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-throttle.h"
  16
  17/* Max dispatch from a group in 1 round */
  18#define THROTL_GRP_QUANTUM 8
  19
  20/* Total max dispatch from all groups in one round */
  21#define THROTL_QUANTUM 32
  22
  23/* Throttling is performed over a slice and after that slice is renewed */
  24#define DFL_THROTL_SLICE (HZ / 10)
  25
  26/* A workqueue to queue throttle related work */
  27static struct workqueue_struct *kthrotld_workqueue;
  28
  29#define rb_entry_tg(node)	rb_entry((node), struct throtl_grp, rb_node)
  30
  31struct throtl_data
  32{
  33	/* service tree for active throtl groups */
  34	struct throtl_service_queue service_queue;
  35
  36	struct request_queue *queue;
  37
  38	/* Total Number of queued bios on READ and WRITE lists */
  39	unsigned int nr_queued[2];
  40
  41	/* Work for dispatching throttled bios */
  42	struct work_struct dispatch_work;
  43};
  44
  45static void throtl_pending_timer_fn(struct timer_list *t);
  46
  47static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
  48{
  49	return pd_to_blkg(&tg->pd);
  50}
  51
  52/**
  53 * sq_to_tg - return the throl_grp the specified service queue belongs to
  54 * @sq: the throtl_service_queue of interest
  55 *
  56 * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
  57 * embedded in throtl_data, %NULL is returned.
  58 */
  59static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
  60{
  61	if (sq && sq->parent_sq)
  62		return container_of(sq, struct throtl_grp, service_queue);
  63	else
  64		return NULL;
  65}
  66
  67/**
  68 * sq_to_td - return throtl_data the specified service queue belongs to
  69 * @sq: the throtl_service_queue of interest
  70 *
  71 * A service_queue can be embedded in either a throtl_grp or throtl_data.
  72 * Determine the associated throtl_data accordingly and return it.
  73 */
  74static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
  75{
  76	struct throtl_grp *tg = sq_to_tg(sq);
  77
  78	if (tg)
  79		return tg->td;
  80	else
  81		return container_of(sq, struct throtl_data, service_queue);
  82}
  83
  84static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
  85{
  86	struct blkcg_gq *blkg = tg_to_blkg(tg);
  87
  88	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
  89		return U64_MAX;
  90
  91	return tg->bps[rw];
  92}
  93
  94static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
  95{
  96	struct blkcg_gq *blkg = tg_to_blkg(tg);
  97
  98	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
  99		return UINT_MAX;
 100
 101	return tg->iops[rw];
 102}
 103
 104/**
 105 * throtl_log - log debug message via blktrace
 106 * @sq: the service_queue being reported
 107 * @fmt: printf format string
 108 * @args: printf args
 109 *
 110 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 111 * throtl_grp; otherwise, just "throtl".
 112 */
 113#define throtl_log(sq, fmt, args...)	do {				\
 114	struct throtl_grp *__tg = sq_to_tg((sq));			\
 115	struct throtl_data *__td = sq_to_td((sq));			\
 116									\
 117	(void)__td;							\
 118	if (likely(!blk_trace_note_message_enabled(__td->queue)))	\
 119		break;							\
 120	if ((__tg)) {							\
 121		blk_add_cgroup_trace_msg(__td->queue,			\
 122			&tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\
 123	} else {							\
 124		blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);	\
 125	}								\
 126} while (0)
 127
 128static inline unsigned int throtl_bio_data_size(struct bio *bio)
 129{
 130	/* assume it's one sector */
 131	if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 132		return 512;
 133	return bio->bi_iter.bi_size;
 134}
 135
 136static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 137{
 138	INIT_LIST_HEAD(&qn->node);
 139	bio_list_init(&qn->bios_bps);
 140	bio_list_init(&qn->bios_iops);
 141	qn->tg = tg;
 142}
 143
 144/**
 145 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 146 * @bio: bio being added
 147 * @qn: qnode to add bio to
 148 * @sq: the service_queue @qn belongs to
 149 *
 150 * Add @bio to @qn and put @qn on @sq->queued if it's not already on.
 151 * @qn->tg's reference count is bumped when @qn is activated.  See the
 152 * comment on top of throtl_qnode definition for details.
 153 */
 154static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 155				 struct throtl_service_queue *sq)
 156{
 157	bool rw = bio_data_dir(bio);
 158
 159	/*
 160	 * Split bios have already been throttled by bps, so they are
 161	 * directly queued into the iops path.
 162	 */
 163	if (bio_flagged(bio, BIO_TG_BPS_THROTTLED) ||
 164	    bio_flagged(bio, BIO_BPS_THROTTLED)) {
 165		bio_list_add(&qn->bios_iops, bio);
 166		sq->nr_queued_iops[rw]++;
 167	} else {
 168		bio_list_add(&qn->bios_bps, bio);
 169		sq->nr_queued_bps[rw]++;
 170	}
 171
 172	if (list_empty(&qn->node)) {
 173		list_add_tail(&qn->node, &sq->queued[rw]);
 174		blkg_get(tg_to_blkg(qn->tg));
 175	}
 176}
 177
 178/**
 179 * throtl_peek_queued - peek the first bio on a qnode list
 180 * @queued: the qnode list to peek
 181 *
 182 * Always take a bio from the head of the iops queue first. If the queue is
 183 * empty, we then take it from the bps queue to maintain the overall idea of
 184 * fetching bios from the head.
 185 */
 186static struct bio *throtl_peek_queued(struct list_head *queued)
 187{
 188	struct throtl_qnode *qn;
 189	struct bio *bio;
 190
 191	if (list_empty(queued))
 192		return NULL;
 193
 194	qn = list_first_entry(queued, struct throtl_qnode, node);
 195	bio = bio_list_peek(&qn->bios_iops);
 196	if (!bio)
 197		bio = bio_list_peek(&qn->bios_bps);
 198	WARN_ON_ONCE(!bio);
 199	return bio;
 200}
 201
 202/**
 203 * throtl_pop_queued - pop the first bio form a qnode list
 204 * @sq: the service_queue to pop a bio from
 205 * @tg_to_put: optional out argument for throtl_grp to put
 206 * @rw: read/write
 207 *
 208 * Pop the first bio from the qnode list @sq->queued. Note that we firstly
 209 * focus on the iops list because bios are ultimately dispatched from it.
 210 * After popping, the first qnode is removed from @sq->queued if empty or moved
 211 * to the end of @sq->queued so that the popping order is round-robin.
 212 *
 213 * When the first qnode is removed, its associated throtl_grp should be put
 214 * too.  If @tg_to_put is NULL, this function automatically puts it;
 215 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 216 * responsible for putting it.
