Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 * Interface for controlling IO bandwidth on a request queue
   3 *
   4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   5 */
   6
   7#include <linux/module.h>
   8#include <linux/slab.h>
   9#include <linux/blkdev.h>
  10#include <linux/bio.h>
  11#include <linux/blktrace_api.h>
  12#include "blk-cgroup.h"
 
  13
  14/* Max dispatch from a group in 1 round */
  15static int throtl_grp_quantum = 8;
  16
  17/* Total max dispatch from all groups in one round */
  18static int throtl_quantum = 32;
  19
  20/* Throttling is performed over 100ms slice and after that slice is renewed */
  21static unsigned long throtl_slice = HZ/10;	/* 100 ms */
  22
 
 
  23/* A workqueue to queue throttle related work */
  24static struct workqueue_struct *kthrotld_workqueue;
  25static void throtl_schedule_delayed_work(struct throtl_data *td,
  26				unsigned long delay);
  27
  28struct throtl_rb_root {
  29	struct rb_root rb;
  30	struct rb_node *left;
  31	unsigned int count;
  32	unsigned long min_disptime;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  33};
  34
  35#define THROTL_RB_ROOT	(struct throtl_rb_root) { .rb = RB_ROOT, .left = NULL, \
  36			.count = 0, .min_disptime = 0}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  37
  38#define rb_entry_tg(node)	rb_entry((node), struct throtl_grp, rb_node)
  39
  40struct throtl_grp {
  41	/* List of throtl groups on the request queue*/
  42	struct hlist_node tg_node;
  43
  44	/* active throtl group service_tree member */
  45	struct rb_node rb_node;
  46
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  47	/*
  48	 * Dispatch time in jiffies. This is the estimated time when group
  49	 * will unthrottle and is ready to dispatch more bio. It is used as
  50	 * key to sort active groups in service tree.
  51	 */
  52	unsigned long disptime;
  53
  54	struct blkio_group blkg;
  55	atomic_t ref;
  56	unsigned int flags;
  57
  58	/* Two lists for READ and WRITE */
  59	struct bio_list bio_lists[2];
  60
  61	/* Number of queued bios on READ and WRITE lists */
  62	unsigned int nr_queued[2];
  63
  64	/* bytes per second rate limits */
  65	uint64_t bps[2];
  66
  67	/* IOPS limits */
  68	unsigned int iops[2];
  69
  70	/* Number of bytes disptached in current slice */
  71	uint64_t bytes_disp[2];
  72	/* Number of bio's dispatched in current slice */
  73	unsigned int io_disp[2];
  74
  75	/* When did we start a new slice */
  76	unsigned long slice_start[2];
  77	unsigned long slice_end[2];
  78
  79	/* Some throttle limits got updated for the group */
  80	int limits_changed;
  81
  82	struct rcu_head rcu_head;
  83};
  84
  85struct throtl_data
  86{
  87	/* List of throtl groups */
  88	struct hlist_head tg_list;
  89
  90	/* service tree for active throtl groups */
  91	struct throtl_rb_root tg_service_tree;
  92
  93	struct throtl_grp *root_tg;
  94	struct request_queue *queue;
  95
  96	/* Total Number of queued bios on READ and WRITE lists */
  97	unsigned int nr_queued[2];
  98
  99	/*
 100	 * number of total undestroyed groups
 101	 */
 102	unsigned int nr_undestroyed_grps;
 103
 104	/* Work for dispatching throttled bios */
 105	struct delayed_work throtl_work;
 106
 107	int limits_changed;
 108};
 109
 110enum tg_state_flags {
 111	THROTL_TG_FLAG_on_rr = 0,	/* on round-robin busy list */
 112};
 113
 114#define THROTL_TG_FNS(name)						\
 115static inline void throtl_mark_tg_##name(struct throtl_grp *tg)		\
 116{									\
 117	(tg)->flags |= (1 << THROTL_TG_FLAG_##name);			\
 118}									\
 119static inline void throtl_clear_tg_##name(struct throtl_grp *tg)	\
 120{									\
 121	(tg)->flags &= ~(1 << THROTL_TG_FLAG_##name);			\
 122}									\
 123static inline int throtl_tg_##name(const struct throtl_grp *tg)		\
 124{									\
 125	return ((tg)->flags & (1 << THROTL_TG_FLAG_##name)) != 0;	\
 126}
 127
 128THROTL_TG_FNS(on_rr);
 129
 130#define throtl_log_tg(td, tg, fmt, args...)				\
 131	blk_add_trace_msg((td)->queue, "throtl %s " fmt,		\
 132				blkg_path(&(tg)->blkg), ##args);      	\
 133
 134#define throtl_log(td, fmt, args...)	\
 135	blk_add_trace_msg((td)->queue, "throtl " fmt, ##args)
 136
 137static inline struct throtl_grp *tg_of_blkg(struct blkio_group *blkg)
 138{
 139	if (blkg)
 140		return container_of(blkg, struct throtl_grp, blkg);
 141
 142	return NULL;
 143}
 144
 145static inline unsigned int total_nr_queued(struct throtl_data *td)
 146{
 147	return td->nr_queued[0] + td->nr_queued[1];
 148}
 149
 150static inline struct throtl_grp *throtl_ref_get_tg(struct throtl_grp *tg)
 151{
 152	atomic_inc(&tg->ref);
 153	return tg;
 154}
 155
 156static void throtl_free_tg(struct rcu_head *head)
 157{
 158	struct throtl_grp *tg;
 159
 160	tg = container_of(head, struct throtl_grp, rcu_head);
 161	free_percpu(tg->blkg.stats_cpu);
 162	kfree(tg);
 163}
 164
 165static void throtl_put_tg(struct throtl_grp *tg)
 166{
 167	BUG_ON(atomic_read(&tg->ref) <= 0);
 168	if (!atomic_dec_and_test(&tg->ref))
 169		return;
 170
 171	/*
 172	 * A group is freed in rcu manner. But having an rcu lock does not
 173	 * mean that one can access all the fields of blkg and assume these
 174	 * are valid. For example, don't try to follow throtl_data and
 175	 * request queue links.
 176	 *
 177	 * Having a reference to blkg under an rcu allows acess to only
 178	 * values local to groups like group stats and group rate limits
 179	 */
 180	call_rcu(&tg->rcu_head, throtl_free_tg);
 181}
 182
 183static void throtl_init_group(struct throtl_grp *tg)
 
 
 
 
 
 
 
 184{
 185	INIT_HLIST_NODE(&tg->tg_node);
 186	RB_CLEAR_NODE(&tg->rb_node);
 187	bio_list_init(&tg->bio_lists[0]);
 188	bio_list_init(&tg->bio_lists[1]);
 189	tg->limits_changed = false;
 190
 191	/* Practically unlimited BW */
 192	tg->bps[0] = tg->bps[1] = -1;
 193	tg->iops[0] = tg->iops[1] = -1;
 194
 195	/*
 196	 * Take the initial reference that will be released on destroy
 197	 * This can be thought of a joint reference by cgroup and
 198	 * request queue which will be dropped by either request queue
 199	 * exit or cgroup deletion path depending on who is exiting first.
 200	 */
 201	atomic_set(&tg->ref, 1);
 202}
 203
 204/* Should be called with rcu read lock held (needed for blkcg) */
 205static void
 206throtl_add_group_to_td_list(struct throtl_data *td, struct throtl_grp *tg)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 207{
 208	hlist_add_head(&tg->tg_node, &td->tg_list);
 209	td->nr_undestroyed_grps++;
 
 
 
 210}
 211
 212static void
 213__throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
 
 
 
 214{
 215	struct backing_dev_info *bdi = &td->queue->backing_dev_info;
 216	unsigned int major, minor;
 217
 218	if (!tg || tg->blkg.dev)
 219		return;
 220
 221	/*
 222	 * Fill in device details for a group which might not have been
 223	 * filled at group creation time as queue was being instantiated
 224	 * and driver had not attached a device yet
 225	 */
 226	if (bdi->dev && dev_name(bdi->dev)) {
 227		sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
 228		tg->blkg.dev = MKDEV(major, minor);
 229	}
 230}
 231
 232/*
 233 * Should be called with without queue lock held. Here queue lock will be
 234 * taken rarely. It will be taken only once during life time of a group
 235 * if need be
 
 
 
 
 
 
 
 
 
 236 */
 237static void
 238throtl_tg_fill_dev_details(struct throtl_data *td, struct throtl_grp *tg)
 239{
 240	if (!tg || tg->blkg.dev)
 241		return;
 242
 243	spin_lock_irq(td->queue->queue_lock);
 244	__throtl_tg_fill_dev_details(td, tg);
 245	spin_unlock_irq(td->queue->queue_lock);
 246}
 247
 248static void throtl_init_add_tg_lists(struct throtl_data *td,
 249			struct throtl_grp *tg, struct blkio_cgroup *blkcg)
 250{
 251	__throtl_tg_fill_dev_details(td, tg);
 252
 253	/* Add group onto cgroup list */
 254	blkiocg_add_blkio_group(blkcg, &tg->blkg, (void *)td,
 255				tg->blkg.dev, BLKIO_POLICY_THROTL);
 
 
 
 
 
 
 256
 257	tg->bps[READ] = blkcg_get_read_bps(blkcg, tg->blkg.dev);
 258	tg->bps[WRITE] = blkcg_get_write_bps(blkcg, tg->blkg.dev);
 259	tg->iops[READ] = blkcg_get_read_iops(blkcg, tg->blkg.dev);
 260	tg->iops[WRITE] = blkcg_get_write_iops(blkcg, tg->blkg.dev);
 261
 262	throtl_add_group_to_td_list(td, tg);
 
 
 
 
 
 
 
 263}
 264
 265/* Should be called without queue lock and outside of rcu period */
 266static struct throtl_grp *throtl_alloc_tg(struct throtl_data *td)
 267{
 268	struct throtl_grp *tg = NULL;
 269	int ret;
 270
 271	tg = kzalloc_node(sizeof(*tg), GFP_ATOMIC, td->queue->node);
 272	if (!tg)
 273		return NULL;
 274
 275	ret = blkio_alloc_blkg_stats(&tg->blkg);
 276
 277	if (ret) {
 278		kfree(tg);
 279		return NULL;
 280	}
 281
 282	throtl_init_group(tg);
 283	return tg;
 
 
 
 
 
 284}
 285
 286static struct
 287throtl_grp *throtl_find_tg(struct throtl_data *td, struct blkio_cgroup *blkcg)
 288{
 289	struct throtl_grp *tg = NULL;
 290	void *key = td;
 
 
 291
 292	/*
 293	 * This is the common case when there are no blkio cgroups.
 294 	 * Avoid lookup in this case
 295 	 */
 296	if (blkcg == &blkio_root_cgroup)
 297		tg = td->root_tg;
 298	else
 299		tg = tg_of_blkg(blkiocg_lookup_group(blkcg, key));
 300
 301	__throtl_tg_fill_dev_details(td, tg);
 302	return tg;
 
 
 
 
 
 
 303}
 304
 305/*
 306 * This function returns with queue lock unlocked in case of error, like
 307 * request queue is no more
 
 308 */
 309static struct throtl_grp * throtl_get_tg(struct throtl_data *td)
 310{
 311	struct throtl_grp *tg = NULL, *__tg = NULL;
 312	struct blkio_cgroup *blkcg;
 313	struct request_queue *q = td->queue;
 314
 315	rcu_read_lock();
 316	blkcg = task_blkio_cgroup(current);
 317	tg = throtl_find_tg(td, blkcg);
 318	if (tg) {
 319		rcu_read_unlock();
 320		return tg;
 321	}
 322
 323	/*
 324	 * Need to allocate a group. Allocation of group also needs allocation
 325	 * of per cpu stats which in-turn takes a mutex() and can block. Hence
 326	 * we need to drop rcu lock and queue_lock before we call alloc
 327	 *
 328	 * Take the request queue reference to make sure queue does not
 329	 * go away once we return from allocation.
 330	 */
 331	blk_get_queue(q);
 332	rcu_read_unlock();
 333	spin_unlock_irq(q->queue_lock);
 334
 335	tg = throtl_alloc_tg(td);
 336	/*
 337	 * We might have slept in group allocation. Make sure queue is not
 338	 * dead
 339	 */
 340	if (unlikely(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) {
 341		blk_put_queue(q);
 342		if (tg)
 343			kfree(tg);
 344
 345		return ERR_PTR(-ENODEV);
 346	}
 347	blk_put_queue(q);
 348
 349	/* Group allocated and queue is still alive. take the lock */
 350	spin_lock_irq(q->queue_lock);
 351
 352	/*
 353	 * Initialize the new group. After sleeping, read the blkcg again.
 354	 */
 355	rcu_read_lock();
 356	blkcg = task_blkio_cgroup(current);
 357
 
