[PATCH 4/7] io-throttle controller infrastructure

[Paul E. McKenney]

On Sun, May 03, 2009 at 01:36:20PM +0200, Andrea Righi wrote:
> This is the core of the io-throttle kernel infrastructure. It creates
> the basic interfaces to the cgroup subsystem and implements the I/O
> measurement and throttling functionality.

Looks good from an RCU perspective!

						Thanx, Paul

> Signed-off-by: Gui Jianfeng <guijianfeng at cn.fujitsu.com>
> Signed-off-by: Andrea Righi <righi.andrea at gmail.com>
> ---
>  block/Makefile                  |    1 +
>  block/blk-io-throttle.c         |  928 +++++++++++++++++++++++++++++++++++++++
>  include/linux/blk-io-throttle.h |  168 +++++++
>  include/linux/cgroup.h          |    1 +
>  include/linux/cgroup_subsys.h   |    6 +
>  init/Kconfig                    |   12 +
>  kernel/cgroup.c                 |    9 +
>  7 files changed, 1125 insertions(+), 0 deletions(-)
>  create mode 100644 block/blk-io-throttle.c
>  create mode 100644 include/linux/blk-io-throttle.h
> 
> diff --git a/block/Makefile b/block/Makefile
> index e9fa4dd..42b6a46 100644
> --- a/block/Makefile
> +++ b/block/Makefile
> @@ -13,5 +13,6 @@ obj-$(CONFIG_IOSCHED_AS)	+= as-iosched.o
>  obj-$(CONFIG_IOSCHED_DEADLINE)	+= deadline-iosched.o
>  obj-$(CONFIG_IOSCHED_CFQ)	+= cfq-iosched.o
> 
> +obj-$(CONFIG_CGROUP_IO_THROTTLE)	+= blk-io-throttle.o
>  obj-$(CONFIG_BLOCK_COMPAT)	+= compat_ioctl.o
>  obj-$(CONFIG_BLK_DEV_INTEGRITY)	+= blk-integrity.o
> diff --git a/block/blk-io-throttle.c b/block/blk-io-throttle.c
> new file mode 100644
> index 0000000..e2dfd24
> --- /dev/null
> +++ b/block/blk-io-throttle.c
> @@ -0,0 +1,928 @@
> +/*
> + * blk-io-throttle.c
> + *
> + * This program is free software; you can redistribute it and/or
> + * modify it under the terms of the GNU General Public
> + * License as published by the Free Software Foundation; either
> + * version 2 of the License, or (at your option) any later version.
> + *
> + * This program is distributed in the hope that it will be useful,
> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> + * General Public License for more details.
> + *
> + * You should have received a copy of the GNU General Public
> + * License along with this program; if not, write to the
> + * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
> + * Boston, MA 021110-1307, USA.
> + *
> + * Copyright (C) 2008 Andrea Righi <righi.andrea at gmail.com>
> + */
> +
> +#include <linux/init.h>
> +#include <linux/module.h>
> +#include <linux/res_counter.h>
> +#include <linux/memcontrol.h>
> +#include <linux/slab.h>
> +#include <linux/gfp.h>
> +#include <linux/err.h>
> +#include <linux/genhd.h>
> +#include <linux/hardirq.h>
> +#include <linux/list.h>
> +#include <linux/seq_file.h>
> +#include <linux/spinlock.h>
> +#include <linux/blk-io-throttle.h>
> +#include <linux/mm.h>
> +#include <linux/mm_inline.h>
> +#include <linux/page_cgroup.h>
> +#include <linux/sched.h>
> +#include <linux/bio.h>
> +
> +/*
> + * Statistics for I/O bandwidth controller.