 217 */
 218static struct bio *throtl_pop_queued(struct throtl_service_queue *sq,
 219				     struct throtl_grp **tg_to_put, bool rw)
 220{
 221	struct list_head *queued = &sq->queued[rw];
 222	struct throtl_qnode *qn;
 223	struct bio *bio;
 224
 225	if (list_empty(queued))
 226		return NULL;
 227
 228	qn = list_first_entry(queued, struct throtl_qnode, node);
 229	bio = bio_list_pop(&qn->bios_iops);
 230	if (bio) {
 231		sq->nr_queued_iops[rw]--;
 232	} else {
 233		bio = bio_list_pop(&qn->bios_bps);
 234		if (bio)
 235			sq->nr_queued_bps[rw]--;
 236	}
 237	WARN_ON_ONCE(!bio);
 238
 239	if (bio_list_empty(&qn->bios_bps) && bio_list_empty(&qn->bios_iops)) {
 240		list_del_init(&qn->node);
 241		if (tg_to_put)
 242			*tg_to_put = qn->tg;
 243		else
 244			blkg_put(tg_to_blkg(qn->tg));
 245	} else {
 246		list_move_tail(&qn->node, queued);
 247	}
 248
 249	return bio;
 250}
 251
 252/* init a service_queue, assumes the caller zeroed it */
 253static void throtl_service_queue_init(struct throtl_service_queue *sq)
 254{
 255	INIT_LIST_HEAD(&sq->queued[READ]);
 256	INIT_LIST_HEAD(&sq->queued[WRITE]);
 257	sq->pending_tree = RB_ROOT_CACHED;
 258	timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 259}
 260
 261static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk,
 262		struct blkcg *blkcg, gfp_t gfp)
 263{
 264	struct throtl_grp *tg;
 265	int rw;
 266
 267	tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id);
 268	if (!tg)
 269		return NULL;
 270
 271	if (blkg_rwstat_init(&tg->stat_bytes, gfp))
 272		goto err_free_tg;
 273
 274	if (blkg_rwstat_init(&tg->stat_ios, gfp))
 275		goto err_exit_stat_bytes;
 276
 277	throtl_service_queue_init(&tg->service_queue);
 278
 279	for (rw = READ; rw <= WRITE; rw++) {
 280		throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 281		throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 282	}
 283
 284	RB_CLEAR_NODE(&tg->rb_node);
 285	tg->bps[READ] = U64_MAX;
 286	tg->bps[WRITE] = U64_MAX;
 287	tg->iops[READ] = UINT_MAX;
 288	tg->iops[WRITE] = UINT_MAX;
 289
 290	return &tg->pd;
 291
 292err_exit_stat_bytes:
 293	blkg_rwstat_exit(&tg->stat_bytes);
 294err_free_tg:
 295	kfree(tg);
 296	return NULL;
 297}
 298
 299static void throtl_pd_init(struct blkg_policy_data *pd)
 300{
 301	struct throtl_grp *tg = pd_to_tg(pd);
 302	struct blkcg_gq *blkg = tg_to_blkg(tg);
 303	struct throtl_data *td = blkg->q->td;
 304	struct throtl_service_queue *sq = &tg->service_queue;
 305
 306	/*
 307	 * If on the default hierarchy, we switch to properly hierarchical
 308	 * behavior where limits on a given throtl_grp are applied to the
 309	 * whole subtree rather than just the group itself.  e.g. If 16M
 310	 * read_bps limit is set on a parent group, summary bps of
 311	 * parent group and its subtree groups can't exceed 16M for the
 312	 * device.
 313	 *
 314	 * If not on the default hierarchy, the broken flat hierarchy
 315	 * behavior is retained where all throtl_grps are treated as if
 316	 * they're all separate root groups right below throtl_data.
 317	 * Limits of a group don't interact with limits of other groups
 318	 * regardless of the position of the group in the hierarchy.
 319	 */
 320	sq->parent_sq = &td->service_queue;
 321	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 322		sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 323	tg->td = td;
 324}
 325
 326/*
 327 * Set has_rules[] if @tg or any of its parents have limits configured.
 328 * This doesn't require walking up to the top of the hierarchy as the
 329 * parent's has_rules[] is guaranteed to be correct.
 330 */
 331static void tg_update_has_rules(struct throtl_grp *tg)
 332{
 333	struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 334	int rw;
 335
 336	for (rw = READ; rw <= WRITE; rw++) {
 337		tg->has_rules_iops[rw] =
 338			(parent_tg && parent_tg->has_rules_iops[rw]) ||
 339			tg_iops_limit(tg, rw) != UINT_MAX;
 340		tg->has_rules_bps[rw] =
 341			(parent_tg && parent_tg->has_rules_bps[rw]) ||
 342			tg_bps_limit(tg, rw) != U64_MAX;
 343	}
 344}
 345
 346static void throtl_pd_online(struct blkg_policy_data *pd)
 347{
 348	struct throtl_grp *tg = pd_to_tg(pd);
 349	/*
 350	 * We don't want new groups to escape the limits of its ancestors.
 351	 * Update has_rules[] after a new group is brought online.
 352	 */
 353	tg_update_has_rules(tg);
 354}
 355
 356static void throtl_pd_free(struct blkg_policy_data *pd)
 357{
 358	struct throtl_grp *tg = pd_to_tg(pd);
 359
 360	timer_delete_sync(&tg->service_queue.pending_timer);
 361	blkg_rwstat_exit(&tg->stat_bytes);
 362	blkg_rwstat_exit(&tg->stat_ios);
 363	kfree(tg);
 364}
 365
 366static struct throtl_grp *
 367throtl_rb_first(struct throtl_service_queue *parent_sq)
 368{
 369	struct rb_node *n;
 370
 371	n = rb_first_cached(&parent_sq->pending_tree);
 372	WARN_ON_ONCE(!n);
 373	if (!n)
 374		return NULL;
 375	return rb_entry_tg(n);
 376}
 377
 378static void throtl_rb_erase(struct rb_node *n,
 379			    struct throtl_service_queue *parent_sq)
 380{
 381	rb_erase_cached(n, &parent_sq->pending_tree);
 382	RB_CLEAR_NODE(n);
 383}
 384
 385static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 386{
 387	struct throtl_grp *tg;
 388
 389	tg = throtl_rb_first(parent_sq);
 390	if (!tg)
 391		return;
 392
 393	parent_sq->first_pending_disptime = tg->disptime;
 394}
 395
 396static void tg_service_queue_add(struct throtl_grp *tg)
 397{
 398	struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 399	struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 400	struct rb_node *parent = NULL;
 401	struct throtl_grp *__tg;
 402	unsigned long key = tg->disptime;
 403	bool leftmost = true;
 404
 405	while (*node != NULL) {
 406		parent = *node;
 407		__tg = rb_entry_tg(parent);
 408
 409		if (time_before(key, __tg->disptime))
 410			node = &parent->rb_left;
 411		else {
 412			node = &parent->rb_right;
 413			leftmost = false;
 414		}
 415	}
 416
 417	rb_link_node(&tg->rb_node, parent, node);
 418	rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 419			       leftmost);
 420}
 421
 422static void throtl_enqueue_tg(struct throtl_grp *tg)
 423{
 424	if (!(tg->flags & THROTL_TG_PENDING)) {
 425		tg_service_queue_add(tg);
 426		tg->flags |= THROTL_TG_PENDING;
 427		tg->service_queue.parent_sq->nr_pending++;
 428	}
 429}
 430
 431static void throtl_dequeue_tg(struct throtl_grp *tg)
 432{
 433	if (tg->flags & THROTL_TG_PENDING) {
 434		struct throtl_service_queue *parent_sq =
 435			tg->service_queue.parent_sq;
 436
 437		throtl_rb_erase(&tg->rb_node, parent_sq);
 438		--parent_sq->nr_pending;
 439		tg->flags &= ~THROTL_TG_PENDING;
 440	}
 441}
 442
 443/* Call with queue lock held */
 444static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 445					  unsigned long expires)
 446{
 447	unsigned long max_expire = jiffies + 8 * DFL_THROTL_SLICE;
 448
 449	/*
 450	 * Since we are adjusting the throttle limit dynamically, the sleep
 451	 * time calculated according to previous limit might be invalid. It's
 452	 * possible the cgroup sleep time is very long and no other cgroups
 453	 * have IO running so notify the limit changes. Make sure the cgroup
 454	 * doesn't sleep too long to avoid the missed notification.
 455	 */
 456	if (time_after(expires, max_expire))
 457		expires = max_expire;
 458	mod_timer(&sq->pending_timer, expires);
 459	throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 460		   expires - jiffies, jiffies);
 461}
 462
 463/**
 464 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 465 * @sq: the service_queue to schedule dispatch for
 466 * @force: force scheduling
 467 *
 468 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 469 * dispatch time of the first pending child.  Returns %true if either timer
 470 * is armed or there's no pending child left.  %false if the current
 471 * dispatch window is still open and the caller should continue
 472 * dispatching.
 473 *
 474 * If @force is %true, the dispatch timer is always scheduled and this
 475 * function is guaranteed to return %true.  This is to be used when the
 476 * caller can't dispatch itself and needs to invoke pending_timer
 477 * unconditionally.  Note that forced scheduling is likely to induce short
 478 * delay before dispatch starts even if @sq->first_pending_disptime is not
 479 * in the future and thus shouldn't be used in hot paths.