 
 358	/*
 359	 * If some other thread already allocated the group while we were
 360	 * not holding queue lock, free up the group
 361	 */
 362	__tg = throtl_find_tg(td, blkcg);
 363
 364	if (__tg) {
 365		kfree(tg);
 366		rcu_read_unlock();
 367		return __tg;
 368	}
 369
 370	/* Group allocation failed. Account the IO to root group */
 371	if (!tg) {
 372		tg = td->root_tg;
 373		return tg;
 374	}
 375
 376	throtl_init_add_tg_lists(td, tg, blkcg);
 377	rcu_read_unlock();
 378	return tg;
 379}
 380
 381static struct throtl_grp *throtl_rb_first(struct throtl_rb_root *root)
 
 382{
 383	/* Service tree is empty */
 384	if (!root->count)
 385		return NULL;
 386
 387	if (!root->left)
 388		root->left = rb_first(&root->rb);
 389
 390	if (root->left)
 391		return rb_entry_tg(root->left);
 392
 393	return NULL;
 394}
 395
 396static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 397{
 398	rb_erase(n, root);
 399	RB_CLEAR_NODE(n);
 400}
 401
 402static void throtl_rb_erase(struct rb_node *n, struct throtl_rb_root *root)
 
 403{
 404	if (root->left == n)
 405		root->left = NULL;
 406	rb_erase_init(n, &root->rb);
 407	--root->count;
 408}
 409
 410static void update_min_dispatch_time(struct throtl_rb_root *st)
 411{
 412	struct throtl_grp *tg;
 413
 414	tg = throtl_rb_first(st);
 415	if (!tg)
 416		return;
 417
 418	st->min_disptime = tg->disptime;
 419}
 420
 421static void
 422tg_service_tree_add(struct throtl_rb_root *st, struct throtl_grp *tg)
 423{
 424	struct rb_node **node = &st->rb.rb_node;
 
 425	struct rb_node *parent = NULL;
 426	struct throtl_grp *__tg;
 427	unsigned long key = tg->disptime;
 428	int left = 1;
 429
 430	while (*node != NULL) {
 431		parent = *node;
 432		__tg = rb_entry_tg(parent);
 433
 434		if (time_before(key, __tg->disptime))
 435			node = &parent->rb_left;
 436		else {
 437			node = &parent->rb_right;
 438			left = 0;
 439		}
 440	}
 441
 442	if (left)
 443		st->left = &tg->rb_node;
 444
 445	rb_link_node(&tg->rb_node, parent, node);
 446	rb_insert_color(&tg->rb_node, &st->rb);
 447}
 448
 449static void __throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
 450{
 451	struct throtl_rb_root *st = &td->tg_service_tree;
 
 
 
 452
 453	tg_service_tree_add(st, tg);
 454	throtl_mark_tg_on_rr(tg);
 455	st->count++;
 
 456}
 457
 458static void throtl_enqueue_tg(struct throtl_data *td, struct throtl_grp *tg)
 459{
 460	if (!throtl_tg_on_rr(tg))
 461		__throtl_enqueue_tg(td, tg);
 462}
 463
 464static void __throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
 465{
 466	throtl_rb_erase(&tg->rb_node, &td->tg_service_tree);
 467	throtl_clear_tg_on_rr(tg);
 468}
 469
 470static void throtl_dequeue_tg(struct throtl_data *td, struct throtl_grp *tg)
 
 
 471{
 472	if (throtl_tg_on_rr(tg))
 473		__throtl_dequeue_tg(td, tg);
 
 474}
 475
 476static void throtl_schedule_next_dispatch(struct throtl_data *td)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 477{
 478	struct throtl_rb_root *st = &td->tg_service_tree;
 
 
 479
 480	/*
 481	 * If there are more bios pending, schedule more work.
 482	 */
 483	if (!total_nr_queued(td))
 484		return;
 485
 486	BUG_ON(!st->count);
 
 
 
 
 487
 488	update_min_dispatch_time(st);
 
 
 489
 490	if (time_before_eq(st->min_disptime, jiffies))
 491		throtl_schedule_delayed_work(td, 0);
 492	else
 493		throtl_schedule_delayed_work(td, (st->min_disptime - jiffies));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 494}
 495
 496static inline void
 497throtl_start_new_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 498{
 499	tg->bytes_disp[rw] = 0;
 500	tg->io_disp[rw] = 0;
 501	tg->slice_start[rw] = jiffies;
 502	tg->slice_end[rw] = jiffies + throtl_slice;
 503	throtl_log_tg(td, tg, "[%c] new slice start=%lu end=%lu jiffies=%lu",
 504			rw == READ ? 'R' : 'W', tg->slice_start[rw],
 505			tg->slice_end[rw], jiffies);
 
 506}
 507
 508static inline void throtl_set_slice_end(struct throtl_data *td,
 509		struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
 510{
 511	tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 512}
 513
 514static inline void throtl_extend_slice(struct throtl_data *td,
 515		struct throtl_grp *tg, bool rw, unsigned long jiffy_end)
 516{
 517	tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 518	throtl_log_tg(td, tg, "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 519			rw == READ ? 'R' : 'W', tg->slice_start[rw],
 520			tg->slice_end[rw], jiffies);
 
 521}
 522
 523/* Determine if previously allocated or extended slice is complete or not */
 524static bool
 525throtl_slice_used(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 526{
 527	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 528		return 0;
 529
 530	return 1;
 531}
 532
 533/* Trim the used slices and adjust slice start accordingly */
 534static inline void
 535throtl_trim_slice(struct throtl_data *td, struct throtl_grp *tg, bool rw)
 536{
 537	unsigned long nr_slices, time_elapsed, io_trim;
 538	u64 bytes_trim, tmp;
 539
 540	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 541
 542	/*
 543	 * If bps are unlimited (-1), then time slice don't get
 544	 * renewed. Don't try to trim the slice if slice is used. A new
 545	 * slice will start when appropriate.
 546	 */
 547	if (throtl_slice_used(td, tg, rw))
 548		return;
 549
 550	/*
 551	 * A bio has been dispatched. Also adjust slice_end. It might happen
 552	 * that initially cgroup limit was very low resulting in high
 553	 * slice_end, but later limit was bumped up and bio was dispached
 554	 * sooner, then we need to reduce slice_end. A high bogus slice_end
 555	 * is bad because it does not allow new slice to start.
 556	 */
 557
 558	throtl_set_slice_end(td, tg, rw, jiffies + throtl_slice);
 559
 560	time_elapsed = jiffies - tg->slice_start[rw];
 561
 562	nr_slices = time_elapsed / throtl_slice;
 563
 564	if (!nr_slices)
 565		return;
 566	tmp = tg->bps[rw] * throtl_slice * nr_slices;
 567	do_div(tmp, HZ);
 568	bytes_trim = tmp;
 569
 570	io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
 571
 572	if (!bytes_trim && !io_trim)
 573		return;
 574
 575	if (tg->bytes_disp[rw] >= bytes_trim)
 576		tg->bytes_disp[rw] -= bytes_trim;
 577	else
 578		tg->bytes_disp[rw] = 0;
 579
 580	if (tg->io_disp[rw] >= io_trim)
 581		tg->io_disp[rw] -= io_trim;
 582	else
 583		tg->io_disp[rw] = 0;
 584
 585	tg->slice_start[rw] += nr_slices * throtl_slice;
 586
 587	throtl_log_tg(td, tg, "[%c] trim slice nr=%lu bytes=%llu io=%lu"
 588			" start=%lu end=%lu jiffies=%lu",
 589			rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 590			tg->slice_start[rw], tg->slice_end[rw], jiffies);
 591}
 592
 593static bool tg_with_in_iops_limit(struct throtl_data *td, struct throtl_grp *tg,
 594		struct bio *bio, unsigned long *wait)
 595{
 596	bool rw = bio_data_dir(bio);
 597	unsigned int io_allowed;
 598	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 599	u64 tmp;
 600
 601	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 602
 603	/* Slice has just started. Consider one slice interval */
 604	if (!jiffy_elapsed)
 605		jiffy_elapsed_rnd = throtl_slice;
 606
 607	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 608
 609	/*
 610	 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 611	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 612	 * will allow dispatch after 1 second and after that slice should
 613	 * have been trimmed.
 614	 */
 615
 616	tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
 617	do_div(tmp, HZ);
 618
 619	if (tmp > UINT_MAX)
 620		io_allowed = UINT_MAX;
 621	else
 622		io_allowed = tmp;
 623
 624	if (tg->io_disp[rw] + 1 <= io_allowed) {
 625		if (wait)
 626			*wait = 0;
 627		return 1;
 628	}
 629
 630	/* Calc approx time to dispatch */
 631	jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
 632
 633	if (jiffy_wait > jiffy_elapsed)
 634		jiffy_wait = jiffy_wait - jiffy_elapsed;
 635	else
 636		jiffy_wait = 1;
 637
 638	if (wait)
 639		*wait = jiffy_wait;
 640	return 0;
 641}
 642
 643static bool tg_with_in_bps_limit(struct throtl_data *td, struct throtl_grp *tg,
 644		struct bio *bio, unsigned long *wait)
 645{
 646	bool rw = bio_data_dir(bio);
 647	u64 bytes_allowed, extra_bytes, tmp;
 648	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 649
 650	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 651
 652	/* Slice has just started. Consider one slice interval */
 653	if (!jiffy_elapsed)
 654		jiffy_elapsed_rnd = throtl_slice;
 655
 656	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 657
 658	tmp = tg->bps[rw] * jiffy_elapsed_rnd;
 659	do_div(tmp, HZ);
 660	bytes_allowed = tmp;
 661
 662	if (tg->bytes_disp[rw] + bio->bi_size <= bytes_allowed) {
 663		if (wait)
 664			*wait = 0;
 665		return 1;
 666	}
 667
 668	/* Calc approx time to dispatch */
 669	extra_bytes = tg->bytes_disp[rw] + bio->bi_size - bytes_allowed;
 670	jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
 671
 672	if (!jiffy_wait)
 673		jiffy_wait = 1;
 674
 675	/*
 676	 * This wait time is without taking into consideration the rounding
 677	 * up we did. Add that time also.
 678	 */
 679	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 680	if (wait)
 681		*wait = jiffy_wait;
 682	return 0;
 683}
 684
 685static bool tg_no_rule_group(struct throtl_grp *tg, bool rw) {
 686	if (tg->bps[rw] == -1 && tg->iops[rw] == -1)
 687		return 1;
 688	return 0;
 689}
 690
 691/*
 692 * Returns whether one can dispatch a bio or not. Also returns approx number
 693 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 694 */
 695static bool tg_may_dispatch(struct throtl_data *td, struct throtl_grp *tg,
 696				struct bio *bio, unsigned long *wait)
 697{
 698	bool rw = bio_data_dir(bio);
 699	unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 700
 701	/*
 702 	 * Currently whole state machine of group depends on first bio
 703	 * queued in the group bio list. So one should not be calling
 704	 * this function with a different bio if there are other bios
 705	 * queued.
 706	 */
 707	BUG_ON(tg->nr_queued[rw] && bio != bio_list_peek(&tg->bio_lists[rw]));
 
 708
 709	/* If tg->bps = -1, then BW is unlimited */
 710	if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
 711		if (wait)
 712			*wait = 0;
 713		return 1;
 714	}
 715
 716	/*
 717	 * If previous slice expired, start a new one otherwise renew/extend
 718	 * existing slice to make sure it is at least throtl_slice interval
 719	 * long since now.
 