> + */
> +enum iothrottle_stat_index {
> +	/* # of times the cgroup has been throttled for bw limit */
> +	IOTHROTTLE_STAT_BW_COUNT,
> +	/* # of jiffies spent to sleep for throttling for bw limit */
> +	IOTHROTTLE_STAT_BW_SLEEP,
> +	/* # of times the cgroup has been throttled for iops limit */
> +	IOTHROTTLE_STAT_IOPS_COUNT,
> +	/* # of jiffies spent to sleep for throttling for iops limit */
> +	IOTHROTTLE_STAT_IOPS_SLEEP,
> +	/* total number of bytes read and written */
> +	IOTHROTTLE_STAT_BYTES_TOT,
> +	/* total number of I/O operations */
> +	IOTHROTTLE_STAT_IOPS_TOT,
> +
> +	IOTHROTTLE_STAT_NSTATS,
> +};
> +
> +struct iothrottle_stat_cpu {
> +	unsigned long long count[IOTHROTTLE_STAT_NSTATS];
> +} ____cacheline_aligned_in_smp;
> +
> +struct iothrottle_stat {
> +	struct iothrottle_stat_cpu cpustat[NR_CPUS];
> +};
> +
> +static void iothrottle_stat_add(struct iothrottle_stat *stat,
> +			enum iothrottle_stat_index type, unsigned long long val)
> +{
> +	int cpu = get_cpu();
> +
> +	stat->cpustat[cpu].count[type] += val;
> +	put_cpu();
> +}
> +
> +static void iothrottle_stat_add_sleep(struct iothrottle_stat *stat,
> +			int type, unsigned long long sleep)
> +{
> +	int cpu = get_cpu();
> +
> +	switch (type) {
> +	case IOTHROTTLE_BANDWIDTH:
> +		stat->cpustat[cpu].count[IOTHROTTLE_STAT_BW_COUNT]++;
> +		stat->cpustat[cpu].count[IOTHROTTLE_STAT_BW_SLEEP] += sleep;
> +		break;
> +	case IOTHROTTLE_IOPS:
> +		stat->cpustat[cpu].count[IOTHROTTLE_STAT_IOPS_COUNT]++;
> +		stat->cpustat[cpu].count[IOTHROTTLE_STAT_IOPS_SLEEP] += sleep;
> +		break;
> +	}
> +	put_cpu();
> +}
> +
> +static unsigned long long iothrottle_read_stat(struct iothrottle_stat *stat,
> +				enum iothrottle_stat_index idx)
> +{
> +	int cpu;
> +	unsigned long long ret = 0;
> +
> +	for_each_possible_cpu(cpu)
> +		ret += stat->cpustat[cpu].count[idx];
> +	return ret;
> +}
> +
> +struct iothrottle_sleep {
> +	unsigned long long bw_sleep;
> +	unsigned long long iops_sleep;
> +};
> +
> +/*
> + * struct iothrottle_node - throttling rule of a single block device
> + * @node: list of per block device throttling rules
> + * @dev: block device number, used as key in the list
> + * @bw: max i/o bandwidth (in bytes/s)
> + * @iops: max i/o operations per second
> + * @stat: throttling statistics
> + *
> + * Define a i/o throttling rule for a single block device.
> + *
> + * NOTE: limiting rules always refer to dev_t; if a block device is unplugged
> + * the limiting rules defined for that device persist and they are still valid
> + * if a new device is plugged and it uses the same dev_t number.
> + */
> +struct iothrottle_node {
> +	struct list_head node;
> +	dev_t dev;
> +	struct res_counter bw;
> +	struct res_counter iops;
> +	struct iothrottle_stat stat;
> +};
> +
> +/**
> + * struct iothrottle - throttling rules for a cgroup
> + * @css: pointer to the cgroup state
> + * @list: list of iothrottle_node elements
> + * @watermark: a watermark in percentage to start/stop i/o throttling
> + *
> + * Define multiple per-block device i/o throttling rules.
> + *
> + * Throttling will begin when the percentage of the consumed disk i/o bandwidth
> + * exceeds the attribute @watermark. If @watermark is 0 throttling is applied
> + * immediately (hard limit).
> + *
> + * Note: the list of the throttling rules is protected by RCU locking:
> + *	 - hold cgroup_lock() for update.
> + *	 - hold rcu_read_lock() for read.