 480 */
 481static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 482					  bool force)
 483{
 484	/* any pending children left? */
 485	if (!sq->nr_pending)
 486		return true;
 487
 488	update_min_dispatch_time(sq);
 489
 490	/* is the next dispatch time in the future? */
 491	if (force || time_after(sq->first_pending_disptime, jiffies)) {
 492		throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 493		return true;
 494	}
 495
 496	/* tell the caller to continue dispatching */
 497	return false;
 498}
 499
 500static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 501		bool rw, unsigned long start)
 502{
 503	tg->bytes_disp[rw] = 0;
 504	tg->io_disp[rw] = 0;
 505
 506	/*
 507	 * Previous slice has expired. We must have trimmed it after last
 508	 * bio dispatch. That means since start of last slice, we never used
 509	 * that bandwidth. Do try to make use of that bandwidth while giving
 510	 * credit.
 511	 */
 512	if (time_after(start, tg->slice_start[rw]))
 513		tg->slice_start[rw] = start;
 514
 515	tg->slice_end[rw] = jiffies + DFL_THROTL_SLICE;
 516	throtl_log(&tg->service_queue,
 517		   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 518		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 519		   tg->slice_end[rw], jiffies);
 520}
 521
 522static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw,
 523					  bool clear)
 524{
 525	if (clear) {
 526		tg->bytes_disp[rw] = 0;
 527		tg->io_disp[rw] = 0;
 528	}
 529	tg->slice_start[rw] = jiffies;
 530	tg->slice_end[rw] = jiffies + DFL_THROTL_SLICE;
 531
 532	throtl_log(&tg->service_queue,
 533		   "[%c] new slice start=%lu end=%lu jiffies=%lu",
 534		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 535		   tg->slice_end[rw], jiffies);
 536}
 537
 538static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 539					unsigned long jiffy_end)
 540{
 541	tg->slice_end[rw] = roundup(jiffy_end, DFL_THROTL_SLICE);
 542}
 543
 544static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 545				       unsigned long jiffy_end)
 546{
 547	if (!time_before(tg->slice_end[rw], jiffy_end))
 548		return;
 549
 550	throtl_set_slice_end(tg, rw, jiffy_end);
 551	throtl_log(&tg->service_queue,
 552		   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 553		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 554		   tg->slice_end[rw], jiffies);
 555}
 556
 557/* Determine if previously allocated or extended slice is complete or not */
 558static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 559{
 560	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 561		return false;
 562
 563	return true;
 564}
 565
 566static unsigned int sq_queued(struct throtl_service_queue *sq, int type)
 567{
 568	return sq->nr_queued_bps[type] + sq->nr_queued_iops[type];
 569}
 570
 571static unsigned int calculate_io_allowed(u32 iops_limit,
 572					 unsigned long jiffy_elapsed)
 573{
 574	unsigned int io_allowed;
 575	u64 tmp;
 576
 577	/*
 578	 * jiffy_elapsed should not be a big value as minimum iops can be
 579	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 580	 * will allow dispatch after 1 second and after that slice should
 581	 * have been trimmed.
 582	 */
 583
 584	tmp = (u64)iops_limit * jiffy_elapsed;
 585	do_div(tmp, HZ);
 586
 587	if (tmp > UINT_MAX)
 588		io_allowed = UINT_MAX;
 589	else
 590		io_allowed = tmp;
 591
 592	return io_allowed;
 593}
 594
 595static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed)
 596{
 597	/*
 598	 * Can result be wider than 64 bits?
 599	 * We check against 62, not 64, due to ilog2 truncation.
 600	 */
 601	if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62)
 602		return U64_MAX;
 603	return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ);
 604}
 605
 606static long long throtl_trim_bps(struct throtl_grp *tg, bool rw,
 607				 unsigned long time_elapsed)
 608{
 609	u64 bps_limit = tg_bps_limit(tg, rw);
 610	long long bytes_trim;
 611
 612	if (bps_limit == U64_MAX)
 613		return 0;
 614
 615	/* Need to consider the case of bytes_allowed overflow. */
 616	bytes_trim = calculate_bytes_allowed(bps_limit, time_elapsed);
 617	if (bytes_trim <= 0 || tg->bytes_disp[rw] < bytes_trim) {
 618		bytes_trim = tg->bytes_disp[rw];
 619		tg->bytes_disp[rw] = 0;
 620	} else {
 621		tg->bytes_disp[rw] -= bytes_trim;
 622	}
 623
 624	return bytes_trim;
 625}
 626
 627static int throtl_trim_iops(struct throtl_grp *tg, bool rw,
 628			    unsigned long time_elapsed)
 629{
 630	u32 iops_limit = tg_iops_limit(tg, rw);
 631	int io_trim;
 632
 633	if (iops_limit == UINT_MAX)
 634		return 0;
 635
 636	/* Need to consider the case of io_allowed overflow. */
 637	io_trim = calculate_io_allowed(iops_limit, time_elapsed);
 638	if (io_trim <= 0 || tg->io_disp[rw] < io_trim) {
 639		io_trim = tg->io_disp[rw];
 640		tg->io_disp[rw] = 0;
 641	} else {
 642		tg->io_disp[rw] -= io_trim;
 643	}
 644
 645	return io_trim;
 646}
 647
 648/* Trim the used slices and adjust slice start accordingly */
 649static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 650{
 651	unsigned long time_elapsed;
 652	long long bytes_trim;
 653	int io_trim;
 654
 655	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 656
 657	/*
 658	 * If bps are unlimited (-1), then time slice don't get
 659	 * renewed. Don't try to trim the slice if slice is used. A new
 660	 * slice will start when appropriate.
 661	 */
 662	if (throtl_slice_used(tg, rw))
 663		return;
 664
 665	/*
 666	 * A bio has been dispatched. Also adjust slice_end. It might happen
 667	 * that initially cgroup limit was very low resulting in high
 668	 * slice_end, but later limit was bumped up and bio was dispatched
 669	 * sooner, then we need to reduce slice_end. A high bogus slice_end
 670	 * is bad because it does not allow new slice to start.
 671	 */
 672	throtl_set_slice_end(tg, rw, jiffies + DFL_THROTL_SLICE);
 673
 674	time_elapsed = rounddown(jiffies - tg->slice_start[rw],
 675				 DFL_THROTL_SLICE);
 676	/* Don't trim slice until at least 2 slices are used */
 677	if (time_elapsed < DFL_THROTL_SLICE * 2)
 678		return;
 679
 680	/*
 681	 * The bio submission time may be a few jiffies more than the expected
 682	 * waiting time, due to 'extra_bytes' can't be divided in
 683	 * tg_within_bps_limit(), and also due to timer wakeup delay. In this
 684	 * case, adjust slice_start will discard the extra wait time, causing
 685	 * lower rate than expected. Therefore, other than the above rounddown,
 686	 * one extra slice is preserved for deviation.
 687	 */
 688	time_elapsed -= DFL_THROTL_SLICE;
 689	bytes_trim = throtl_trim_bps(tg, rw, time_elapsed);
 690	io_trim = throtl_trim_iops(tg, rw, time_elapsed);
 691	if (!bytes_trim && !io_trim)
 692		return;
 693
 694	tg->slice_start[rw] += time_elapsed;
 695
 696	throtl_log(&tg->service_queue,
 697		   "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu",
 698		   rw == READ ? 'R' : 'W', time_elapsed / DFL_THROTL_SLICE,
 699		   bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw],
 700		   jiffies);
 701}
 702
 703static void __tg_update_carryover(struct throtl_grp *tg, bool rw,
 704				  long long *bytes, int *ios)
 705{
 706	unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw];
 707	u64 bps_limit = tg_bps_limit(tg, rw);
 708	u32 iops_limit = tg_iops_limit(tg, rw);
 709	long long bytes_allowed;
 710	int io_allowed;
 711
 712	/*
 713	 * If the queue is empty, carryover handling is not needed. In such cases,
 714	 * tg->[bytes/io]_disp should be reset to 0 to avoid impacting the dispatch
 715	 * of subsequent bios. The same handling applies when the previous BPS/IOPS
 716	 * limit was set to max.