 
 720	 */
 721	if (throtl_slice_used(td, tg, rw))
 722		throtl_start_new_slice(td, tg, rw);
 723	else {
 724		if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
 725			throtl_extend_slice(td, tg, rw, jiffies + throtl_slice);
 726	}
 727
 728	if (tg_with_in_bps_limit(td, tg, bio, &bps_wait)
 729	    && tg_with_in_iops_limit(td, tg, bio, &iops_wait)) {
 730		if (wait)
 731			*wait = 0;
 732		return 1;
 733	}
 734
 735	max_wait = max(bps_wait, iops_wait);
 736
 737	if (wait)
 738		*wait = max_wait;
 739
 740	if (time_before(tg->slice_end[rw], jiffies + max_wait))
 741		throtl_extend_slice(td, tg, rw, jiffies + max_wait);
 742
 743	return 0;
 744}
 745
 746static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 747{
 748	bool rw = bio_data_dir(bio);
 749	bool sync = rw_is_sync(bio->bi_rw);
 750
 751	/* Charge the bio to the group */
 752	tg->bytes_disp[rw] += bio->bi_size;
 753	tg->io_disp[rw]++;
 754
 755	blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size, rw, sync);
 
 
 
 
 
 
 
 756}
 757
 758static void throtl_add_bio_tg(struct throtl_data *td, struct throtl_grp *tg,
 759			struct bio *bio)
 
 
 
 
 
 
 
 
 
 760{
 
 761	bool rw = bio_data_dir(bio);
 762
 763	bio_list_add(&tg->bio_lists[rw], bio);
 764	/* Take a bio reference on tg */
 765	throtl_ref_get_tg(tg);
 766	tg->nr_queued[rw]++;
 767	td->nr_queued[rw]++;
 768	throtl_enqueue_tg(td, tg);
 
 
 
 
 
 
 
 
 
 
 769}
 770
 771static void tg_update_disptime(struct throtl_data *td, struct throtl_grp *tg)
 772{
 
 773	unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
 774	struct bio *bio;
 775
 776	if ((bio = bio_list_peek(&tg->bio_lists[READ])))
 777		tg_may_dispatch(td, tg, bio, &read_wait);
 778
 779	if ((bio = bio_list_peek(&tg->bio_lists[WRITE])))
 780		tg_may_dispatch(td, tg, bio, &write_wait);
 781
 782	min_wait = min(read_wait, write_wait);
 783	disptime = jiffies + min_wait;
 784
 785	/* Update dispatch time */
 786	throtl_dequeue_tg(td, tg);
 787	tg->disptime = disptime;
 788	throtl_enqueue_tg(td, tg);
 
 
 
 789}
 790
 791static void tg_dispatch_one_bio(struct throtl_data *td, struct throtl_grp *tg,
 792				bool rw, struct bio_list *bl)
 793{
 794	struct bio *bio;
 
 
 
 795
 796	bio = bio_list_pop(&tg->bio_lists[rw]);
 797	tg->nr_queued[rw]--;
 798	/* Drop bio reference on tg */
 799	throtl_put_tg(tg);
 800
 801	BUG_ON(td->nr_queued[rw] <= 0);
 802	td->nr_queued[rw]--;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 803
 804	throtl_charge_bio(tg, bio);
 805	bio_list_add(bl, bio);
 806	bio->bi_rw |= REQ_THROTTLED;
 807
 808	throtl_trim_slice(td, tg, rw);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 809}
 810
 811static int throtl_dispatch_tg(struct throtl_data *td, struct throtl_grp *tg,
 812				struct bio_list *bl)
 813{
 
 814	unsigned int nr_reads = 0, nr_writes = 0;
 815	unsigned int max_nr_reads = throtl_grp_quantum*3/4;
 816	unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
 817	struct bio *bio;
 818
 819	/* Try to dispatch 75% READS and 25% WRITES */
 820
 821	while ((bio = bio_list_peek(&tg->bio_lists[READ]))
 822		&& tg_may_dispatch(td, tg, bio, NULL)) {
 823
 824		tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
 825		nr_reads++;
 826
 827		if (nr_reads >= max_nr_reads)
 828			break;
 829	}
 830
 831	while ((bio = bio_list_peek(&tg->bio_lists[WRITE]))
 832		&& tg_may_dispatch(td, tg, bio, NULL)) {
 833
 834		tg_dispatch_one_bio(td, tg, bio_data_dir(bio), bl);
 835		nr_writes++;
 836
 837		if (nr_writes >= max_nr_writes)
 838			break;
 839	}
 840
 841	return nr_reads + nr_writes;
 842}
 843
 844static int throtl_select_dispatch(struct throtl_data *td, struct bio_list *bl)
 845{
 846	unsigned int nr_disp = 0;
 847	struct throtl_grp *tg;
 848	struct throtl_rb_root *st = &td->tg_service_tree;
 849
 850	while (1) {
 851		tg = throtl_rb_first(st);
 
 852
 853		if (!tg)
 854			break;
 855
 856		if (time_before(jiffies, tg->disptime))
 857			break;
 858
 859		throtl_dequeue_tg(td, tg);
 860
 861		nr_disp += throtl_dispatch_tg(td, tg, bl);
 862
 863		if (tg->nr_queued[0] || tg->nr_queued[1]) {
 864			tg_update_disptime(td, tg);
 865			throtl_enqueue_tg(td, tg);
 866		}
 867
 868		if (nr_disp >= throtl_quantum)
 869			break;
 870	}
 871
 872	return nr_disp;
 873}
 874
 875static void throtl_process_limit_change(struct throtl_data *td)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 876{
 877	struct throtl_grp *tg;
 878	struct hlist_node *pos, *n;
 879
 880	if (!td->limits_changed)
 881		return;
 882
 883	xchg(&td->limits_changed, false);
 884
 885	throtl_log(td, "limits changed");
 886
 887	hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
 888		if (!tg->limits_changed)
 889			continue;
 890
 891		if (!xchg(&tg->limits_changed, false))
 892			continue;
 893
 894		throtl_log_tg(td, tg, "limit change rbps=%llu wbps=%llu"
 895			" riops=%u wiops=%u", tg->bps[READ], tg->bps[WRITE],
 896			tg->iops[READ], tg->iops[WRITE]);
 
 
 
 
 
 
 
 
 
 
 
 
 897
 898		/*
 899		 * Restart the slices for both READ and WRITES. It
 900		 * might happen that a group's limit are dropped
 901		 * suddenly and we don't want to account recently
 902		 * dispatched IO with new low rate
 903		 */
 904		throtl_start_new_slice(td, tg, 0);
 905		throtl_start_new_slice(td, tg, 1);
 906
 907		if (throtl_tg_on_rr(tg))
 908			tg_update_disptime(td, tg);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 909	}
 
 
 910}
 911
 912/* Dispatch throttled bios. Should be called without queue lock held. */
 913static int throtl_dispatch(struct request_queue *q)
 
 
 
 
 
 
 
 914{
 915	struct throtl_data *td = q->td;
 916	unsigned int nr_disp = 0;
 
 
 917	struct bio_list bio_list_on_stack;
 918	struct bio *bio;
 919	struct blk_plug plug;
 920
 921	spin_lock_irq(q->queue_lock);
 922
 923	throtl_process_limit_change(td);
 924
 925	if (!total_nr_queued(td))
 926		goto out;
 927
 928	bio_list_init(&bio_list_on_stack);
 929
 930	throtl_log(td, "dispatch nr_queued=%u read=%u write=%u",
 931			total_nr_queued(td), td->nr_queued[READ],
 932			td->nr_queued[WRITE]);
 933
 934	nr_disp = throtl_select_dispatch(td, &bio_list_on_stack);
 935
 936	if (nr_disp)
 937		throtl_log(td, "bios disp=%u", nr_disp);
 938
 939	throtl_schedule_next_dispatch(td);
 940out:
 941	spin_unlock_irq(q->queue_lock);
 942
 943	/*
 944	 * If we dispatched some requests, unplug the queue to make sure
 945	 * immediate dispatch
 946	 */
 947	if (nr_disp) {
 948		blk_start_plug(&plug);
 949		while((bio = bio_list_pop(&bio_list_on_stack)))
 950			generic_make_request(bio);
 951		blk_finish_plug(&plug);
 952	}
 953	return nr_disp;
 954}
 955
 956void blk_throtl_work(struct work_struct *work)
 
 957{
 958	struct throtl_data *td = container_of(work, struct throtl_data,
 959					throtl_work.work);
 960	struct request_queue *q = td->queue;
 961
 962	throtl_dispatch(q);
 
 
 963}
 964
 965/* Call with queue lock held */
 966static void
 967throtl_schedule_delayed_work(struct throtl_data *td, unsigned long delay)
 968{
 
 
 969
 970	struct delayed_work *dwork = &td->throtl_work;
 
 
 
 971
 972	/* schedule work if limits changed even if no bio is queued */
 973	if (total_nr_queued(td) || td->limits_changed) {
 974		/*
 975		 * We might have a work scheduled to be executed in future.
 976		 * Cancel that and schedule a new one.
 977		 */
 978		__cancel_delayed_work(dwork);
 979		queue_delayed_work(kthrotld_workqueue, dwork, delay);
 980		throtl_log(td, "schedule work. delay=%lu jiffies=%lu",
 981				delay, jiffies);
 982	}
 
 983}
 984
 985static void
 986throtl_destroy_tg(struct throtl_data *td, struct throtl_grp *tg)
 987{
 988	/* Something wrong if we are trying to remove same group twice */
 989	BUG_ON(hlist_unhashed(&tg->tg_node));
 
 990
 991	hlist_del_init(&tg->tg_node);
 
 
 
 992
 993	/*
 994	 * Put the reference taken at the time of creation so that when all
 995	 * queues are gone, group can be destroyed.
 
 
 
 996	 */
 997	throtl_put_tg(tg);
 998	td->nr_undestroyed_grps--;
 999}
1000
1001static void throtl_release_tgs(struct throtl_data *td)
1002{
1003	struct hlist_node *pos, *n;
1004	struct throtl_grp *tg;
 
 
 
 
 
 
1005
1006	hlist_for_each_entry_safe(tg, pos, n, &td->tg_list, tg_node) {
1007		/*
1008		 * If cgroup removal path got to blk_group first and removed
1009		 * it from cgroup list, then it will take care of destroying
1010		 * cfqg also.
1011		 */
1012		if (!blkiocg_del_blkio_group(&tg->blkg))
1013			throtl_destroy_tg(td, tg);
1014	}
1015}
1016
1017static void throtl_td_free(struct throtl_data *td)
 
1018{
1019	kfree(td);
1020}
 
 
 
1021
1022/*
1023 * Blk cgroup controller notification saying that blkio_group object is being
1024 * delinked as associated cgroup object is going away. That also means that
1025 * no new IO will come in this group. So get rid of this group as soon as
1026 * any pending IO in the group is finished.
1027 *
1028 * This function is called under rcu_read_lock(). key is the rcu protected
1029 * pointer. That means "key" is a valid throtl_data pointer as long as we are
1030 * rcu read lock.
1031 *
1032 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1033 * it should not be NULL as even if queue was going away, cgroup deltion
1034 * path got to it first.
1035 */
1036void throtl_unlink_blkio_group(void *key, struct blkio_group *blkg)
1037{
1038	unsigned long flags;
1039	struct throtl_data *td = key;
1040
1041	spin_lock_irqsave(td->queue->queue_lock, flags);
1042	throtl_destroy_tg(td, tg_of_blkg(blkg));
1043	spin_unlock_irqrestore(td->queue->queue_lock, flags);
 
 
 
 
 
 
 
1044}
1045
1046static void throtl_update_blkio_group_common(struct throtl_data *td,
1047				struct throtl_grp *tg)
1048{
1049	xchg(&tg->limits_changed, true);
1050	xchg(&td->limits_changed, true);
1051	/* Schedule a work now to process the limit change */
1052	throtl_schedule_delayed_work(td, 0);
1053}
1054
1055/*
1056 * For all update functions, key should be a valid pointer because these
1057 * update functions are called under blkcg_lock, that means, blkg is
1058 * valid and in turn key is valid. queue exit path can not race because
1059 * of blkcg_lock
1060 *
1061 * Can not take queue lock in update functions as queue lock under blkcg_lock
1062 * is not allowed. Under other paths we take blkcg_lock under queue_lock.
1063 */
1064static void throtl_update_blkio_group_read_bps(void *key,
1065				struct blkio_group *blkg, u64 read_bps)
1066{
1067	struct throtl_data *td = key;
1068	struct throtl_grp *tg = tg_of_blkg(blkg);
1069
1070	tg->bps[READ] = read_bps;
1071	throtl_update_blkio_group_common(td, tg);
1072}
1073
1074static void throtl_update_blkio_group_write_bps(void *key,
1075				struct blkio_group *blkg, u64 write_bps)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1076{
1077	struct throtl_data *td = key;
1078	struct throtl_grp *tg = tg_of_blkg(blkg);
 
1079
1080	tg->bps[WRITE] = write_bps;
1081	throtl_update_blkio_group_common(td, tg);
1082}
 