> + */
> +struct iothrottle {
> +	struct cgroup_subsys_state css;
> +	struct list_head list;
> +	unsigned long watermark;
> +};
> +static struct iothrottle init_iothrottle;
> +
> +static inline struct iothrottle *cgroup_to_iothrottle(struct cgroup *cgrp)
> +{
> +	return container_of(cgroup_subsys_state(cgrp, iothrottle_subsys_id),
> +			    struct iothrottle, css);
> +}
> +
> +/*
> + * Note: called with rcu_read_lock() held.
> + */
> +static inline struct iothrottle *task_to_iothrottle(struct task_struct *task)
> +{
> +	return container_of(task_subsys_state(task, iothrottle_subsys_id),
> +			    struct iothrottle, css);
> +}
> +
> +/*
> + * Note: called with rcu_read_lock() or cgroup_lock() held.
> + */
> +static struct iothrottle_node *
> +iothrottle_search_node(const struct iothrottle *iot, dev_t dev)
> +{
> +	struct iothrottle_node *n;
> +
> +	if (list_empty(&iot->list))
> +		return NULL;
> +	list_for_each_entry_rcu(n, &iot->list, node)
> +		if (n->dev == dev)
> +			return n;
> +	return NULL;
> +}
> +
> +/*
> + * Note: called with cgroup_lock() held.
> + */
> +static void iothrottle_insert_node(struct iothrottle *iot,
> +					struct iothrottle_node *n)
> +{
> +	WARN_ON_ONCE(!cgroup_is_locked());
> +	list_add_rcu(&n->node, &iot->list);
> +}
> +
> +/*
> + * Note: called with cgroup_lock() held.
> + */
> +static void
> +iothrottle_replace_node(struct iothrottle *iot, struct iothrottle_node *old,
> +			struct iothrottle_node *new)
> +{
> +	WARN_ON_ONCE(!cgroup_is_locked());
> +	list_replace_rcu(&old->node, &new->node);
> +}
> +
> +/*
> + * Note: called with cgroup_lock() held.
> + */
> +static void
> +iothrottle_delete_node(struct iothrottle *iot, struct iothrottle_node *n)
> +{
> +	WARN_ON_ONCE(!cgroup_is_locked());
> +	list_del_rcu(&n->node);
> +}
> +
> +/*
> + * Note: called from kernel/cgroup.c with cgroup_lock() held.
> + */
> +static struct cgroup_subsys_state *
> +iothrottle_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
> +{
> +	struct iothrottle *iot;
> +
> +	if (unlikely((cgrp->parent) == NULL)) {
> +		iot = &init_iothrottle;
> +	} else {
> +		iot = kzalloc(sizeof(*iot), GFP_KERNEL);
> +		if (unlikely(!iot))
> +			return ERR_PTR(-ENOMEM);
> +	}
> +	INIT_LIST_HEAD(&iot->list);
> +
> +	return &iot->css;
> +}
> +
> +/*
> + * Note: called from kernel/cgroup.c with cgroup_lock() held.
> + */
> +static void iothrottle_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
> +{
> +	struct iothrottle_node *n, *p;
> +	struct iothrottle *iot = cgroup_to_iothrottle(cgrp);
> +
> +	free_css_id(&iothrottle_subsys, &iot->css);
> +	/*
> +	 * don't worry about locking here, at this point there must be not any
> +	 * reference to the list.
> +	 */
> +	if (!list_empty(&iot->list))
> +		list_for_each_entry_safe(n, p, &iot->list, node)
> +			kfree(n);
> +	kfree(iot);
> +}
> +
> +/*
> + * NOTE: called with rcu_read_lock() held.
> + *
> + * do not care too much about locking for single res_counter values here.
> + */
> +static void iothrottle_show_limit(struct seq_file *m, dev_t dev,
> +			struct res_counter *res)
> +{
> +	if (!res->limit)
> +		return;
> +	/* maj min bw-limit ratelimit-policy usage bucket-size delta-time */
> +	seq_printf(m, "%u %u %llu %lu %lli %llu %li\n",
> +		MAJOR(dev), MINOR(dev),
> +		res->limit, res_counter_flagged(res, RES_COUNTER_POLICY),
> +		(long long)res->usage, res->max_usage,
> +		jiffies_to_clock_t(res_counter_ratelimit_delta_t(res)));
> +}
> +
> +/*
> + * NOTE: called with rcu_read_lock() held.