 717	 */
 718	if (sq_queued(&tg->service_queue, rw) == 0) {
 719		tg->bytes_disp[rw] = 0;
 720		tg->io_disp[rw] = 0;
 721		return;
 722	}
 723
 724	/*
 725	 * If config is updated while bios are still throttled, calculate and
 726	 * accumulate how many bytes/ios are waited across changes. And use the
 727	 * calculated carryover (@bytes/@ios) to update [bytes/io]_disp, which
 728	 * will be used to calculate new wait time under new configuration.
 729	 * And we need to consider the case of bytes/io_allowed overflow.
 730	 */
 731	if (bps_limit != U64_MAX) {
 732		bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed);
 733		if (bytes_allowed > 0)
 734			*bytes = bytes_allowed - tg->bytes_disp[rw];
 735	}
 736	if (iops_limit != UINT_MAX) {
 737		io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed);
 738		if (io_allowed > 0)
 739			*ios = io_allowed - tg->io_disp[rw];
 740	}
 741
 742	tg->bytes_disp[rw] = -*bytes;
 743	tg->io_disp[rw] = -*ios;
 744}
 745
 746static void tg_update_carryover(struct throtl_grp *tg)
 747{
 748	long long bytes[2] = {0};
 749	int ios[2] = {0};
 750
 751	__tg_update_carryover(tg, READ, &bytes[READ], &ios[READ]);
 752	__tg_update_carryover(tg, WRITE, &bytes[WRITE], &ios[WRITE]);
 753
 754	/* see comments in struct throtl_grp for meaning of carryover. */
 755	throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__,
 756		   bytes[READ], bytes[WRITE], ios[READ], ios[WRITE]);
 757}
 758
 759static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio,
 760				 u32 iops_limit)
 761{
 762	bool rw = bio_data_dir(bio);
 763	int io_allowed;
 764	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 765
 766	jiffy_elapsed = jiffies - tg->slice_start[rw];
 767
 768	/* Round up to the next throttle slice, wait time must be nonzero */
 769	jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, DFL_THROTL_SLICE);
 770	io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd);
 771	if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed)
 772		return 0;
 773
 774	/* Calc approx time to dispatch */
 775	jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 776
 777	/* make sure at least one io can be dispatched after waiting */
 778	jiffy_wait = max(jiffy_wait, HZ / iops_limit + 1);
 779	return jiffy_wait;
 780}
 781
 782static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio,
 783				u64 bps_limit)
 784{
 785	bool rw = bio_data_dir(bio);
 786	long long bytes_allowed;
 787	u64 extra_bytes;
 788	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 789	unsigned int bio_size = throtl_bio_data_size(bio);
 790
 791	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 792
 793	/* Slice has just started. Consider one slice interval */
 794	if (!jiffy_elapsed)
 795		jiffy_elapsed_rnd = DFL_THROTL_SLICE;
 796
 797	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, DFL_THROTL_SLICE);
 798	bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd);
 799	/* Need to consider the case of bytes_allowed overflow. */
 800	if ((bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed)
 801	    || bytes_allowed < 0)
 802		return 0;
 803
 804	/* Calc approx time to dispatch */
 805	extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 806	jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 807
 808	if (!jiffy_wait)
 809		jiffy_wait = 1;
 810
 811	/*
 812	 * This wait time is without taking into consideration the rounding
 813	 * up we did. Add that time also.
 814	 */
 815	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 816	return jiffy_wait;
 817}
 818
 819static void throtl_charge_bps_bio(struct throtl_grp *tg, struct bio *bio)
 820{
 821	unsigned int bio_size = throtl_bio_data_size(bio);
 822
 823	/* Charge the bio to the group */
 824	if (!bio_flagged(bio, BIO_BPS_THROTTLED) &&
 825	    !bio_flagged(bio, BIO_TG_BPS_THROTTLED)) {
 826		bio_set_flag(bio, BIO_TG_BPS_THROTTLED);
 827		tg->bytes_disp[bio_data_dir(bio)] += bio_size;
 828	}
 829}
 830
 831static void throtl_charge_iops_bio(struct throtl_grp *tg, struct bio *bio)
 832{
 833	bio_clear_flag(bio, BIO_TG_BPS_THROTTLED);
 834	tg->io_disp[bio_data_dir(bio)]++;
 835}
 836
 837/*
 838 * If previous slice expired, start a new one otherwise renew/extend existing
 839 * slice to make sure it is at least throtl_slice interval long since now. New
 840 * slice is started only for empty throttle group. If there is queued bio, that
 841 * means there should be an active slice and it should be extended instead.
 842 */
 843static void tg_update_slice(struct throtl_grp *tg, bool rw)
 844{
 845	if (throtl_slice_used(tg, rw) &&
 846	    sq_queued(&tg->service_queue, rw) == 0)
 847		throtl_start_new_slice(tg, rw, true);
 848	else
 849		throtl_extend_slice(tg, rw, jiffies + DFL_THROTL_SLICE);
 850}
 851
 852static unsigned long tg_dispatch_bps_time(struct throtl_grp *tg, struct bio *bio)
 853{
 854	bool rw = bio_data_dir(bio);
 855	u64 bps_limit = tg_bps_limit(tg, rw);
 856	unsigned long bps_wait;
 857
 858	/* no need to throttle if this bio's bytes have been accounted */
 859	if (bps_limit == U64_MAX || tg->flags & THROTL_TG_CANCELING ||
 860	    bio_flagged(bio, BIO_BPS_THROTTLED) ||
 861	    bio_flagged(bio, BIO_TG_BPS_THROTTLED))
 862		return 0;
 863
 864	tg_update_slice(tg, rw);
 865	bps_wait = tg_within_bps_limit(tg, bio, bps_limit);
 866	throtl_extend_slice(tg, rw, jiffies + bps_wait);
 867
 868	return bps_wait;
 869}
 870
 871static unsigned long tg_dispatch_iops_time(struct throtl_grp *tg, struct bio *bio)
 872{
 873	bool rw = bio_data_dir(bio);
 874	u32 iops_limit = tg_iops_limit(tg, rw);
 875	unsigned long iops_wait;
 876
 877	if (iops_limit == UINT_MAX || tg->flags & THROTL_TG_CANCELING)
 878		return 0;
 879
 880	tg_update_slice(tg, rw);
 881	iops_wait = tg_within_iops_limit(tg, bio, iops_limit);
 882	throtl_extend_slice(tg, rw, jiffies + iops_wait);
 883
 884	return iops_wait;
 885}
 886
 887/*
 888 * Returns approx number of jiffies to wait before this bio is with-in IO rate
 889 * and can be moved to other queue or dispatched.
 890 */
 891static unsigned long tg_dispatch_time(struct throtl_grp *tg, struct bio *bio)
 892{
 893	bool rw = bio_data_dir(bio);
 894	unsigned long wait;
 895
 896	/*
 897 	 * Currently whole state machine of group depends on first bio
 898	 * queued in the group bio list. So one should not be calling
 899	 * this function with a different bio if there are other bios
 900	 * queued.
 901	 */
 902	BUG_ON(sq_queued(&tg->service_queue, rw) &&
 903	       bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 904
 905	wait = tg_dispatch_bps_time(tg, bio);
 906	if (wait != 0)
 907		return wait;
 908
 909	/*
 910	 * Charge bps here because @bio will be directly placed into the
 911	 * iops queue afterward.
 912	 */
 913	throtl_charge_bps_bio(tg, bio);
 914
 915	return tg_dispatch_iops_time(tg, bio);
 916}
 917
 918/**
 919 * throtl_add_bio_tg - add a bio to the specified throtl_grp
 920 * @bio: bio to add
 921 * @qn: qnode to use
 922 * @tg: the target throtl_grp
 923 *
 924 * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 925 * tg->qnode_on_self[] is used.