 
1083
1084static void throtl_update_blkio_group_read_iops(void *key,
1085			struct blkio_group *blkg, unsigned int read_iops)
1086{
1087	struct throtl_data *td = key;
1088	struct throtl_grp *tg = tg_of_blkg(blkg);
 
 
 
1089
1090	tg->iops[READ] = read_iops;
1091	throtl_update_blkio_group_common(td, tg);
 
1092}
1093
1094static void throtl_update_blkio_group_write_iops(void *key,
1095			struct blkio_group *blkg, unsigned int write_iops)
1096{
1097	struct throtl_data *td = key;
1098	struct throtl_grp *tg = tg_of_blkg(blkg);
1099
1100	tg->iops[WRITE] = write_iops;
1101	throtl_update_blkio_group_common(td, tg);
1102}
1103
1104static void throtl_shutdown_wq(struct request_queue *q)
 
1105{
1106	struct throtl_data *td = q->td;
 
 
 
 
1107
1108	cancel_delayed_work_sync(&td->throtl_work);
1109}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1110
1111static struct blkio_policy_type blkio_policy_throtl = {
1112	.ops = {
1113		.blkio_unlink_group_fn = throtl_unlink_blkio_group,
1114		.blkio_update_group_read_bps_fn =
1115					throtl_update_blkio_group_read_bps,
1116		.blkio_update_group_write_bps_fn =
1117					throtl_update_blkio_group_write_bps,
1118		.blkio_update_group_read_iops_fn =
1119					throtl_update_blkio_group_read_iops,
1120		.blkio_update_group_write_iops_fn =
1121					throtl_update_blkio_group_write_iops,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1122	},
1123	.plid = BLKIO_POLICY_THROTL,
1124};
1125
1126int blk_throtl_bio(struct request_queue *q, struct bio **biop)
1127{
1128	struct throtl_data *td = q->td;
1129	struct throtl_grp *tg;
1130	struct bio *bio = *biop;
1131	bool rw = bio_data_dir(bio), update_disptime = true;
1132	struct blkio_cgroup *blkcg;
1133
1134	if (bio->bi_rw & REQ_THROTTLED) {
1135		bio->bi_rw &= ~REQ_THROTTLED;
1136		return 0;
1137	}
1138
1139	/*
1140	 * A throtl_grp pointer retrieved under rcu can be used to access
1141	 * basic fields like stats and io rates. If a group has no rules,
1142	 * just update the dispatch stats in lockless manner and return.
1143	 */
1144
1145	rcu_read_lock();
1146	blkcg = task_blkio_cgroup(current);
1147	tg = throtl_find_tg(td, blkcg);
1148	if (tg) {
1149		throtl_tg_fill_dev_details(td, tg);
1150
1151		if (tg_no_rule_group(tg, rw)) {
1152			blkiocg_update_dispatch_stats(&tg->blkg, bio->bi_size,
1153					rw, rw_is_sync(bio->bi_rw));
1154			rcu_read_unlock();
1155			return 0;
1156		}
1157	}
1158	rcu_read_unlock();
1159
1160	/*
1161	 * Either group has not been allocated yet or it is not an unlimited
1162	 * IO group
1163	 */
 
 
 
 
 
 
 
 
 
 
1164
1165	spin_lock_irq(q->queue_lock);
1166	tg = throtl_get_tg(td);
1167
1168	if (IS_ERR(tg)) {
1169		if (PTR_ERR(tg)	== -ENODEV) {
1170			/*
1171			 * Queue is gone. No queue lock held here.
1172			 */
1173			return -ENODEV;
1174		}
1175	}
1176
1177	if (tg->nr_queued[rw]) {
1178		/*
1179		 * There is already another bio queued in same dir. No
1180		 * need to update dispatch time.
1181		 */
1182		update_disptime = false;
1183		goto queue_bio;
1184
1185	}
 
 
 
 
 
 
 
1186
1187	/* Bio is with-in rate limit of group */
1188	if (tg_may_dispatch(td, tg, bio, NULL)) {
1189		throtl_charge_bio(tg, bio);
1190
1191		/*
1192		 * We need to trim slice even when bios are not being queued
1193		 * otherwise it might happen that a bio is not queued for
1194		 * a long time and slice keeps on extending and trim is not
1195		 * called for a long time. Now if limits are reduced suddenly
1196		 * we take into account all the IO dispatched so far at new
1197		 * low rate and * newly queued IO gets a really long dispatch
1198		 * time.
1199		 *
1200		 * So keep on trimming slice even if bio is not queued.
1201		 */
1202		throtl_trim_slice(td, tg, rw);
1203		goto out;
 
 
 
 
 
 
 
 
 
 
1204	}
1205
1206queue_bio:
1207	throtl_log_tg(td, tg, "[%c] bio. bdisp=%llu sz=%u bps=%llu"
1208			" iodisp=%u iops=%u queued=%d/%d",
1209			rw == READ ? 'R' : 'W',
1210			tg->bytes_disp[rw], bio->bi_size, tg->bps[rw],
1211			tg->io_disp[rw], tg->iops[rw],
1212			tg->nr_queued[READ], tg->nr_queued[WRITE]);
1213
1214	throtl_add_bio_tg(q->td, tg, bio);
1215	*biop = NULL;
1216
1217	if (update_disptime) {
1218		tg_update_disptime(td, tg);
1219		throtl_schedule_next_dispatch(td);
 
 
 
 
 
 
 
1220	}
1221
1222out:
1223	spin_unlock_irq(q->queue_lock);
1224	return 0;
 
 
 
 
 
 
 
 
1225}
1226
1227int blk_throtl_init(struct request_queue *q)
 
 
 
 
 
1228{
1229	struct throtl_data *td;
1230	struct throtl_grp *tg;
1231
1232	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1233	if (!td)
1234		return -ENOMEM;
1235
1236	INIT_HLIST_HEAD(&td->tg_list);
1237	td->tg_service_tree = THROTL_RB_ROOT;
1238	td->limits_changed = false;
1239	INIT_DELAYED_WORK(&td->throtl_work, blk_throtl_work);
1240
1241	/* alloc and Init root group. */
1242	td->queue = q;
1243	tg = throtl_alloc_tg(td);
1244
1245	if (!tg) {
1246		kfree(td);
1247		return -ENOMEM;
 
1248	}
 
1249
1250	td->root_tg = tg;
 
 
 
 
 
 
 
 
 
 
 
 
 
1251
 
1252	rcu_read_lock();
1253	throtl_init_add_tg_lists(td, tg, &blkio_root_cgroup);
1254	rcu_read_unlock();
1255
1256	/* Attach throtl data to request queue */
1257	q->td = td;
1258	return 0;
1259}
 
 
 
 
1260
1261void blk_throtl_exit(struct request_queue *q)
1262{
1263	struct throtl_data *td = q->td;
1264	bool wait = false;
1265
1266	BUG_ON(!td);
 
1267
1268	throtl_shutdown_wq(q);
 
 
 
 
1269
1270	spin_lock_irq(q->queue_lock);
1271	throtl_release_tgs(td);
1272
1273	/* If there are other groups */
1274	if (td->nr_undestroyed_grps > 0)
1275		wait = true;
 
1276
1277	spin_unlock_irq(q->queue_lock);
 
 
1278
1279	/*
1280	 * Wait for tg->blkg->key accessors to exit their grace periods.
1281	 * Do this wait only if there are other undestroyed groups out
1282	 * there (other than root group). This can happen if cgroup deletion
1283	 * path claimed the responsibility of cleaning up a group before
1284	 * queue cleanup code get to the group.
1285	 *
1286	 * Do not call synchronize_rcu() unconditionally as there are drivers
1287	 * which create/delete request queue hundreds of times during scan/boot
1288	 * and synchronize_rcu() can take significant time and slow down boot.
1289	 */
1290	if (wait)
1291		synchronize_rcu();
1292
1293	/*
1294	 * Just being safe to make sure after previous flush if some body did
1295	 * update limits through cgroup and another work got queued, cancel
1296	 * it.
1297	 */
 
 
 
 
 
 
 
 
1298	throtl_shutdown_wq(q);
1299	throtl_td_free(td);
 
1300}
1301
1302static int __init throtl_init(void)
1303{
1304	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1305	if (!kthrotld_workqueue)
1306		panic("Failed to create kthrotld\n");
1307
1308	blkio_policy_register(&blkio_policy_throtl);
1309	return 0;
1310}
1311
1312module_init(throtl_init);
v4.10.11
   1/*
   2 * Interface for controlling IO bandwidth on a request queue
   3 *
   4 * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   5 */
   6
   7#include <linux/module.h>
   8#include <linux/slab.h>
   9#include <linux/blkdev.h>
  10#include <linux/bio.h>
  11#include <linux/blktrace_api.h>
  12#include <linux/blk-cgroup.h>
  13#include "blk.h"
  14
  15/* Max dispatch from a group in 1 round */
  16static int throtl_grp_quantum = 8;
  17
  18/* Total max dispatch from all groups in one round */
  19static int throtl_quantum = 32;
  20
  21/* Throttling is performed over 100ms slice and after that slice is renewed */
  22static unsigned long throtl_slice = HZ/10;	/* 100 ms */
  23
  24static struct blkcg_policy blkcg_policy_throtl;
  25
  26/* A workqueue to queue throttle related work */
  27static struct workqueue_struct *kthrotld_workqueue;
 
 
  28
  29/*
  30 * To implement hierarchical throttling, throtl_grps form a tree and bios
  31 * are dispatched upwards level by level until they reach the top and get
  32 * issued.  When dispatching bios from the children and local group at each
  33 * level, if the bios are dispatched into a single bio_list, there's a risk
  34 * of a local or child group which can queue many bios at once filling up
  35 * the list starving others.
  36 *
  37 * To avoid such starvation, dispatched bios are queued separately
  38 * according to where they came from.  When they are again dispatched to
  39 * the parent, they're popped in round-robin order so that no single source
  40 * hogs the dispatch window.
  41 *
  42 * throtl_qnode is used to keep the queued bios separated by their sources.
  43 * Bios are queued to throtl_qnode which in turn is queued to
  44 * throtl_service_queue and then dispatched in round-robin order.
  45 *
  46 * It's also used to track the reference counts on blkg's.  A qnode always
  47 * belongs to a throtl_grp and gets queued on itself or the parent, so
  48 * incrementing the reference of the associated throtl_grp when a qnode is
  49 * queued and decrementing when dequeued is enough to keep the whole blkg
  50 * tree pinned while bios are in flight.
  51 */
  52struct throtl_qnode {
  53	struct list_head	node;		/* service_queue->queued[] */
  54	struct bio_list		bios;		/* queued bios */
  55	struct throtl_grp	*tg;		/* tg this qnode belongs to */
  56};
  57
  58struct throtl_service_queue {
  59	struct throtl_service_queue *parent_sq;	/* the parent service_queue */
  60
  61	/*
  62	 * Bios queued directly to this service_queue or dispatched from
  63	 * children throtl_grp's.
  64	 */
  65	struct list_head	queued[2];	/* throtl_qnode [READ/WRITE] */
  66	unsigned int		nr_queued[2];	/* number of queued bios */
  67
  68	/*
  69	 * RB tree of active children throtl_grp's, which are sorted by
  70	 * their ->disptime.
  71	 */
  72	struct rb_root		pending_tree;	/* RB tree of active tgs */
  73	struct rb_node		*first_pending;	/* first node in the tree */
  74	unsigned int		nr_pending;	/* # queued in the tree */
  75	unsigned long		first_pending_disptime;	/* disptime of the first tg */
  76	struct timer_list	pending_timer;	/* fires on first_pending_disptime */
  77};
  78
  79enum tg_state_flags {
  80	THROTL_TG_PENDING	= 1 << 0,	/* on parent's pending tree */
  81	THROTL_TG_WAS_EMPTY	= 1 << 1,	/* bio_lists[] became non-empty */
  82};
  83
  84#define rb_entry_tg(node)	rb_entry((node), struct throtl_grp, rb_node)
  85
  86struct throtl_grp {
  87	/* must be the first member */
  88	struct blkg_policy_data pd;
  89
  90	/* active throtl group service_queue member */
  91	struct rb_node rb_node;
  92
  93	/* throtl_data this group belongs to */
  94	struct throtl_data *td;
  95
  96	/* this group's service queue */
  97	struct throtl_service_queue service_queue;
  98
  99	/*
 100	 * qnode_on_self is used when bios are directly queued to this
 101	 * throtl_grp so that local bios compete fairly with bios
 102	 * dispatched from children.  qnode_on_parent is used when bios are
 103	 * dispatched from this throtl_grp into its parent and will compete
 104	 * with the sibling qnode_on_parents and the parent's
 105	 * qnode_on_self.
 106	 */
 107	struct throtl_qnode qnode_on_self[2];
 108	struct throtl_qnode qnode_on_parent[2];
 109
 110	/*
 111	 * Dispatch time in jiffies. This is the estimated time when group
 112	 * will unthrottle and is ready to dispatch more bio. It is used as
 113	 * key to sort active groups in service tree.
 114	 */
 115	unsigned long disptime;
 116
 