> + *
> + */
> +static void iothrottle_show_failcnt(struct seq_file *m, dev_t dev,
> +				struct iothrottle_stat *stat)
> +{
> +	unsigned long long bw_count, bw_sleep, iops_count, iops_sleep;
> +
> +	bw_count = iothrottle_read_stat(stat, IOTHROTTLE_STAT_BW_COUNT);
> +	bw_sleep = iothrottle_read_stat(stat, IOTHROTTLE_STAT_BW_SLEEP);
> +	iops_count = iothrottle_read_stat(stat, IOTHROTTLE_STAT_IOPS_COUNT);
> +	iops_sleep = iothrottle_read_stat(stat, IOTHROTTLE_STAT_IOPS_SLEEP);
> +
> +	seq_printf(m, "%u %u %llu %li %llu %li\n", MAJOR(dev), MINOR(dev),
> +		bw_count, jiffies_to_clock_t(bw_sleep),
> +		iops_count, jiffies_to_clock_t(iops_sleep));
> +}
> +
> +/*
> + * NOTE: called with rcu_read_lock() held.
> + */
> +static void iothrottle_show_stat(struct seq_file *m, dev_t dev,
> +				struct iothrottle_stat *stat)
> +{
> +	unsigned long long bytes, iops;
> +
> +	bytes = iothrottle_read_stat(stat, IOTHROTTLE_STAT_BYTES_TOT);
> +	iops = iothrottle_read_stat(stat, IOTHROTTLE_STAT_IOPS_TOT);
> +
> +	seq_printf(m, "%u %u %llu %llu\n", MAJOR(dev), MINOR(dev), bytes, iops);
> +}
> +
> +static int iothrottle_read(struct cgroup *cgrp, struct cftype *cft,
> +				struct seq_file *m)
> +{
> +	struct iothrottle *iot = cgroup_to_iothrottle(cgrp);
> +	struct iothrottle_node *n;
> +
> +	rcu_read_lock();
> +	if (list_empty(&iot->list))
> +		goto unlock_and_return;
> +	list_for_each_entry_rcu(n, &iot->list, node) {
> +		BUG_ON(!n->dev);
> +		switch (cft->private) {
> +		case IOTHROTTLE_BANDWIDTH:
> +			iothrottle_show_limit(m, n->dev, &n->bw);
> +			break;
> +		case IOTHROTTLE_IOPS:
> +			iothrottle_show_limit(m, n->dev, &n->iops);
> +			break;
> +		case IOTHROTTLE_FAILCNT:
> +			iothrottle_show_failcnt(m, n->dev, &n->stat);
> +			break;
> +		case IOTHROTTLE_STAT:
> +			iothrottle_show_stat(m, n->dev, &n->stat);
> +			break;
> +		}
> +	}
> +unlock_and_return:
> +	rcu_read_unlock();
> +	return 0;
> +}
> +
> +static dev_t devname2dev_t(const char *buf)
> +{
> +	struct block_device *bdev;
> +	dev_t dev = 0;
> +	struct gendisk *disk;
> +	int part;
> +
> +	/* use a lookup to validate the block device */
> +	bdev = lookup_bdev(buf);
> +	if (IS_ERR(bdev))
> +		return 0;
> +	/* only entire devices are allowed, not single partitions */
> +	disk = get_gendisk(bdev->bd_dev, &part);
> +	if (disk && !part) {
> +		BUG_ON(!bdev->bd_inode);
> +		dev = bdev->bd_inode->i_rdev;
> +	}
> +	bdput(bdev);
> +
> +	return dev;
> +}
> +
> +/*
> + * The userspace input string must use one of the following syntaxes:
> + *
> + * dev:0			<- delete an i/o limiting rule
> + * dev:io-limit:0		<- set a leaky bucket throttling rule
> + * dev:io-limit:1:bucket-size	<- set a token bucket throttling rule
> + * dev:io-limit:1		<- set a token bucket throttling rule using
> + *				   bucket-size == io-limit
> + */
> +static int iothrottle_parse_args(char *buf, size_t nbytes, int filetype,