 926 */
 927static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 928			      struct throtl_grp *tg)
 929{
 930	struct throtl_service_queue *sq = &tg->service_queue;
 931	bool rw = bio_data_dir(bio);
 932
 933	if (!qn)
 934		qn = &tg->qnode_on_self[rw];
 935
 936	/*
 937	 * If @tg doesn't currently have any bios queued in the same
 938	 * direction, queueing @bio can change when @tg should be
 939	 * dispatched.  Mark that @tg was empty.  This is automatically
 940	 * cleared on the next tg_update_disptime().
 941	 */
 942	if (sq_queued(sq, rw) == 0)
 943		tg->flags |= THROTL_TG_WAS_EMPTY;
 944
 945	throtl_qnode_add_bio(bio, qn, sq);
 946
 947	/*
 948	 * Since we have split the queues, when the iops queue is
 949	 * previously empty and a new @bio is added into the first @qn,
 950	 * we also need to update the @tg->disptime.
 951	 */
 952	if (bio_flagged(bio, BIO_BPS_THROTTLED) &&
 953	    bio == throtl_peek_queued(&sq->queued[rw]))
 954		tg->flags |= THROTL_TG_IOPS_WAS_EMPTY;
 955
 956	throtl_enqueue_tg(tg);
 957}
 958
 959static void tg_update_disptime(struct throtl_grp *tg)
 960{
 961	struct throtl_service_queue *sq = &tg->service_queue;
 962	unsigned long read_wait = -1, write_wait = -1, min_wait, disptime;
 963	struct bio *bio;
 964
 965	bio = throtl_peek_queued(&sq->queued[READ]);
 966	if (bio)
 967		read_wait = tg_dispatch_time(tg, bio);
 968
 969	bio = throtl_peek_queued(&sq->queued[WRITE]);
 970	if (bio)
 971		write_wait = tg_dispatch_time(tg, bio);
 972
 973	min_wait = min(read_wait, write_wait);
 974	disptime = jiffies + min_wait;
 975
 976	/* Update dispatch time */
 977	throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 978	tg->disptime = disptime;
 979	tg_service_queue_add(tg);
 980
 981	/* see throtl_add_bio_tg() */
 982	tg->flags &= ~THROTL_TG_WAS_EMPTY;
 983	tg->flags &= ~THROTL_TG_IOPS_WAS_EMPTY;
 984}
 985
 986static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
 987					struct throtl_grp *parent_tg, bool rw)
 988{
 989	if (throtl_slice_used(parent_tg, rw)) {
 990		throtl_start_new_slice_with_credit(parent_tg, rw,
 991				child_tg->slice_start[rw]);
 992	}
 993
 994}
 995
 996static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
 997{
 998	struct throtl_service_queue *sq = &tg->service_queue;
 999	struct throtl_service_queue *parent_sq = sq->parent_sq;
1000	struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1001	struct throtl_grp *tg_to_put = NULL;
1002	struct bio *bio;
1003
1004	/*
1005	 * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1006	 * from @tg may put its reference and @parent_sq might end up
1007	 * getting released prematurely.  Remember the tg to put and put it
1008	 * after @bio is transferred to @parent_sq.
1009	 */
1010	bio = throtl_pop_queued(sq, &tg_to_put, rw);
1011
1012	throtl_charge_iops_bio(tg, bio);
1013
1014	/*
1015	 * If our parent is another tg, we just need to transfer @bio to
1016	 * the parent using throtl_add_bio_tg().  If our parent is
1017	 * @td->service_queue, @bio is ready to be issued.  Put it on its
1018	 * bio_lists[] and decrease total number queued.  The caller is
1019	 * responsible for issuing these bios.
1020	 */
1021	if (parent_tg) {
1022		throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1023		start_parent_slice_with_credit(tg, parent_tg, rw);
1024	} else {
1025		bio_set_flag(bio, BIO_BPS_THROTTLED);
1026		throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1027				     parent_sq);
1028		BUG_ON(tg->td->nr_queued[rw] <= 0);
1029		tg->td->nr_queued[rw]--;
1030	}
1031
1032	throtl_trim_slice(tg, rw);
1033
1034	if (tg_to_put)
1035		blkg_put(tg_to_blkg(tg_to_put));
1036}
1037
1038static int throtl_dispatch_tg(struct throtl_grp *tg)
1039{
1040	struct throtl_service_queue *sq = &tg->service_queue;
1041	unsigned int nr_reads = 0, nr_writes = 0;
1042	unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1043	unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1044	struct bio *bio;
1045
1046	/* Try to dispatch 75% READS and 25% WRITES */
1047
1048	while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1049	       tg_dispatch_time(tg, bio) == 0) {
1050
1051		tg_dispatch_one_bio(tg, READ);
1052		nr_reads++;
1053
1054		if (nr_reads >= max_nr_reads)
1055			break;
1056	}
1057
1058	while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1059	       tg_dispatch_time(tg, bio) == 0) {
1060
1061		tg_dispatch_one_bio(tg, WRITE);
1062		nr_writes++;
1063
1064		if (nr_writes >= max_nr_writes)
1065			break;
1066	}
1067
1068	return nr_reads + nr_writes;
1069}
1070
1071static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1072{
1073	unsigned int nr_disp = 0;
1074
1075	while (1) {
1076		struct throtl_grp *tg;
1077		struct throtl_service_queue *sq;
1078
1079		if (!parent_sq->nr_pending)
1080			break;
1081
1082		tg = throtl_rb_first(parent_sq);
1083		if (!tg)
1084			break;
1085
1086		if (time_before(jiffies, tg->disptime))
1087			break;
1088
1089		nr_disp += throtl_dispatch_tg(tg);
1090
1091		sq = &tg->service_queue;
1092		if (sq_queued(sq, READ) || sq_queued(sq, WRITE))
1093			tg_update_disptime(tg);
1094		else
1095			throtl_dequeue_tg(tg);
1096
1097		if (nr_disp >= THROTL_QUANTUM)
1098			break;
1099	}
1100
1101	return nr_disp;
1102}
1103
1104/**
1105 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1106 * @t: the pending_timer member of the throtl_service_queue being serviced
1107 *
1108 * This timer is armed when a child throtl_grp with active bio's become
1109 * pending and queued on the service_queue's pending_tree and expires when
1110 * the first child throtl_grp should be dispatched.  This function
1111 * dispatches bio's from the children throtl_grps to the parent
1112 * service_queue.
1113 *
1114 * If the parent's parent is another throtl_grp, dispatching is propagated
1115 * by either arming its pending_timer or repeating dispatch directly.  If
1116 * the top-level service_tree is reached, throtl_data->dispatch_work is
1117 * kicked so that the ready bio's are issued.