 
 117	unsigned int flags;
 118
 119	/* are there any throtl rules between this group and td? */
 120	bool has_rules[2];
 
 
 
 121
 122	/* bytes per second rate limits */
 123	uint64_t bps[2];
 124
 125	/* IOPS limits */
 126	unsigned int iops[2];
 127
 128	/* Number of bytes disptached in current slice */
 129	uint64_t bytes_disp[2];
 130	/* Number of bio's dispatched in current slice */
 131	unsigned int io_disp[2];
 132
 133	/* When did we start a new slice */
 134	unsigned long slice_start[2];
 135	unsigned long slice_end[2];
 
 
 
 
 
 136};
 137
 138struct throtl_data
 139{
 
 
 
 140	/* service tree for active throtl groups */
 141	struct throtl_service_queue service_queue;
 142
 
 143	struct request_queue *queue;
 144
 145	/* Total Number of queued bios on READ and WRITE lists */
 146	unsigned int nr_queued[2];
 147
 
 
 
 
 
 148	/* Work for dispatching throttled bios */
 149	struct work_struct dispatch_work;
 
 
 
 
 
 
 150};
 151
 152static void throtl_pending_timer_fn(unsigned long arg);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 153
 154static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 155{
 156	return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 
 
 
 157}
 158
 159static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 160{
 161	return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 162}
 163
 164static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 165{
 166	return pd_to_blkg(&tg->pd);
 
 167}
 168
 169/**
 170 * sq_to_tg - return the throl_grp the specified service queue belongs to
 171 * @sq: the throtl_service_queue of interest
 172 *
 173 * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 174 * embedded in throtl_data, %NULL is returned.
 175 */
 176static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 
 
 177{
 178	if (sq && sq->parent_sq)
 179		return container_of(sq, struct throtl_grp, service_queue);
 180	else
 181		return NULL;
 
 
 
 
 
 
 
 
 
 
 182}
 183
 184/**
 185 * sq_to_td - return throtl_data the specified service queue belongs to
 186 * @sq: the throtl_service_queue of interest
 187 *
 188 * A service_queue can be embeded in either a throtl_grp or throtl_data.
 189 * Determine the associated throtl_data accordingly and return it.
 190 */
 191static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 192{
 193	struct throtl_grp *tg = sq_to_tg(sq);
 
 
 
 
 194
 195	if (tg)
 196		return tg->td;
 197	else
 198		return container_of(sq, struct throtl_data, service_queue);
 
 
 
 
 
 
 
 199}
 200
 201/**
 202 * throtl_log - log debug message via blktrace
 203 * @sq: the service_queue being reported
 204 * @fmt: printf format string
 205 * @args: printf args
 206 *
 207 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 208 * throtl_grp; otherwise, just "throtl".
 209 */
 210#define throtl_log(sq, fmt, args...)	do {				\
 211	struct throtl_grp *__tg = sq_to_tg((sq));			\
 212	struct throtl_data *__td = sq_to_td((sq));			\
 213									\
 214	(void)__td;							\
 215	if (likely(!blk_trace_note_message_enabled(__td->queue)))	\
 216		break;							\
 217	if ((__tg)) {							\
 218		char __pbuf[128];					\
 219									\
 220		blkg_path(tg_to_blkg(__tg), __pbuf, sizeof(__pbuf));	\
 221		blk_add_trace_msg(__td->queue, "throtl %s " fmt, __pbuf, ##args); \
 222	} else {							\
 223		blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);	\
 224	}								\
 225} while (0)
 226
 227static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 228{
 229	INIT_LIST_HEAD(&qn->node);
 230	bio_list_init(&qn->bios);
 231	qn->tg = tg;
 232}
 233
 234/**
 235 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 236 * @bio: bio being added
 237 * @qn: qnode to add bio to
 238 * @queued: the service_queue->queued[] list @qn belongs to
 239 *
 240 * Add @bio to @qn and put @qn on @queued if it's not already on.
 241 * @qn->tg's reference count is bumped when @qn is activated.  See the
 242 * comment on top of throtl_qnode definition for details.
 243 */
 244static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 245				 struct list_head *queued)
 246{
 247	bio_list_add(&qn->bios, bio);
 248	if (list_empty(&qn->node)) {
 249		list_add_tail(&qn->node, queued);
 250		blkg_get(tg_to_blkg(qn->tg));
 251	}
 252}
 253
 254/**
 255 * throtl_peek_queued - peek the first bio on a qnode list
 256 * @queued: the qnode list to peek
 257 */
 258static struct bio *throtl_peek_queued(struct list_head *queued)
 259{
 260	struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 261	struct bio *bio;
 262
 263	if (list_empty(queued))
 264		return NULL;
 265
 266	bio = bio_list_peek(&qn->bios);
 267	WARN_ON_ONCE(!bio);
 268	return bio;
 
 
 
 
 
 
 269}
 270
 271/**
 272 * throtl_pop_queued - pop the first bio form a qnode list
 273 * @queued: the qnode list to pop a bio from
 274 * @tg_to_put: optional out argument for throtl_grp to put
 275 *
 276 * Pop the first bio from the qnode list @queued.  After popping, the first
 277 * qnode is removed from @queued if empty or moved to the end of @queued so
 278 * that the popping order is round-robin.
 279 *
 280 * When the first qnode is removed, its associated throtl_grp should be put
 281 * too.  If @tg_to_put is NULL, this function automatically puts it;
 282 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 283 * responsible for putting it.
 284 */
 285static struct bio *throtl_pop_queued(struct list_head *queued,
 286				     struct throtl_grp **tg_to_put)
 287{
 288	struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 289	struct bio *bio;
 290
 291	if (list_empty(queued))
 292		return NULL;
 
 
 293
 294	bio = bio_list_pop(&qn->bios);
 295	WARN_ON_ONCE(!bio);
 
 
 296
 297	if (bio_list_empty(&qn->bios)) {
 298		list_del_init(&qn->node);
 299		if (tg_to_put)
 300			*tg_to_put = qn->tg;
 301		else
 302			blkg_put(tg_to_blkg(qn->tg));
 303	} else {
 304		list_move_tail(&qn->node, queued);
 305	}
 306
 307	return bio;
 308}
 
 
 309
 310/* init a service_queue, assumes the caller zeroed it */
 311static void throtl_service_queue_init(struct throtl_service_queue *sq)
 312{
 313	INIT_LIST_HEAD(&sq->queued[0]);
 314	INIT_LIST_HEAD(&sq->queued[1]);
 315	sq->pending_tree = RB_ROOT;
 316	setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
 317		    (unsigned long)sq);
 318}
 319
 320static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
 
 321{
 322	struct throtl_grp *tg;
 323	int rw;
 324
 325	tg = kzalloc_node(sizeof(*tg), gfp, node);
 326	if (!tg)
 327		return NULL;
 328
 329	throtl_service_queue_init(&tg->service_queue);
 330
 331	for (rw = READ; rw <= WRITE; rw++) {
 332		throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 333		throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 334	}
 335
 336	RB_CLEAR_NODE(&tg->rb_node);
 337	tg->bps[READ] = -1;
 338	tg->bps[WRITE] = -1;
 339	tg->iops[READ] = -1;
 340	tg->iops[WRITE] = -1;
 341
 342	return &tg->pd;
 343}
 344
 345static void throtl_pd_init(struct blkg_policy_data *pd)
 
 346{
 347	struct throtl_grp *tg = pd_to_tg(pd);
 348	struct blkcg_gq *blkg = tg_to_blkg(tg);
 349	struct throtl_data *td = blkg->q->td;
 350	struct throtl_service_queue *sq = &tg->service_queue;
 351
 352	/*
 353	 * If on the default hierarchy, we switch to properly hierarchical
 354	 * behavior where limits on a given throtl_grp are applied to the
 355	 * whole subtree rather than just the group itself.  e.g. If 16M
 356	 * read_bps limit is set on the root group, the whole system can't
 357	 * exceed 16M for the device.
 358	 *
 359	 * If not on the default hierarchy, the broken flat hierarchy
 360	 * behavior is retained where all throtl_grps are treated as if
 361	 * they're all separate root groups right below throtl_data.
 362	 * Limits of a group don't interact with limits of other groups
 363	 * regardless of the position of the group in the hierarchy.
 364	 */
 365	sq->parent_sq = &td->service_queue;
 366	if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 367		sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 368	tg->td = td;
 369}
 370
 371/*
 372 * Set has_rules[] if @tg or any of its parents have limits configured.
 373 * This doesn't require walking up to the top of the hierarchy as the
 374 * parent's has_rules[] is guaranteed to be correct.
 375 */
 376static void tg_update_has_rules(struct throtl_grp *tg)
 377{
 378	struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 379	int rw;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 380
 381	for (rw = READ; rw <= WRITE; rw++)
 382		tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 383				    (tg->bps[rw] != -1 || tg->iops[rw] != -1);
 384}
 
 
 
 
 
 
 
 
 385
 386static void throtl_pd_online(struct blkg_policy_data *pd)
 387{
 388	/*
 389	 * We don't want new groups to escape the limits of its ancestors.
 390	 * Update has_rules[] after a new group is brought online.
 391	 */
 392	tg_update_has_rules(pd_to_tg(pd));
 393}
 
 
 
 
 
 394
 395static void throtl_pd_free(struct blkg_policy_data *pd)
 396{
 397	struct throtl_grp *tg = pd_to_tg(pd);
 
 
 398
 399	del_timer_sync(&tg->service_queue.pending_timer);
 400	kfree(tg);
 
 401}
 402
 403static struct throtl_grp *
 404throtl_rb_first(struct throtl_service_queue *parent_sq)
 405{
 406	/* Service tree is empty */
 407	if (!parent_sq->nr_pending)
 408		return NULL;
 409
 410	if (!parent_sq->first_pending)
 411		parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
 412
 413	if (parent_sq->first_pending)
 414		return rb_entry_tg(parent_sq->first_pending);
 415
 416	return NULL;
 417}
 418
 419static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 420{
 421	rb_erase(n, root);
 422	RB_CLEAR_NODE(n);
 423}
 424
 425static void throtl_rb_erase(struct rb_node *n,
 426			    struct throtl_service_queue *parent_sq)
 427{
 428	if (parent_sq->first_pending == n)
 429		parent_sq->first_pending = NULL;
 430	rb_erase_init(n, &parent_sq->pending_tree);
 431	--parent_sq->nr_pending;
 432}
 433
 434static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 435{
 436	struct throtl_grp *tg;
 437
 438	tg = throtl_rb_first(parent_sq);
 439	if (!tg)
 440		return;
 441
 442	parent_sq->first_pending_disptime = tg->disptime;
 443}
 444
 445static void tg_service_queue_add(struct throtl_grp *tg)
 