> +			dev_t *dev, unsigned long long *iolimit,
> +			unsigned long *strategy,
> +			unsigned long long *bucket_size)
> +{
> +	char *p;
> +	int count = 0;
> +	char *s[4];
> +	int ret;
> +
> +	memset(s, 0, sizeof(s));
> +	*dev = 0;
> +	*iolimit = 0;
> +	*strategy = 0;
> +	*bucket_size = 0;
> +
> +	/* split the colon-delimited input string into its elements */
> +	while (count < ARRAY_SIZE(s)) {
> +		p = strsep(&buf, ":");
> +		if (!p)
> +			break;
> +		if (!*p)
> +			continue;
> +		s[count++] = p;
> +	}
> +
> +	/* i/o limit */
> +	if (!s[1])
> +		return -EINVAL;
> +	ret = strict_strtoull(s[1], 10, iolimit);
> +	if (ret < 0)
> +		return ret;
> +	if (!*iolimit)
> +		goto out;
> +	/* throttling strategy (leaky bucket / token bucket) */
> +	if (!s[2])
> +		return -EINVAL;
> +	ret = strict_strtoul(s[2], 10, strategy);
> +	if (ret < 0)
> +		return ret;
> +	switch (*strategy) {
> +	case RATELIMIT_LEAKY_BUCKET:
> +		goto out;
> +	case RATELIMIT_TOKEN_BUCKET:
> +		break;
> +	default:
> +		return -EINVAL;
> +	}
> +	/* bucket size */
> +	if (!s[3])
> +		*bucket_size = *iolimit;
> +	else {
> +		ret = strict_strtoll(s[3], 10, bucket_size);
> +		if (ret < 0)
> +			return ret;
> +	}
> +	if (*bucket_size <= 0)
> +		return -EINVAL;
> +out:
> +	/* block device number */
> +	*dev = devname2dev_t(s[0]);
> +	return *dev ? 0 : -EINVAL;
> +}
> +
> +static int iothrottle_write(struct cgroup *cgrp, struct cftype *cft,
> +				const char *buffer)
> +{
> +	struct iothrottle *iot;
> +	struct iothrottle_node *n, *newn = NULL;
> +	dev_t dev;
> +	unsigned long long iolimit, bucket_size;
> +	unsigned long strategy;
> +	char *buf;
> +	size_t nbytes = strlen(buffer);
> +	int ret = 0;
> +
> +	/*
> +	 * We need to allocate a new buffer here, because
> +	 * iothrottle_parse_args() can modify it and the buffer provided by
> +	 * write_string is supposed to be const.
> +	 */
> +	buf = kmalloc(nbytes + 1, GFP_KERNEL);
> +	if (!buf)
> +		return -ENOMEM;
> +	memcpy(buf, buffer, nbytes + 1);
> +
> +	ret = iothrottle_parse_args(buf, nbytes, cft->private, &dev, &iolimit,
> +				&strategy, &bucket_size);
> +	if (ret)
> +		goto out1;
> +	newn = kzalloc(sizeof(*newn), GFP_KERNEL);
> +	if (!newn) {
> +		ret = -ENOMEM;
> +		goto out1;
> +	}
> +	newn->dev = dev;
> +	res_counter_init(&newn->bw, NULL);
> +	res_counter_init(&newn->iops, NULL);
> +
> +	switch (cft->private) {
> +	case IOTHROTTLE_BANDWIDTH:
> +		res_counter_ratelimit_set_limit(&newn->iops, 0, 0, 0);
> +		res_counter_ratelimit_set_limit(&newn->bw, strategy,
> +				ALIGN(iolimit, 1024), ALIGN(bucket_size, 1024));
> +		break;
> +	case IOTHROTTLE_IOPS:
> +		res_counter_ratelimit_set_limit(&newn->bw, 0, 0, 0);
> +		/*
> +		 * scale up iops cost by a factor of 1000, this allows to apply
> +		 * a more fine grained sleeps, and throttling results more
> +		 * precise this way.