1118 */
1119static void throtl_pending_timer_fn(struct timer_list *t)
1120{
1121	struct throtl_service_queue *sq = timer_container_of(sq, t,
1122							     pending_timer);
1123	struct throtl_grp *tg = sq_to_tg(sq);
1124	struct throtl_data *td = sq_to_td(sq);
1125	struct throtl_service_queue *parent_sq;
1126	struct request_queue *q;
1127	bool dispatched;
1128	int ret;
1129
1130	/* throtl_data may be gone, so figure out request queue by blkg */
1131	if (tg)
1132		q = tg->pd.blkg->q;
1133	else
1134		q = td->queue;
1135
1136	spin_lock_irq(&q->queue_lock);
1137
1138	if (!q->root_blkg)
1139		goto out_unlock;
1140
1141again:
1142	parent_sq = sq->parent_sq;
1143	dispatched = false;
1144
1145	while (true) {
1146		unsigned int __maybe_unused bio_cnt_r = sq_queued(sq, READ);
1147		unsigned int __maybe_unused bio_cnt_w = sq_queued(sq, WRITE);
1148
1149		throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1150			   bio_cnt_r + bio_cnt_w, bio_cnt_r, bio_cnt_w);
1151
1152		ret = throtl_select_dispatch(sq);
1153		if (ret) {
1154			throtl_log(sq, "bios disp=%u", ret);
1155			dispatched = true;
1156		}
1157
1158		if (throtl_schedule_next_dispatch(sq, false))
1159			break;
1160
1161		/* this dispatch windows is still open, relax and repeat */
1162		spin_unlock_irq(&q->queue_lock);
1163		cpu_relax();
1164		spin_lock_irq(&q->queue_lock);
1165	}
1166
1167	if (!dispatched)
1168		goto out_unlock;
1169
1170	if (parent_sq) {
1171		/* @parent_sq is another throl_grp, propagate dispatch */
1172		if (tg->flags & THROTL_TG_WAS_EMPTY ||
1173		    tg->flags & THROTL_TG_IOPS_WAS_EMPTY) {
1174			tg_update_disptime(tg);
1175			if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1176				/* window is already open, repeat dispatching */
1177				sq = parent_sq;
1178				tg = sq_to_tg(sq);
1179				goto again;
1180			}
1181		}
1182	} else {
1183		/* reached the top-level, queue issuing */
1184		queue_work(kthrotld_workqueue, &td->dispatch_work);
1185	}
1186out_unlock:
1187	spin_unlock_irq(&q->queue_lock);
1188}
1189
1190/**
1191 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1192 * @work: work item being executed
1193 *
1194 * This function is queued for execution when bios reach the bio_lists[]
1195 * of throtl_data->service_queue.  Those bios are ready and issued by this
1196 * function.
1197 */
1198static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1199{
1200	struct throtl_data *td = container_of(work, struct throtl_data,
1201					      dispatch_work);
1202	struct throtl_service_queue *td_sq = &td->service_queue;
1203	struct request_queue *q = td->queue;
1204	struct bio_list bio_list_on_stack;
1205	struct bio *bio;
1206	struct blk_plug plug;
1207	int rw;
1208
1209	bio_list_init(&bio_list_on_stack);
1210
1211	spin_lock_irq(&q->queue_lock);
1212	for (rw = READ; rw <= WRITE; rw++)
1213		while ((bio = throtl_pop_queued(td_sq, NULL, rw)))
1214			bio_list_add(&bio_list_on_stack, bio);
1215	spin_unlock_irq(&q->queue_lock);
1216
1217	if (!bio_list_empty(&bio_list_on_stack)) {
1218		blk_start_plug(&plug);
1219		while ((bio = bio_list_pop(&bio_list_on_stack)))
1220			submit_bio_noacct_nocheck(bio, false);
1221		blk_finish_plug(&plug);
1222	}
1223}
1224
1225static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1226			      int off)
1227{
1228	struct throtl_grp *tg = pd_to_tg(pd);
1229	u64 v = *(u64 *)((void *)tg + off);
1230
1231	if (v == U64_MAX)
1232		return 0;
1233	return __blkg_prfill_u64(sf, pd, v);
1234}
1235
1236static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1237			       int off)
1238{
1239	struct throtl_grp *tg = pd_to_tg(pd);
1240	unsigned int v = *(unsigned int *)((void *)tg + off);
1241
1242	if (v == UINT_MAX)
1243		return 0;
1244	return __blkg_prfill_u64(sf, pd, v);
1245}
1246
1247static int tg_print_conf_u64(struct seq_file *sf, void *v)
1248{
1249	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1250			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1251	return 0;
1252}
1253
1254static int tg_print_conf_uint(struct seq_file *sf, void *v)
1255{
1256	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1257			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1258	return 0;
1259}
1260
1261static void tg_conf_updated(struct throtl_grp *tg, bool global)
1262{
1263	struct throtl_service_queue *sq = &tg->service_queue;
1264	struct cgroup_subsys_state *pos_css;
1265	struct blkcg_gq *blkg;
1266
1267	throtl_log(&tg->service_queue,
1268		   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1269		   tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1270		   tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1271
1272	rcu_read_lock();
1273	/*
1274	 * Update has_rules[] flags for the updated tg's subtree.  A tg is
1275	 * considered to have rules if either the tg itself or any of its
1276	 * ancestors has rules.  This identifies groups without any
1277	 * restrictions in the whole hierarchy and allows them to bypass
1278	 * blk-throttle.
1279	 */
1280	blkg_for_each_descendant_pre(blkg, pos_css,
1281			global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1282		struct throtl_grp *this_tg = blkg_to_tg(blkg);
1283
1284		tg_update_has_rules(this_tg);
1285		/* ignore root/second level */
1286		if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1287		    !blkg->parent->parent)
1288			continue;
1289	}
1290	rcu_read_unlock();
1291
1292	/*
1293	 * We're already holding queue_lock and know @tg is valid.  Let's
1294	 * apply the new config directly.
1295	 *
1296	 * Restart the slices for both READ and WRITES. It might happen
1297	 * that a group's limit are dropped suddenly and we don't want to
1298	 * account recently dispatched IO with new low rate.
1299	 */
1300	throtl_start_new_slice(tg, READ, false);
1301	throtl_start_new_slice(tg, WRITE, false);
1302
1303	if (tg->flags & THROTL_TG_PENDING) {
1304		tg_update_disptime(tg);
1305		throtl_schedule_next_dispatch(sq->parent_sq, true);
1306	}
1307}
1308
1309static int blk_throtl_init(struct gendisk *disk)
1310{
1311	struct request_queue *q = disk->queue;
1312	struct throtl_data *td;
1313	unsigned int memflags;
1314	int ret;
1315
1316	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1317	if (!td)
1318		return -ENOMEM;
1319
1320	INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1321	throtl_service_queue_init(&td->service_queue);
1322
1323	memflags = blk_mq_freeze_queue(disk->queue);
1324	blk_mq_quiesce_queue(disk->queue);
1325
1326	q->td = td;
1327	td->queue = q;
1328
1329	/* activate policy, blk_throtl_activated() will return true */
1330	ret = blkcg_activate_policy(disk, &blkcg_policy_throtl);
1331	if (ret) {
1332		q->td = NULL;
1333		kfree(td);
1334	}
1335
1336	blk_mq_unquiesce_queue(disk->queue);
1337	blk_mq_unfreeze_queue(disk->queue, memflags);
1338
1339	return ret;
1340}
1341
1342
1343static ssize_t tg_set_conf(struct kernfs_open_file *of,
1344			   char *buf, size_t nbytes, loff_t off, bool is_u64)
1345{
1346	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1347	struct blkg_conf_ctx ctx;
1348	struct throtl_grp *tg;
1349	int ret;
1350	u64 v;
1351
1352	blkg_conf_init(&ctx, buf);
1353
1354	ret = blkg_conf_open_bdev(&ctx);
1355	if (ret)
1356		goto out_finish;
1357
1358	if (!blk_throtl_activated(ctx.bdev->bd_queue)) {
1359		ret = blk_throtl_init(ctx.bdev->bd_disk);
1360		if (ret)
1361			goto out_finish;
1362	}
1363
1364	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1365	if (ret)