 446{
 447	struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 448	struct rb_node **node = &parent_sq->pending_tree.rb_node;
 449	struct rb_node *parent = NULL;
 450	struct throtl_grp *__tg;
 451	unsigned long key = tg->disptime;
 452	int left = 1;
 453
 454	while (*node != NULL) {
 455		parent = *node;
 456		__tg = rb_entry_tg(parent);
 457
 458		if (time_before(key, __tg->disptime))
 459			node = &parent->rb_left;
 460		else {
 461			node = &parent->rb_right;
 462			left = 0;
 463		}
 464	}
 465
 466	if (left)
 467		parent_sq->first_pending = &tg->rb_node;
 468
 469	rb_link_node(&tg->rb_node, parent, node);
 470	rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
 471}
 472
 473static void __throtl_enqueue_tg(struct throtl_grp *tg)
 474{
 475	tg_service_queue_add(tg);
 476	tg->flags |= THROTL_TG_PENDING;
 477	tg->service_queue.parent_sq->nr_pending++;
 478}
 479
 480static void throtl_enqueue_tg(struct throtl_grp *tg)
 481{
 482	if (!(tg->flags & THROTL_TG_PENDING))
 483		__throtl_enqueue_tg(tg);
 484}
 485
 486static void __throtl_dequeue_tg(struct throtl_grp *tg)
 487{
 488	throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 489	tg->flags &= ~THROTL_TG_PENDING;
 490}
 491
 492static void throtl_dequeue_tg(struct throtl_grp *tg)
 493{
 494	if (tg->flags & THROTL_TG_PENDING)
 495		__throtl_dequeue_tg(tg);
 496}
 497
 498/* Call with queue lock held */
 499static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 500					  unsigned long expires)
 501{
 502	mod_timer(&sq->pending_timer, expires);
 503	throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 504		   expires - jiffies, jiffies);
 505}
 506
 507/**
 508 * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 509 * @sq: the service_queue to schedule dispatch for
 510 * @force: force scheduling
 511 *
 512 * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 513 * dispatch time of the first pending child.  Returns %true if either timer
 514 * is armed or there's no pending child left.  %false if the current
 515 * dispatch window is still open and the caller should continue
 516 * dispatching.
 517 *
 518 * If @force is %true, the dispatch timer is always scheduled and this
 519 * function is guaranteed to return %true.  This is to be used when the
 520 * caller can't dispatch itself and needs to invoke pending_timer
 521 * unconditionally.  Note that forced scheduling is likely to induce short
 522 * delay before dispatch starts even if @sq->first_pending_disptime is not
 523 * in the future and thus shouldn't be used in hot paths.
 524 */
 525static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 526					  bool force)
 527{
 528	/* any pending children left? */
 529	if (!sq->nr_pending)
 530		return true;
 531
 532	update_min_dispatch_time(sq);
 
 
 
 
 533
 534	/* is the next dispatch time in the future? */
 535	if (force || time_after(sq->first_pending_disptime, jiffies)) {
 536		throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 537		return true;
 538	}
 539
 540	/* tell the caller to continue dispatching */
 541	return false;
 542}
 543
 544static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 545		bool rw, unsigned long start)
 546{
 547	tg->bytes_disp[rw] = 0;
 548	tg->io_disp[rw] = 0;
 549
 550	/*
 551	 * Previous slice has expired. We must have trimmed it after last
 552	 * bio dispatch. That means since start of last slice, we never used
 553	 * that bandwidth. Do try to make use of that bandwidth while giving
 554	 * credit.
 555	 */
 556	if (time_after_eq(start, tg->slice_start[rw]))
 557		tg->slice_start[rw] = start;
 558
 559	tg->slice_end[rw] = jiffies + throtl_slice;
 560	throtl_log(&tg->service_queue,
 561		   "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 562		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 563		   tg->slice_end[rw], jiffies);
 564}
 565
 566static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 
 567{
 568	tg->bytes_disp[rw] = 0;
 569	tg->io_disp[rw] = 0;
 570	tg->slice_start[rw] = jiffies;
 571	tg->slice_end[rw] = jiffies + throtl_slice;
 572	throtl_log(&tg->service_queue,
 573		   "[%c] new slice start=%lu end=%lu jiffies=%lu",
 574		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 575		   tg->slice_end[rw], jiffies);
 576}
 577
 578static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 579					unsigned long jiffy_end)
 580{
 581	tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 582}
 583
 584static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 585				       unsigned long jiffy_end)
 586{
 587	tg->slice_end[rw] = roundup(jiffy_end, throtl_slice);
 588	throtl_log(&tg->service_queue,
 589		   "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 590		   rw == READ ? 'R' : 'W', tg->slice_start[rw],
 591		   tg->slice_end[rw], jiffies);
 592}
 593
 594/* Determine if previously allocated or extended slice is complete or not */
 595static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 
 596{
 597	if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 598		return false;
 599
 600	return 1;
 601}
 602
 603/* Trim the used slices and adjust slice start accordingly */
 604static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 
 605{
 606	unsigned long nr_slices, time_elapsed, io_trim;
 607	u64 bytes_trim, tmp;
 608
 609	BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 610
 611	/*
 612	 * If bps are unlimited (-1), then time slice don't get
 613	 * renewed. Don't try to trim the slice if slice is used. A new
 614	 * slice will start when appropriate.
 615	 */
 616	if (throtl_slice_used(tg, rw))
 617		return;
 618
 619	/*
 620	 * A bio has been dispatched. Also adjust slice_end. It might happen
 621	 * that initially cgroup limit was very low resulting in high
 622	 * slice_end, but later limit was bumped up and bio was dispached
 623	 * sooner, then we need to reduce slice_end. A high bogus slice_end
 624	 * is bad because it does not allow new slice to start.
 625	 */
 626
 627	throtl_set_slice_end(tg, rw, jiffies + throtl_slice);
 628
 629	time_elapsed = jiffies - tg->slice_start[rw];
 630
 631	nr_slices = time_elapsed / throtl_slice;
 632
 633	if (!nr_slices)
 634		return;
 635	tmp = tg->bps[rw] * throtl_slice * nr_slices;
 636	do_div(tmp, HZ);
 637	bytes_trim = tmp;
 638
 639	io_trim = (tg->iops[rw] * throtl_slice * nr_slices)/HZ;
 640
 641	if (!bytes_trim && !io_trim)
 642		return;
 643
 644	if (tg->bytes_disp[rw] >= bytes_trim)
 645		tg->bytes_disp[rw] -= bytes_trim;
 646	else
 647		tg->bytes_disp[rw] = 0;
 648
 649	if (tg->io_disp[rw] >= io_trim)
 650		tg->io_disp[rw] -= io_trim;
 651	else
 652		tg->io_disp[rw] = 0;
 653
 654	tg->slice_start[rw] += nr_slices * throtl_slice;
 655
 656	throtl_log(&tg->service_queue,
 657		   "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 658		   rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 659		   tg->slice_start[rw], tg->slice_end[rw], jiffies);
 660}
 661
 662static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 663				  unsigned long *wait)
 664{
 665	bool rw = bio_data_dir(bio);
 666	unsigned int io_allowed;
 667	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 668	u64 tmp;
 669
 670	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 671
 672	/* Slice has just started. Consider one slice interval */
 673	if (!jiffy_elapsed)
 674		jiffy_elapsed_rnd = throtl_slice;
 675
 676	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 677
 678	/*
 679	 * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 680	 * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 681	 * will allow dispatch after 1 second and after that slice should
 682	 * have been trimmed.
 683	 */
 684
 685	tmp = (u64)tg->iops[rw] * jiffy_elapsed_rnd;
 686	do_div(tmp, HZ);
 687
 688	if (tmp > UINT_MAX)
 689		io_allowed = UINT_MAX;
 690	else
 691		io_allowed = tmp;
 692
 693	if (tg->io_disp[rw] + 1 <= io_allowed) {
 694		if (wait)
 695			*wait = 0;
 696		return true;
 697	}
 698
 699	/* Calc approx time to dispatch */
 700	jiffy_wait = ((tg->io_disp[rw] + 1) * HZ)/tg->iops[rw] + 1;
 701
 702	if (jiffy_wait > jiffy_elapsed)
 703		jiffy_wait = jiffy_wait - jiffy_elapsed;
 704	else
 705		jiffy_wait = 1;
 706
 707	if (wait)
 708		*wait = jiffy_wait;
 709	return 0;
 710}
 711
 712static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 713				 unsigned long *wait)
 714{
 715	bool rw = bio_data_dir(bio);
 716	u64 bytes_allowed, extra_bytes, tmp;
 717	unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 718
 719	jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 720
 721	/* Slice has just started. Consider one slice interval */
 722	if (!jiffy_elapsed)
 723		jiffy_elapsed_rnd = throtl_slice;
 724
 725	jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, throtl_slice);
 726
 727	tmp = tg->bps[rw] * jiffy_elapsed_rnd;
 728	do_div(tmp, HZ);
 729	bytes_allowed = tmp;
 730
 731	if (tg->bytes_disp[rw] + bio->bi_iter.bi_size <= bytes_allowed) {
 732		if (wait)
 733			*wait = 0;
 734		return true;
 735	}
 736
 737	/* Calc approx time to dispatch */
 738	extra_bytes = tg->bytes_disp[rw] + bio->bi_iter.bi_size - bytes_allowed;
 739	jiffy_wait = div64_u64(extra_bytes * HZ, tg->bps[rw]);
 740
 741	if (!jiffy_wait)
 742		jiffy_wait = 1;
 743
 744	/*
 745	 * This wait time is without taking into consideration the rounding
 746	 * up we did. Add that time also.
 747	 */
 748	jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 749	if (wait)
 750		*wait = jiffy_wait;
 751	return 0;
 752}
 753
 
 
 
 
 
 
 754/*
 755 * Returns whether one can dispatch a bio or not. Also returns approx number
 756 * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 757 */
 758static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 759			    unsigned long *wait)
 760{
 761	bool rw = bio_data_dir(bio);
 762	unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 763
 764	/*
 765 	 * Currently whole state machine of group depends on first bio
 766	 * queued in the group bio list. So one should not be calling
 767	 * this function with a different bio if there are other bios
 768	 * queued.
 769	 */
 770	BUG_ON(tg->service_queue.nr_queued[rw] &&
 771	       bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 772
 773	/* If tg->bps = -1, then BW is unlimited */
 774	if (tg->bps[rw] == -1 && tg->iops[rw] == -1) {
 775		if (wait)
 776			*wait = 0;
 777		return true;
 778	}
 779
 780	/*
 781	 * If previous slice expired, start a new one otherwise renew/extend
 782	 * existing slice to make sure it is at least throtl_slice interval
 783	 * long since now. New slice is started only for empty throttle group.
 784	 * If there is queued bio, that means there should be an active
 785	 * slice and it should be extended instead.
 786	 */
 787	if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 788		throtl_start_new_slice(tg, rw);
 789	else {
 790		if (time_before(tg->slice_end[rw], jiffies + throtl_slice))
 791			throtl_extend_slice(tg, rw, jiffies + throtl_slice);
 792	}
 793
 794	if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
 795	    tg_with_in_iops_limit(tg, bio, &iops_wait)) {
 796		if (wait)
 797			*wait = 0;
 798		return 1;
 799	}
 800
 801	max_wait = max(bps_wait, iops_wait);
 802
 803	if (wait)
 804		*wait = max_wait;
 805
 806	if (time_before(tg->slice_end[rw], jiffies + max_wait))
 807		throtl_extend_slice(tg, rw, jiffies + max_wait);
 808
 809	return 0;
 810}
 811
 812static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 813{
 814	bool rw = bio_data_dir(bio);
 
 815
 816	/* Charge the bio to the group */
 817	tg->bytes_disp[rw] += bio->bi_iter.bi_size;
 818	tg->io_disp[rw]++;
 819
 820	/*
 821	 * BIO_THROTTLED is used to prevent the same bio to be throttled
 822	 * more than once as a throttled bio will go through blk-throtl the
 823	 * second time when it eventually gets issued.  Set it when a bio
 824	 * is being charged to a tg.
 825	 */
 826	if (!bio_flagged(bio, BIO_THROTTLED))
 827		bio_set_flag(bio, BIO_THROTTLED);
 828}
 829
 830/**
 831 * throtl_add_bio_tg - add a bio to the specified throtl_grp
 832 * @bio: bio to add
 833 * @qn: qnode to use
 834 * @tg: the target throtl_grp
 835 *
 836 * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 837 * tg->qnode_on_self[] is used.
 838 */
 839static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 840			      struct throtl_grp *tg)
 841{
 842	struct throtl_service_queue *sq = &tg->service_queue;
 843	bool rw = bio_data_dir(bio);
 844
 845	if (!qn)
 846		qn = &tg->qnode_on_self[rw];
 847
 848	/*
 849	 * If @tg doesn't currently have any bios queued in the same
 850	 * direction, queueing @bio can change when @tg should be
 851	 * dispatched.  Mark that @tg was empty.  This is automatically
 852	 * cleaered on the next tg_update_disptime().
 853	 */
 854	if (!sq->nr_queued[rw])
 855		tg->flags |= THROTL_TG_WAS_EMPTY;
 856
 857	throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
 858
 859	sq->nr_queued[rw]++;
 860	throtl_enqueue_tg(tg);
 861}
 862
 863static void tg_update_disptime(struct throtl_grp *tg)
 864{
 865	struct throtl_service_queue *sq = &tg->service_queue;
 866	unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
 867	struct bio *bio;
 868
 869	if ((bio = throtl_peek_queued(&sq->queued[READ])))
 870		tg_may_dispatch(tg, bio, &read_wait);
 871
 872	if ((bio = throtl_peek_queued(&sq->queued[WRITE])))
 873		tg_may_dispatch(tg, bio, &write_wait);
 874
 875	min_wait = min(read_wait, write_wait);
 876	disptime = jiffies + min_wait;
 877
 878	/* Update dispatch time */
 879	throtl_dequeue_tg(tg);
 880	tg->disptime = disptime;
 881	throtl_enqueue_tg(tg);
 882
 883	/* see throtl_add_bio_tg() */
 884	tg->flags &= ~THROTL_TG_WAS_EMPTY;
 885}
 886
 887static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
 888					struct throtl_grp *parent_tg, bool rw)
 889{
 890	if (throtl_slice_used(parent_tg, rw)) {
 891		throtl_start_new_slice_with_credit(parent_tg, rw,
 892				child_tg->slice_start[rw]);
 893	}
 894
 895}
 