> +		 */
> +		res_counter_ratelimit_set_limit(&newn->iops, strategy,
> +				iolimit * 1000, bucket_size * 1000);
> +		break;
> +	default:
> +		WARN_ON(1);
> +		break;
> +	}
> +
> +	if (!cgroup_lock_live_group(cgrp)) {
> +		ret = -ENODEV;
> +		goto out1;
> +	}
> +	iot = cgroup_to_iothrottle(cgrp);
> +
> +	n = iothrottle_search_node(iot, dev);
> +	if (!n) {
> +		if (iolimit) {
> +			/* Add a new block device limiting rule */
> +			iothrottle_insert_node(iot, newn);
> +			newn = NULL;
> +		}
> +		goto out2;
> +	}
> +	switch (cft->private) {
> +	case IOTHROTTLE_BANDWIDTH:
> +		if (!iolimit && !n->iops.limit) {
> +			/* Delete a block device limiting rule */
> +			iothrottle_delete_node(iot, n);
> +			goto out2;
> +		}
> +		if (!n->iops.limit)
> +			break;
> +		/* Update a block device limiting rule */
> +		newn->iops = n->iops;
> +		break;
> +	case IOTHROTTLE_IOPS:
> +		if (!iolimit && !n->bw.limit) {
> +			/* Delete a block device limiting rule */
> +			iothrottle_delete_node(iot, n);
> +			goto out2;
> +		}
> +		if (!n->bw.limit)
> +			break;
> +		/* Update a block device limiting rule */
> +		newn->bw = n->bw;
> +		break;
> +	}
> +	iothrottle_replace_node(iot, n, newn);
> +	newn = NULL;
> +out2:
> +	cgroup_unlock();
> +	if (n) {
> +		synchronize_rcu();
> +		kfree(n);
> +	}
> +out1:
> +	kfree(newn);
> +	kfree(buf);
> +	return ret;
> +}
> +
> +static u64 iothrottle_watermark_read(struct cgroup *cgrp, struct cftype *cft)
> +{
> +	struct iothrottle *iot = cgroup_to_iothrottle(cgrp);
> +
> +	return iot->watermark;
> +}
> +
> +static int iothrottle_watermark_write(struct cgroup *cgrp,
> +				struct cftype *cftype, u64 val)
> +{
> +	struct iothrottle *iot = cgroup_to_iothrottle(cgrp);
> +
> +	if (val > 100)
> +		return -EINVAL;
> +	iot->watermark = val;
> +	return 0;
> +}
> +
> +static struct cftype files[] = {
> +	{
> +		.name = "bandwidth-max",
> +		.read_seq_string = iothrottle_read,
> +		.write_string = iothrottle_write,
> +		.max_write_len = 256,
> +		.private = IOTHROTTLE_BANDWIDTH,
> +	},
> +	{
> +		.name = "iops-max",
> +		.read_seq_string = iothrottle_read,
> +		.write_string = iothrottle_write,
> +		.max_write_len = 256,
> +		.private = IOTHROTTLE_IOPS,
> +	},
> +	{
> +		.name = "throttlecnt",
> +		.read_seq_string = iothrottle_read,
> +		.private = IOTHROTTLE_FAILCNT,
> +	},
> +	{
> +		.name = "stat",
> +		.read_seq_string = iothrottle_read,
> +		.private = IOTHROTTLE_STAT,
> +	},
> +	{
> +		.name = "watermark",
> +		.read_u64 = iothrottle_watermark_read,
> +		.write_u64 = iothrottle_watermark_write,
> +	},
> +};
> +
> +static int iothrottle_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
> +{
> +	return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
> +}
> +
> +struct cgroup_subsys iothrottle_subsys = {
> +	.name = "blockio",
> +	.create = iothrottle_create,
> +	.destroy = iothrottle_destroy,
> +	.populate = iothrottle_populate,
> +	.subsys_id = iothrottle_subsys_id,
> +	.early_init = 1,
> +	.use_id = 1,
> +};
> +
> +/*
> + * NOTE: called with rcu_read_lock() held.
