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