1366		goto out_finish;
1367
1368	ret = -EINVAL;
1369	if (sscanf(ctx.body, "%llu", &v) != 1)
1370		goto out_finish;
1371	if (!v)
1372		v = U64_MAX;
1373
1374	tg = blkg_to_tg(ctx.blkg);
1375	tg_update_carryover(tg);
1376
1377	if (is_u64)
1378		*(u64 *)((void *)tg + of_cft(of)->private) = v;
1379	else
1380		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1381
1382	tg_conf_updated(tg, false);
1383	ret = 0;
1384out_finish:
1385	blkg_conf_exit(&ctx);
1386	return ret ?: nbytes;
1387}
1388
1389static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1390			       char *buf, size_t nbytes, loff_t off)
1391{
1392	return tg_set_conf(of, buf, nbytes, off, true);
1393}
1394
1395static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1396				char *buf, size_t nbytes, loff_t off)
1397{
1398	return tg_set_conf(of, buf, nbytes, off, false);
1399}
1400
1401static int tg_print_rwstat(struct seq_file *sf, void *v)
1402{
1403	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1404			  blkg_prfill_rwstat, &blkcg_policy_throtl,
1405			  seq_cft(sf)->private, true);
1406	return 0;
1407}
1408
1409static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1410				      struct blkg_policy_data *pd, int off)
1411{
1412	struct blkg_rwstat_sample sum;
1413
1414	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1415				  &sum);
1416	return __blkg_prfill_rwstat(sf, pd, &sum);
1417}
1418
1419static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1420{
1421	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1422			  tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1423			  seq_cft(sf)->private, true);
1424	return 0;
1425}
1426
1427static struct cftype throtl_legacy_files[] = {
1428	{
1429		.name = "throttle.read_bps_device",
1430		.private = offsetof(struct throtl_grp, bps[READ]),
1431		.seq_show = tg_print_conf_u64,
1432		.write = tg_set_conf_u64,
1433	},
1434	{
1435		.name = "throttle.write_bps_device",
1436		.private = offsetof(struct throtl_grp, bps[WRITE]),
1437		.seq_show = tg_print_conf_u64,
1438		.write = tg_set_conf_u64,
1439	},
1440	{
1441		.name = "throttle.read_iops_device",
1442		.private = offsetof(struct throtl_grp, iops[READ]),
1443		.seq_show = tg_print_conf_uint,
1444		.write = tg_set_conf_uint,
1445	},
1446	{
1447		.name = "throttle.write_iops_device",
1448		.private = offsetof(struct throtl_grp, iops[WRITE]),
1449		.seq_show = tg_print_conf_uint,
1450		.write = tg_set_conf_uint,
1451	},
1452	{
1453		.name = "throttle.io_service_bytes",
1454		.private = offsetof(struct throtl_grp, stat_bytes),
1455		.seq_show = tg_print_rwstat,
1456	},
1457	{
1458		.name = "throttle.io_service_bytes_recursive",
1459		.private = offsetof(struct throtl_grp, stat_bytes),
1460		.seq_show = tg_print_rwstat_recursive,
1461	},
1462	{
1463		.name = "throttle.io_serviced",
1464		.private = offsetof(struct throtl_grp, stat_ios),
1465		.seq_show = tg_print_rwstat,
1466	},
1467	{
1468		.name = "throttle.io_serviced_recursive",
1469		.private = offsetof(struct throtl_grp, stat_ios),
1470		.seq_show = tg_print_rwstat_recursive,
1471	},
1472	{ }	/* terminate */
1473};
1474
1475static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1476			 int off)
1477{
1478	struct throtl_grp *tg = pd_to_tg(pd);
1479	const char *dname = blkg_dev_name(pd->blkg);
1480	u64 bps_dft;
1481	unsigned int iops_dft;
1482
1483	if (!dname)
1484		return 0;
1485
1486	bps_dft = U64_MAX;
1487	iops_dft = UINT_MAX;
1488
1489	if (tg->bps[READ] == bps_dft &&
1490	    tg->bps[WRITE] == bps_dft &&
1491	    tg->iops[READ] == iops_dft &&
1492	    tg->iops[WRITE] == iops_dft)
1493		return 0;
1494
1495	seq_printf(sf, "%s", dname);
1496	if (tg->bps[READ] == U64_MAX)
1497		seq_printf(sf, " rbps=max");
1498	else
1499		seq_printf(sf, " rbps=%llu", tg->bps[READ]);
1500
1501	if (tg->bps[WRITE] == U64_MAX)
1502		seq_printf(sf, " wbps=max");
1503	else
1504		seq_printf(sf, " wbps=%llu", tg->bps[WRITE]);
1505
1506	if (tg->iops[READ] == UINT_MAX)
1507		seq_printf(sf, " riops=max");
1508	else
1509		seq_printf(sf, " riops=%u", tg->iops[READ]);
1510
1511	if (tg->iops[WRITE] == UINT_MAX)
1512		seq_printf(sf, " wiops=max");
1513	else
1514		seq_printf(sf, " wiops=%u", tg->iops[WRITE]);
1515
1516	seq_printf(sf, "\n");
1517	return 0;
1518}
1519
1520static int tg_print_limit(struct seq_file *sf, void *v)
1521{
1522	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1523			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1524	return 0;
1525}
1526
1527static ssize_t tg_set_limit(struct kernfs_open_file *of,
1528			  char *buf, size_t nbytes, loff_t off)
1529{
1530	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1531	struct blkg_conf_ctx ctx;
1532	struct throtl_grp *tg;
1533	u64 v[4];
1534	int ret;
1535
1536	blkg_conf_init(&ctx, buf);
1537
1538	ret = blkg_conf_open_bdev(&ctx);
1539	if (ret)
1540		goto out_finish;
1541
1542	if (!blk_throtl_activated(ctx.bdev->bd_queue)) {
1543		ret = blk_throtl_init(ctx.bdev->bd_disk);
1544		if (ret)
1545			goto out_finish;
1546	}
1547
1548	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx);
1549	if (ret)
1550		goto out_finish;
1551
1552	tg = blkg_to_tg(ctx.blkg);
1553	tg_update_carryover(tg);
1554
1555	v[0] = tg->bps[READ];
1556	v[1] = tg->bps[WRITE];
1557	v[2] = tg->iops[READ];
1558	v[3] = tg->iops[WRITE];
1559
1560	while (true) {
1561		char tok[27];	/* wiops=18446744073709551616 */
1562		char *p;
1563		u64 val = U64_MAX;
1564		int len;
1565
1566		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1567			break;
1568		if (tok[0] == '\0')
1569			break;
1570		ctx.body += len;
1571
1572		ret = -EINVAL;
1573		p = tok;
1574		strsep(&p, "=");
1575		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1576			goto out_finish;
1577
1578		ret = -ERANGE;
1579		if (!val)
1580			goto out_finish;
1581
1582		ret = -EINVAL;
1583		if (!strcmp(tok, "rbps"))
1584			v[0] = val;
1585		else if (!strcmp(tok, "wbps"))
1586			v[1] = val;
1587		else if (!strcmp(tok, "riops"))
1588			v[2] = min_t(u64, val, UINT_MAX);
1589		else if (!strcmp(tok, "wiops"))
1590			v[3] = min_t(u64, val, UINT_MAX);
1591		else
1592			goto out_finish;
1593	}
1594
1595	tg->bps[READ] = v[0];
1596	tg->bps[WRITE] = v[1];
1597	tg->iops[READ] = v[2];
1598	tg->iops[WRITE] = v[3];
1599
1600	tg_conf_updated(tg, false);
1601	ret = 0;
1602out_finish:
1603	blkg_conf_exit(&ctx);
1604	return ret ?: nbytes;
1605}
1606
1607static struct cftype throtl_files[] = {
1608	{
1609		.name = "max",
1610		.flags = CFTYPE_NOT_ON_ROOT,
1611		.seq_show = tg_print_limit,
1612		.write = tg_set_limit,
1613	},
1614	{ }	/* terminate */
1615};
1616
1617static void throtl_shutdown_wq(struct request_queue *q)
1618{
1619	struct throtl_data *td = q->td;
1620
1621	cancel_work_sync(&td->dispatch_work);
1622}
1623
1624static void tg_flush_bios(struct throtl_grp *tg)
1625{
1626	struct throtl_service_queue *sq = &tg->service_queue;
1627
1628	if (tg->flags & THROTL_TG_CANCELING)
1629		return;
1630	/*
1631	 * Set the flag to make sure throtl_pending_timer_fn() won't
1632	 * stop until all throttled bios are dispatched.
1633	 */
1634	tg->flags |= THROTL_TG_CANCELING;
1635
1636	/*
1637	 * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup
1638	 * will be inserted to service queue without THROTL_TG_PENDING
1639	 * set in tg_update_disptime below. Then IO dispatched from
1640	 * child in tg_dispatch_one_bio will trigger double insertion
1641	 * and corrupt the tree.