 
 
 896
 897static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
 898{
 899	struct throtl_service_queue *sq = &tg->service_queue;
 900	struct throtl_service_queue *parent_sq = sq->parent_sq;
 901	struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
 902	struct throtl_grp *tg_to_put = NULL;
 903	struct bio *bio;
 904
 905	/*
 906	 * @bio is being transferred from @tg to @parent_sq.  Popping a bio
 907	 * from @tg may put its reference and @parent_sq might end up
 908	 * getting released prematurely.  Remember the tg to put and put it
 909	 * after @bio is transferred to @parent_sq.
 910	 */
 911	bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
 912	sq->nr_queued[rw]--;
 913
 914	throtl_charge_bio(tg, bio);
 
 
 915
 916	/*
 917	 * If our parent is another tg, we just need to transfer @bio to
 918	 * the parent using throtl_add_bio_tg().  If our parent is
 919	 * @td->service_queue, @bio is ready to be issued.  Put it on its
 920	 * bio_lists[] and decrease total number queued.  The caller is
 921	 * responsible for issuing these bios.
 922	 */
 923	if (parent_tg) {
 924		throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
 925		start_parent_slice_with_credit(tg, parent_tg, rw);
 926	} else {
 927		throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
 928				     &parent_sq->queued[rw]);
 929		BUG_ON(tg->td->nr_queued[rw] <= 0);
 930		tg->td->nr_queued[rw]--;
 931	}
 932
 933	throtl_trim_slice(tg, rw);
 934
 935	if (tg_to_put)
 936		blkg_put(tg_to_blkg(tg_to_put));
 937}
 938
 939static int throtl_dispatch_tg(struct throtl_grp *tg)
 
 940{
 941	struct throtl_service_queue *sq = &tg->service_queue;
 942	unsigned int nr_reads = 0, nr_writes = 0;
 943	unsigned int max_nr_reads = throtl_grp_quantum*3/4;
 944	unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
 945	struct bio *bio;
 946
 947	/* Try to dispatch 75% READS and 25% WRITES */
 948
 949	while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
 950	       tg_may_dispatch(tg, bio, NULL)) {
 951
 952		tg_dispatch_one_bio(tg, bio_data_dir(bio));
 953		nr_reads++;
 954
 955		if (nr_reads >= max_nr_reads)
 956			break;
 957	}
 958
 959	while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
 960	       tg_may_dispatch(tg, bio, NULL)) {
 961
 962		tg_dispatch_one_bio(tg, bio_data_dir(bio));
 963		nr_writes++;
 964
 965		if (nr_writes >= max_nr_writes)
 966			break;
 967	}
 968
 969	return nr_reads + nr_writes;
 970}
 971
 972static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
 973{
 974	unsigned int nr_disp = 0;
 
 
 975
 976	while (1) {
 977		struct throtl_grp *tg = throtl_rb_first(parent_sq);
 978		struct throtl_service_queue *sq = &tg->service_queue;
 979
 980		if (!tg)
 981			break;
 982
 983		if (time_before(jiffies, tg->disptime))
 984			break;
 985
 986		throtl_dequeue_tg(tg);
 987
 988		nr_disp += throtl_dispatch_tg(tg);
 989
 990		if (sq->nr_queued[0] || sq->nr_queued[1])
 991			tg_update_disptime(tg);
 
 
 992
 993		if (nr_disp >= throtl_quantum)
 994			break;
 995	}
 996
 997	return nr_disp;
 998}
 999
1000/**
1001 * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1002 * @arg: the throtl_service_queue being serviced
1003 *
1004 * This timer is armed when a child throtl_grp with active bio's become
1005 * pending and queued on the service_queue's pending_tree and expires when
1006 * the first child throtl_grp should be dispatched.  This function
1007 * dispatches bio's from the children throtl_grps to the parent
1008 * service_queue.
1009 *
1010 * If the parent's parent is another throtl_grp, dispatching is propagated
1011 * by either arming its pending_timer or repeating dispatch directly.  If
1012 * the top-level service_tree is reached, throtl_data->dispatch_work is
1013 * kicked so that the ready bio's are issued.
1014 */
1015static void throtl_pending_timer_fn(unsigned long arg)
1016{
1017	struct throtl_service_queue *sq = (void *)arg;
1018	struct throtl_grp *tg = sq_to_tg(sq);
1019	struct throtl_data *td = sq_to_td(sq);
1020	struct request_queue *q = td->queue;
1021	struct throtl_service_queue *parent_sq;
1022	bool dispatched;
1023	int ret;
 
 
 
 
 
 
 
 
 
1024
1025	spin_lock_irq(q->queue_lock);
1026again:
1027	parent_sq = sq->parent_sq;
1028	dispatched = false;
1029
1030	while (true) {
1031		throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1032			   sq->nr_queued[READ] + sq->nr_queued[WRITE],
1033			   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1034
1035		ret = throtl_select_dispatch(sq);
1036		if (ret) {
1037			throtl_log(sq, "bios disp=%u", ret);
1038			dispatched = true;
1039		}
1040
1041		if (throtl_schedule_next_dispatch(sq, false))
1042			break;
 
 
 
 
 
 
1043
1044		/* this dispatch windows is still open, relax and repeat */
1045		spin_unlock_irq(q->queue_lock);
1046		cpu_relax();
1047		spin_lock_irq(q->queue_lock);
1048	}
1049
1050	if (!dispatched)
1051		goto out_unlock;
1052
1053	if (parent_sq) {
1054		/* @parent_sq is another throl_grp, propagate dispatch */
1055		if (tg->flags & THROTL_TG_WAS_EMPTY) {
1056			tg_update_disptime(tg);
1057			if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1058				/* window is already open, repeat dispatching */
1059				sq = parent_sq;
1060				tg = sq_to_tg(sq);
1061				goto again;
1062			}
1063		}
1064	} else {
1065		/* reached the top-level, queue issueing */
1066		queue_work(kthrotld_workqueue, &td->dispatch_work);
1067	}
1068out_unlock:
1069	spin_unlock_irq(q->queue_lock);
1070}
1071
1072/**
1073 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1074 * @work: work item being executed
1075 *
1076 * This function is queued for execution when bio's reach the bio_lists[]
1077 * of throtl_data->service_queue.  Those bio's are ready and issued by this
1078 * function.
1079 */
1080static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1081{
1082	struct throtl_data *td = container_of(work, struct throtl_data,
1083					      dispatch_work);
1084	struct throtl_service_queue *td_sq = &td->service_queue;
1085	struct request_queue *q = td->queue;
1086	struct bio_list bio_list_on_stack;
1087	struct bio *bio;
1088	struct blk_plug plug;
1089	int rw;
 
 
 
 
 
 
1090
1091	bio_list_init(&bio_list_on_stack);
1092
1093	spin_lock_irq(q->queue_lock);
1094	for (rw = READ; rw <= WRITE; rw++)
1095		while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1096			bio_list_add(&bio_list_on_stack, bio);
 
 
 
 
 
 
 
1097	spin_unlock_irq(q->queue_lock);
1098
1099	if (!bio_list_empty(&bio_list_on_stack)) {
 
 
 
 
1100		blk_start_plug(&plug);
1101		while((bio = bio_list_pop(&bio_list_on_stack)))
1102			generic_make_request(bio);
1103		blk_finish_plug(&plug);
1104	}
 
1105}
1106
1107static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1108			      int off)
1109{
1110	struct throtl_grp *tg = pd_to_tg(pd);
1111	u64 v = *(u64 *)((void *)tg + off);
 
1112
1113	if (v == -1)
1114		return 0;
1115	return __blkg_prfill_u64(sf, pd, v);
1116}
1117
1118static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1119			       int off)
 
1120{
1121	struct throtl_grp *tg = pd_to_tg(pd);
1122	unsigned int v = *(unsigned int *)((void *)tg + off);
1123
1124	if (v == -1)
1125		return 0;
1126	return __blkg_prfill_u64(sf, pd, v);
1127}
1128
1129static int tg_print_conf_u64(struct seq_file *sf, void *v)
1130{
1131	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1132			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1133	return 0;
1134}
1135
1136static int tg_print_conf_uint(struct seq_file *sf, void *v)
1137{
1138	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1139			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1140	return 0;
1141}
1142
1143static void tg_conf_updated(struct throtl_grp *tg)
 
1144{
1145	struct throtl_service_queue *sq = &tg->service_queue;
1146	struct cgroup_subsys_state *pos_css;
1147	struct blkcg_gq *blkg;
1148
1149	throtl_log(&tg->service_queue,
1150		   "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1151		   tg->bps[READ], tg->bps[WRITE],
1152		   tg->iops[READ], tg->iops[WRITE]);
1153
1154	/*
1155	 * Update has_rules[] flags for the updated tg's subtree.  A tg is
1156	 * considered to have rules if either the tg itself or any of its
1157	 * ancestors has rules.  This identifies groups without any
1158	 * restrictions in the whole hierarchy and allows them to bypass
1159	 * blk-throttle.
1160	 */
1161	blkg_for_each_descendant_pre(blkg, pos_css, tg_to_blkg(tg))
1162		tg_update_has_rules(blkg_to_tg(blkg));
 
1163
1164	/*
1165	 * We're already holding queue_lock and know @tg is valid.  Let's
1166	 * apply the new config directly.
1167	 *
1168	 * Restart the slices for both READ and WRITES. It might happen
1169	 * that a group's limit are dropped suddenly and we don't want to
1170	 * account recently dispatched IO with new low rate.
1171	 */
1172	throtl_start_new_slice(tg, 0);
1173	throtl_start_new_slice(tg, 1);
1174
1175	if (tg->flags & THROTL_TG_PENDING) {
1176		tg_update_disptime(tg);
1177		throtl_schedule_next_dispatch(sq->parent_sq, true);
 
 
 
 
 
1178	}
1179}
1180
1181static ssize_t tg_set_conf(struct kernfs_open_file *of,
1182			   char *buf, size_t nbytes, loff_t off, bool is_u64)
1183{
1184	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1185	struct blkg_conf_ctx ctx;
1186	struct throtl_grp *tg;
1187	int ret;
1188	u64 v;
1189
1190	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1191	if (ret)
1192		return ret;
1193
1194	ret = -EINVAL;
1195	if (sscanf(ctx.body, "%llu", &v) != 1)
1196		goto out_finish;
1197	if (!v)
1198		v = -1;
1199
1200	tg = blkg_to_tg(ctx.blkg);
 
 
 
 
 
 
 
1201
1202	if (is_u64)
1203		*(u64 *)((void *)tg + of_cft(of)->private) = v;
1204	else
1205		*(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1206
1207	tg_conf_updated(tg);
1208	ret = 0;
1209out_finish:
1210	blkg_conf_finish(&ctx);
1211	return ret ?: nbytes;
1212}
1213
1214static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1215			       char *buf, size_t nbytes, loff_t off)
1216{
1217	return tg_set_conf(of, buf, nbytes, off, true);
 
 
 
1218}
1219
1220static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1221				char *buf, size_t nbytes, loff_t off)
 
 
 
 
 
 
 
 
 
1222{
1223	return tg_set_conf(of, buf, nbytes, off, false);
 
 
 
