigned long elapsed_time;
1980         unsigned long now = jiffies;
1981
1982         if (tg->td->limit_index != LIMIT_MAX ||
1983             !tg->td->limit_valid[LIMIT_LOW])
1984                 return;
1985         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1986                 return;
1987         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1988                 return;
1989
1990         elapsed_time = now - tg->last_check_time;
1991         tg->last_check_time = now;
1992
1993         if (time_before(now, tg_last_low_overflow_time(tg) +
1994                         tg->td->throtl_slice))
1995                 return;
1996
1997         if (tg->bps[READ][LIMIT_LOW]) {
1998                 bps = tg->last_bytes_disp[READ] * HZ;
1999                 do_div(bps, elapsed_time);
2000                 if (bps >= tg->bps[READ][LIMIT_LOW])
2001                         tg->last_low_overflow_time[READ] = now;
2002         }
2003
2004         if (tg->bps[WRITE][LIMIT_LOW]) {
2005                 bps = tg->last_bytes_disp[WRITE] * HZ;
2006                 do_div(bps, elapsed_time);
2007                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2008                         tg->last_low_overflow_time[WRITE] = now;
2009         }
2010
2011         if (tg->iops[READ][LIMIT_LOW]) {
2012                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2013                 if (iops >= tg->iops[READ][LIMIT_LOW])
2014                         tg->last_low_overflow_time[READ] = now;
2015         }
2016
2017         if (tg->iops[WRITE][LIMIT_LOW]) {
2018                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2019                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2020                         tg->last_low_overflow_time[WRITE] = now;
2021         }
2022
2023         /*
2024          * If cgroup is below low limit, consider downgrade and throttle other
2025          * cgroups
2026          */
2027         if (throtl_hierarchy_can_downgrade(tg))
2028                 throtl_downgrade_state(tg->td);
2029
2030         tg->last_bytes_disp[READ] = 0;
2031         tg->last_bytes_disp[WRITE] = 0;
2032         tg->last_io_disp[READ] = 0;
2033         tg->last_io_disp[WRITE] = 0;
2034 }
2035
2036 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2037 {
2038         unsigned long now;
2039         unsigned long last_finish_time = tg->last_finish_time;
2040
2041         if (last_finish_time == 0)
2042                 return;
2043
2044         now = ktime_get_ns() >> 10;
2045         if (now <= last_finish_time ||
2046             last_finish_time == tg->checked_last_finish_time)
2047                 return;
2048
2049         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2050         tg->checked_last_finish_time = last_finish_time;
2051 }
2052
2053 static void throtl_update_latency_buckets(struct throtl_data *td)
2054 {
2055         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2056         int i, cpu, rw;
2057         unsigned long last_latency[2] = { 0 };
2058         unsigned long latency[2];
2059
2060         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2061                 return;
2062         if (time_before(jiffies, td->last_calculate_time + HZ))
2063                 return;
2064         td->last_calculate_time = jiffies;
2065
2066         memset(avg_latency, 0, sizeof(avg_latency));
2067         for (rw = READ; rw <= WRITE; rw++) {
2068                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2069                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2070
2071                         for_each_possible_cpu(cpu) {
2072                                 struct latency_bucket *bucket;
2073
2074                                 /* this isn't race free, but ok in practice */
2075                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2076                                         cpu);
2077                                 tmp->total_latency += bucket[i].total_latency;
2078                                 tmp->samples += bucket[i].samples;
2079                                 bucket[i].total_latency = 0;
2080                                 bucket[i].samples = 0;
2081                         }
2082
2083                         if (tmp->samples >= 32) {
2084                                 int samples = tmp->samples;
2085
2086                                 latency[rw] = tmp->total_latency;
2087
2088                                 tmp->total_latency = 0;
2089                                 tmp->samples = 0;
2090                                 latency[rw] /= samples;
2091                                 if (latency[rw] == 0)
2092                                         continue;
2093                                 avg_latency[rw][i].latency = latency[rw];
2094                         }
2095                 }
2096         }
2097
2098         for (rw = READ; rw <= WRITE; rw++) {
2099                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2100                         if (!avg_latency[rw][i].latency) {
2101                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2102                                         td->avg_buckets[rw][i].latency =
2103                                                 last_latency[rw];
2104                                 continue;
2105                         }
2106
2107                         if (!td->avg_buckets[rw][i].valid)
2108                                 latency[rw] = avg_latency[rw][i].latency;
2109                         else
2110                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2111                                         avg_latency[rw][i].latency) >> 3;
2112
2113                         td->avg_buckets[rw][i].latency = max(latency[rw],
2114                                 last_latency[rw]);
2115                         td->avg_buckets[rw][i].valid = true;
2116                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2117                 }
2118         }
2119
2120         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2121                 throtl_log(&td->service_queue,
2122                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2123                         "write latency=%ld, write valid=%d", i,
2124                         td->avg_buckets[READ][i].latency,
2125                         td->avg_buckets[READ][i].valid,
2126                         td->avg_buckets[WRITE][i].latency,
2127                         td->avg_buckets[WRITE][i].valid);
2128 }
2129 #else
2130 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2131 {
2132 }
2133
2134 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2135 {
2136 }
2137
2138 static void throtl_downgrade_check(struct throtl_grp *tg)
2139 {
2140 }
2141
2142 static void throtl_upgrade_check(struct throtl_grp *tg)
2143 {
2144 }
2145
2146 static bool throtl_can_upgrade(struct throtl_data *td,
2147         struct throtl_grp *this_tg)
2148 {
2149         return false;
2150 }
2151
2152 static void throtl_upgrade_state(struct throtl_data *td)
2153 {
2154 }
2155 #endif
2156
2157 bool __blk_throtl_bio(struct bio *bio)
2158 {
2159         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2160         struct blkcg_gq *blkg = bio->bi_blkg;
2161         struct throtl_qnode *qn = NULL;
2162         struct throtl_grp *tg = blkg_to_tg(blkg);
2163         struct throtl_service_queue *sq;
2164         bool rw = bio_data_dir(bio);
2165         bool throttled = false;
2166         struct throtl_data *td = tg->td;
2167
2168         rcu_read_lock();
2169
2170         if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2171                 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2172                                 bio->bi_iter.bi_size);
2173                 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2174         }
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(q, &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(q, &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 #endif
2436
2437         td->track_bio_latency = !queue_is_mq(q);
2438         if (!td->track_bio_latency)
2439                 blk_stat_enable_accounting(q);
2440 }
2441
2442 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2443 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2444 {
2445         if (!q->td)
2446                 return -EINVAL;
2447         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2448 }
2449
2450 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2451         const char *page, size_t count)
2452 {
2453         unsigned long v;
2454         unsigned long t;
2455
2456         if (!q->td)
2457                 return -EINVAL;
2458         if (kstrtoul(page, 10, &v))
2459                 return -EINVAL;
2460         t = msecs_to_jiffies(v);
2461         if (t == 0 || t > MAX_THROTL_SLICE)
2462                 return -EINVAL;
2463         q->td->throtl_slice = t;
2464         return count;
2465 }
2466 #endif
2467
2468 static int __init throtl_init(void)
2469 {
2470         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2471         if (!kthrotld_workqueue)
2472                 panic("Failed to create kthrotld\n");
2473
2474         return blkcg_policy_register(&blkcg_policy_throtl);
2475 }
2476
2477 module_init(throtl_init);