1642	 */
1643	if (!(tg->flags & THROTL_TG_PENDING))
1644		return;
1645
1646	/*
1647	 * Update disptime after setting the above flag to make sure
1648	 * throtl_select_dispatch() won't exit without dispatching.
1649	 */
1650	tg_update_disptime(tg);
1651
1652	throtl_schedule_pending_timer(sq, jiffies + 1);
1653}
1654
1655static void throtl_pd_offline(struct blkg_policy_data *pd)
1656{
1657	tg_flush_bios(pd_to_tg(pd));
1658}
1659
1660struct blkcg_policy blkcg_policy_throtl = {
1661	.dfl_cftypes		= throtl_files,
1662	.legacy_cftypes		= throtl_legacy_files,
1663
1664	.pd_alloc_fn		= throtl_pd_alloc,
1665	.pd_init_fn		= throtl_pd_init,
1666	.pd_online_fn		= throtl_pd_online,
1667	.pd_offline_fn		= throtl_pd_offline,
1668	.pd_free_fn		= throtl_pd_free,
1669};
1670
1671void blk_throtl_cancel_bios(struct gendisk *disk)
1672{
1673	struct request_queue *q = disk->queue;
1674	struct cgroup_subsys_state *pos_css;
1675	struct blkcg_gq *blkg;
1676
1677	if (!blk_throtl_activated(q))
1678		return;
1679
1680	spin_lock_irq(&q->queue_lock);
1681	/*
1682	 * queue_lock is held, rcu lock is not needed here technically.
1683	 * However, rcu lock is still held to emphasize that following
1684	 * path need RCU protection and to prevent warning from lockdep.
1685	 */
1686	rcu_read_lock();
1687	blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) {
1688		/*
1689		 * disk_release will call pd_offline_fn to cancel bios.
1690		 * However, disk_release can't be called if someone get
1691		 * the refcount of device and issued bios which are
1692		 * inflight after del_gendisk.
1693		 * Cancel bios here to ensure no bios are inflight after
1694		 * del_gendisk.
1695		 */
1696		tg_flush_bios(blkg_to_tg(blkg));
1697	}
1698	rcu_read_unlock();
1699	spin_unlock_irq(&q->queue_lock);
1700}
1701
1702static bool tg_within_limit(struct throtl_grp *tg, struct bio *bio, bool rw)
1703{
1704	struct throtl_service_queue *sq = &tg->service_queue;
1705
1706	/*
1707	 * For a split bio, we need to specifically distinguish whether the
1708	 * iops queue is empty.
1709	 */
1710	if (bio_flagged(bio, BIO_BPS_THROTTLED))
1711		return sq->nr_queued_iops[rw] == 0 &&
1712				tg_dispatch_iops_time(tg, bio) == 0;
1713
1714	/*
1715	 * Throtl is FIFO - if bios are already queued, should queue.
1716	 * If the bps queue is empty and @bio is within the bps limit, charge
1717	 * bps here for direct placement into the iops queue.
1718	 */
1719	if (sq_queued(&tg->service_queue, rw)) {
1720		if (sq->nr_queued_bps[rw] == 0 &&
1721		    tg_dispatch_bps_time(tg, bio) == 0)
1722			throtl_charge_bps_bio(tg, bio);
1723
1724		return false;
1725	}
1726
1727	return tg_dispatch_time(tg, bio) == 0;
1728}
1729
1730bool __blk_throtl_bio(struct bio *bio)
1731{
1732	struct request_queue *q = bdev_get_queue(bio->bi_bdev);
1733	struct blkcg_gq *blkg = bio->bi_blkg;
1734	struct throtl_qnode *qn = NULL;
1735	struct throtl_grp *tg = blkg_to_tg(blkg);
1736	struct throtl_service_queue *sq;
1737	bool rw = bio_data_dir(bio);
1738	bool throttled = false;
1739	struct throtl_data *td = tg->td;
1740
1741	rcu_read_lock();
1742	spin_lock_irq(&q->queue_lock);
1743	sq = &tg->service_queue;
1744
1745	while (true) {
1746		if (tg_within_limit(tg, bio, rw)) {
1747			/* within limits, let's charge and dispatch directly */
1748			throtl_charge_iops_bio(tg, bio);
1749
1750			/*
1751			 * We need to trim slice even when bios are not being
1752			 * queued otherwise it might happen that a bio is not
1753			 * queued for a long time and slice keeps on extending
1754			 * and trim is not called for a long time. Now if limits
1755			 * are reduced suddenly we take into account all the IO
1756			 * dispatched so far at new low rate and * newly queued
1757			 * IO gets a really long dispatch time.
1758			 *
1759			 * So keep on trimming slice even if bio is not queued.
1760			 */
1761			throtl_trim_slice(tg, rw);
1762		} else if (bio_issue_as_root_blkg(bio)) {
1763			/*
1764			 * IOs which may cause priority inversions are
1765			 * dispatched directly, even if they're over limit.
1766			 *
1767			 * Charge and dispatch directly, and our throttle
1768			 * control algorithm is adaptive, and extra IO bytes
1769			 * will be throttled for paying the debt
1770			 */
1771			throtl_charge_bps_bio(tg, bio);
1772			throtl_charge_iops_bio(tg, bio);
1773		} else {
1774			/* if above limits, break to queue */
1775			break;
1776		}
1777
1778		/*
1779		 * @bio passed through this layer without being throttled.
1780		 * Climb up the ladder.  If we're already at the top, it
1781		 * can be executed directly.
1782		 */
1783		qn = &tg->qnode_on_parent[rw];
1784		sq = sq->parent_sq;
1785		tg = sq_to_tg(sq);
1786		if (!tg) {
1787			bio_set_flag(bio, BIO_BPS_THROTTLED);
1788			goto out_unlock;
1789		}
1790	}
1791
1792	/* out-of-limit, queue to @tg */
1793	throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1794		   rw == READ ? 'R' : 'W',
1795		   tg->bytes_disp[rw], bio->bi_iter.bi_size,
1796		   tg_bps_limit(tg, rw),
1797		   tg->io_disp[rw], tg_iops_limit(tg, rw),
1798		   sq_queued(sq, READ), sq_queued(sq, WRITE));
1799
1800	td->nr_queued[rw]++;
1801	throtl_add_bio_tg(bio, qn, tg);
1802	throttled = true;
1803
1804	/*
1805	 * Update @tg's dispatch time and force schedule dispatch if @tg
1806	 * was empty before @bio, or the iops queue is empty and @bio will
1807	 * add to.  The forced scheduling isn't likely to cause undue
1808	 * delay as @bio is likely to be dispatched directly if its @tg's
1809	 * disptime is not in the future.
1810	 */
1811	if (tg->flags & THROTL_TG_WAS_EMPTY ||
1812	    tg->flags & THROTL_TG_IOPS_WAS_EMPTY) {
1813		tg_update_disptime(tg);
1814		throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1815	}
1816
1817out_unlock:
1818	spin_unlock_irq(&q->queue_lock);
1819
1820	rcu_read_unlock();
1821	return throttled;
1822}
1823
1824void blk_throtl_exit(struct gendisk *disk)
1825{
1826	struct request_queue *q = disk->queue;
1827
1828	/*
1829	 * blkg_destroy_all() already deactivate throtl policy, just check and
1830	 * free throtl data.
1831	 */
1832	if (!q->td)
1833		return;
1834
1835	timer_delete_sync(&q->td->service_queue.pending_timer);
1836	throtl_shutdown_wq(q);
1837	kfree(q->td);
1838}
1839
1840static int __init throtl_init(void)
1841{
1842	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1843	if (!kthrotld_workqueue)
1844		panic("Failed to create kthrotld\n");
1845
1846	return blkcg_policy_register(&blkcg_policy_throtl);
1847}
1848
1849module_init(throtl_init);