 
1224}
1225
1226static struct cftype throtl_legacy_files[] = {
1227	{
1228		.name = "throttle.read_bps_device",
1229		.private = offsetof(struct throtl_grp, bps[READ]),
1230		.seq_show = tg_print_conf_u64,
1231		.write = tg_set_conf_u64,
1232	},
1233	{
1234		.name = "throttle.write_bps_device",
1235		.private = offsetof(struct throtl_grp, bps[WRITE]),
1236		.seq_show = tg_print_conf_u64,
1237		.write = tg_set_conf_u64,
1238	},
1239	{
1240		.name = "throttle.read_iops_device",
1241		.private = offsetof(struct throtl_grp, iops[READ]),
1242		.seq_show = tg_print_conf_uint,
1243		.write = tg_set_conf_uint,
1244	},
1245	{
1246		.name = "throttle.write_iops_device",
1247		.private = offsetof(struct throtl_grp, iops[WRITE]),
1248		.seq_show = tg_print_conf_uint,
1249		.write = tg_set_conf_uint,
1250	},
1251	{
1252		.name = "throttle.io_service_bytes",
1253		.private = (unsigned long)&blkcg_policy_throtl,
1254		.seq_show = blkg_print_stat_bytes,
1255	},
1256	{
1257		.name = "throttle.io_serviced",
1258		.private = (unsigned long)&blkcg_policy_throtl,
1259		.seq_show = blkg_print_stat_ios,
1260	},
1261	{ }	/* terminate */
1262};
1263
1264static u64 tg_prfill_max(struct seq_file *sf, struct blkg_policy_data *pd,
1265			 int off)
1266{
1267	struct throtl_grp *tg = pd_to_tg(pd);
1268	const char *dname = blkg_dev_name(pd->blkg);
1269	char bufs[4][21] = { "max", "max", "max", "max" };
1270
1271	if (!dname)
1272		return 0;
1273	if (tg->bps[READ] == -1 && tg->bps[WRITE] == -1 &&
1274	    tg->iops[READ] == -1 && tg->iops[WRITE] == -1)
1275		return 0;
1276
1277	if (tg->bps[READ] != -1)
1278		snprintf(bufs[0], sizeof(bufs[0]), "%llu", tg->bps[READ]);
1279	if (tg->bps[WRITE] != -1)
1280		snprintf(bufs[1], sizeof(bufs[1]), "%llu", tg->bps[WRITE]);
1281	if (tg->iops[READ] != -1)
1282		snprintf(bufs[2], sizeof(bufs[2]), "%u", tg->iops[READ]);
1283	if (tg->iops[WRITE] != -1)
1284		snprintf(bufs[3], sizeof(bufs[3]), "%u", tg->iops[WRITE]);
1285
1286	seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s\n",
1287		   dname, bufs[0], bufs[1], bufs[2], bufs[3]);
1288	return 0;
1289}
1290
1291static int tg_print_max(struct seq_file *sf, void *v)
 
1292{
1293	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_max,
1294			  &blkcg_policy_throtl, seq_cft(sf)->private, false);
1295	return 0;
 
 
1296}
1297
1298static ssize_t tg_set_max(struct kernfs_open_file *of,
1299			  char *buf, size_t nbytes, loff_t off)
1300{
1301	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1302	struct blkg_conf_ctx ctx;
1303	struct throtl_grp *tg;
1304	u64 v[4];
1305	int ret;
1306
1307	ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1308	if (ret)
1309		return ret;
1310
1311	tg = blkg_to_tg(ctx.blkg);
1312
1313	v[0] = tg->bps[READ];
1314	v[1] = tg->bps[WRITE];
1315	v[2] = tg->iops[READ];
1316	v[3] = tg->iops[WRITE];
1317
1318	while (true) {
1319		char tok[27];	/* wiops=18446744073709551616 */
1320		char *p;
1321		u64 val = -1;
1322		int len;
1323
1324		if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1325			break;
1326		if (tok[0] == '\0')
1327			break;
1328		ctx.body += len;
1329
1330		ret = -EINVAL;
1331		p = tok;
1332		strsep(&p, "=");
1333		if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1334			goto out_finish;
1335
1336		ret = -ERANGE;
1337		if (!val)
1338			goto out_finish;
1339
1340		ret = -EINVAL;
1341		if (!strcmp(tok, "rbps"))
1342			v[0] = val;
1343		else if (!strcmp(tok, "wbps"))
1344			v[1] = val;
1345		else if (!strcmp(tok, "riops"))
1346			v[2] = min_t(u64, val, UINT_MAX);
1347		else if (!strcmp(tok, "wiops"))
1348			v[3] = min_t(u64, val, UINT_MAX);
1349		else
1350			goto out_finish;
1351	}
1352
1353	tg->bps[READ] = v[0];
1354	tg->bps[WRITE] = v[1];
1355	tg->iops[READ] = v[2];
1356	tg->iops[WRITE] = v[3];
1357
1358	tg_conf_updated(tg);
1359	ret = 0;
1360out_finish:
1361	blkg_conf_finish(&ctx);
1362	return ret ?: nbytes;
1363}
1364
1365static struct cftype throtl_files[] = {
1366	{
1367		.name = "max",
1368		.flags = CFTYPE_NOT_ON_ROOT,
1369		.seq_show = tg_print_max,
1370		.write = tg_set_max,
1371	},
1372	{ }	/* terminate */
1373};
1374
1375static void throtl_shutdown_wq(struct request_queue *q)
1376{
1377	struct throtl_data *td = q->td;
 
 
 
 
1378
1379	cancel_work_sync(&td->dispatch_work);
1380}
 
 
1381
1382static struct blkcg_policy blkcg_policy_throtl = {
1383	.dfl_cftypes		= throtl_files,
1384	.legacy_cftypes		= throtl_legacy_files,
 
 
1385
1386	.pd_alloc_fn		= throtl_pd_alloc,
1387	.pd_init_fn		= throtl_pd_init,
1388	.pd_online_fn		= throtl_pd_online,
1389	.pd_free_fn		= throtl_pd_free,
1390};
 
 
 
 
 
 
 
 
 
1391
1392bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
1393		    struct bio *bio)
1394{
1395	struct throtl_qnode *qn = NULL;
1396	struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
1397	struct throtl_service_queue *sq;
1398	bool rw = bio_data_dir(bio);
1399	bool throttled = false;
1400
1401	WARN_ON_ONCE(!rcu_read_lock_held());
1402
1403	/* see throtl_charge_bio() */
1404	if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
1405		goto out;
1406
1407	spin_lock_irq(q->queue_lock);
 
1408
1409	if (unlikely(blk_queue_bypass(q)))
1410		goto out_unlock;
 
 
 
 
 
 
1411
1412	sq = &tg->service_queue;
 
 
 
 
 
 
1413
1414	while (true) {
1415		/* throtl is FIFO - if bios are already queued, should queue */
1416		if (sq->nr_queued[rw])
1417			break;
1418
1419		/* if above limits, break to queue */
1420		if (!tg_may_dispatch(tg, bio, NULL))
1421			break;
1422
1423		/* within limits, let's charge and dispatch directly */
 
1424		throtl_charge_bio(tg, bio);
1425
1426		/*
1427		 * We need to trim slice even when bios are not being queued
1428		 * otherwise it might happen that a bio is not queued for
1429		 * a long time and slice keeps on extending and trim is not
1430		 * called for a long time. Now if limits are reduced suddenly
1431		 * we take into account all the IO dispatched so far at new
1432		 * low rate and * newly queued IO gets a really long dispatch
1433		 * time.
1434		 *
1435		 * So keep on trimming slice even if bio is not queued.
1436		 */
1437		throtl_trim_slice(tg, rw);
1438
1439		/*
1440		 * @bio passed through this layer without being throttled.
1441		 * Climb up the ladder.  If we''re already at the top, it
1442		 * can be executed directly.
1443		 */
1444		qn = &tg->qnode_on_parent[rw];
1445		sq = sq->parent_sq;
1446		tg = sq_to_tg(sq);
1447		if (!tg)
1448			goto out_unlock;
1449	}
1450
1451	/* out-of-limit, queue to @tg */
1452	throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
1453		   rw == READ ? 'R' : 'W',
1454		   tg->bytes_disp[rw], bio->bi_iter.bi_size, tg->bps[rw],
1455		   tg->io_disp[rw], tg->iops[rw],
1456		   sq->nr_queued[READ], sq->nr_queued[WRITE]);
1457
1458	bio_associate_current(bio);
1459	tg->td->nr_queued[rw]++;
1460	throtl_add_bio_tg(bio, qn, tg);
1461	throttled = true;
1462
1463	/*
1464	 * Update @tg's dispatch time and force schedule dispatch if @tg
1465	 * was empty before @bio.  The forced scheduling isn't likely to
1466	 * cause undue delay as @bio is likely to be dispatched directly if
1467	 * its @tg's disptime is not in the future.
1468	 */
1469	if (tg->flags & THROTL_TG_WAS_EMPTY) {
1470		tg_update_disptime(tg);
1471		throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
1472	}
1473
1474out_unlock:
1475	spin_unlock_irq(q->queue_lock);
1476out:
1477	/*
1478	 * As multiple blk-throtls may stack in the same issue path, we
1479	 * don't want bios to leave with the flag set.  Clear the flag if
1480	 * being issued.
1481	 */
1482	if (!throttled)
1483		bio_clear_flag(bio, BIO_THROTTLED);
1484	return throttled;
1485}
1486
1487/*
1488 * Dispatch all bios from all children tg's queued on @parent_sq.  On
1489 * return, @parent_sq is guaranteed to not have any active children tg's
1490 * and all bios from previously active tg's are on @parent_sq->bio_lists[].
1491 */
1492static void tg_drain_bios(struct throtl_service_queue *parent_sq)
1493{
 
1494	struct throtl_grp *tg;
1495
1496	while ((tg = throtl_rb_first(parent_sq))) {
1497		struct throtl_service_queue *sq = &tg->service_queue;
1498		struct bio *bio;
1499
1500		throtl_dequeue_tg(tg);
 
 
 
 
 
 
 
1501
1502		while ((bio = throtl_peek_queued(&sq->queued[READ])))
1503			tg_dispatch_one_bio(tg, bio_data_dir(bio));
1504		while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
1505			tg_dispatch_one_bio(tg, bio_data_dir(bio));
1506	}
1507}
1508
1509/**
1510 * blk_throtl_drain - drain throttled bios
1511 * @q: request_queue to drain throttled bios for
1512 *
1513 * Dispatch all currently throttled bios on @q through ->make_request_fn().
1514 */
1515void blk_throtl_drain(struct request_queue *q)
1516	__releases(q->queue_lock) __acquires(q->queue_lock)
1517{
1518	struct throtl_data *td = q->td;
1519	struct blkcg_gq *blkg;
1520	struct cgroup_subsys_state *pos_css;
1521	struct bio *bio;
1522	int rw;
1523
1524	queue_lockdep_assert_held(q);
1525	rcu_read_lock();
 
 
1526
1527	/*
1528	 * Drain each tg while doing post-order walk on the blkg tree, so
1529	 * that all bios are propagated to td->service_queue.  It'd be
1530	 * better to walk service_queue tree directly but blkg walk is
1531	 * easier.
1532	 */
1533	blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
1534		tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
1535
1536	/* finally, transfer bios from top-level tg's into the td */
1537	tg_drain_bios(&td->service_queue);
 
 
1538
1539	rcu_read_unlock();
1540	spin_unlock_irq(q->queue_lock);
1541
1542	/* all bios now should be in td->service_queue, issue them */
1543	for (rw = READ; rw <= WRITE; rw++)
1544		while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
1545						NULL)))
1546			generic_make_request(bio);
1547
1548	spin_lock_irq(q->queue_lock);
1549}
1550
1551int blk_throtl_init(struct request_queue *q)
1552{
1553	struct throtl_data *td;
1554	int ret;
1555
1556	td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
1557	if (!td)
1558		return -ENOMEM;
1559
1560	INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
1561	throtl_service_queue_init(&td->service_queue);
 
 
 
 
 
 
 
 
 
 
 
1562
1563	q->td = td;
1564	td->queue = q;
1565
1566	/* activate policy */
1567	ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
1568	if (ret)
1569		kfree(td);
1570	return ret;
1571}
1572
1573void blk_throtl_exit(struct request_queue *q)
1574{
1575	BUG_ON(!q->td);
1576	throtl_shutdown_wq(q);
1577	blkcg_deactivate_policy(q, &blkcg_policy_throtl);
1578	kfree(q->td);
1579}
1580
1581static int __init throtl_init(void)
1582{
1583	kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
1584	if (!kthrotld_workqueue)
1585		panic("Failed to create kthrotld\n");
1586
1587	return blkcg_policy_register(&blkcg_policy_throtl);
 
1588}
1589
1590module_init(throtl_init);