> + */
> +static void iothrottle_evaluate_sleep(struct iothrottle_sleep *sleep,
> +				struct iothrottle *iot,
> +				struct block_device *bdev, ssize_t bytes)
> +{
> +	struct iothrottle_node *n;
> +	dev_t dev;
> +
> +	BUG_ON(!iot);
> +
> +	/* accounting and throttling is done only on entire block devices */
> +	dev = MKDEV(MAJOR(bdev->bd_inode->i_rdev), bdev->bd_disk->first_minor);
> +	n = iothrottle_search_node(iot, dev);
> +	if (!n)
> +		return;
> +
> +	/* Update statistics */
> +	iothrottle_stat_add(&n->stat, IOTHROTTLE_STAT_BYTES_TOT, bytes);
> +	if (bytes)
> +		iothrottle_stat_add(&n->stat, IOTHROTTLE_STAT_IOPS_TOT, 1);
> +
> +	/* Evaluate sleep values */
> +	sleep->bw_sleep = res_counter_ratelimit_sleep(&n->bw, bytes);
> +	/*
> +	 * scale up iops cost by a factor of 1000, this allows to apply
> +	 * a more fine grained sleeps, and throttling works better in
> +	 * this way.
> +	 *
> +	 * Note: do not account any i/o operation if bytes is negative or zero.
> +	 */
> +	sleep->iops_sleep = res_counter_ratelimit_sleep(&n->iops,
> +						bytes ? 1000 : 0);
> +}
> +
> +/*
> + * NOTE: called with rcu_read_lock() held.
> + */
> +static void iothrottle_acct_stat(struct iothrottle *iot,
> +			struct block_device *bdev, int type,
> +			unsigned long long sleep)
> +{
> +	struct iothrottle_node *n;
> +	dev_t dev = MKDEV(MAJOR(bdev->bd_inode->i_rdev),
> +			bdev->bd_disk->first_minor);
> +
> +	n = iothrottle_search_node(iot, dev);
> +	if (!n)
> +		return;
> +	iothrottle_stat_add_sleep(&n->stat, type, sleep);
> +}
> +
> +static void iothrottle_acct_task_stat(int type, unsigned long long sleep)
> +{
> +	/*
> +	 * XXX: per-task statistics may be inaccurate (this is not a
> +	 * critical issue, anyway, respect to introduce locking
> +	 * overhead or increase the size of task_struct).
> +	 */
> +	switch (type) {
> +	case IOTHROTTLE_BANDWIDTH:
> +		current->io_throttle_bw_cnt++;
> +		current->io_throttle_bw_sleep += sleep;
> +		break;
> +
> +	case IOTHROTTLE_IOPS:
> +		current->io_throttle_iops_cnt++;
> +		current->io_throttle_iops_sleep += sleep;
> +		break;
> +	}
> +}
> +
> +/*
> + * A helper function to get iothrottle from css id.
> + *
> + * NOTE: must be called under rcu_read_lock(). The caller must check
> + * css_is_removed() or some if it's concern.
> + */
> +static struct iothrottle *iothrottle_lookup(unsigned long id)
> +{
> +	struct cgroup_subsys_state *css;
> +
> +	if (!id)
> +		return NULL;
> +	css = css_lookup(&iothrottle_subsys, id);
> +	if (!css)
> +		return NULL;
> +	return container_of(css, struct iothrottle, css);
> +}
> +
> +static struct iothrottle *get_iothrottle_from_page(struct page *page)
> +{
> +	struct iothrottle *iot;
> +	unsigned long id;
> +
> +	BUG_ON(!page);
> +	id = page_cgroup_get_owner(page);
> +
> +	rcu_read_lock();
> +	iot = iothrottle_lookup(id);
> +	if (!iot)
> +		goto out;
> +	css_get(&iot->css);
> +out:
> +	rcu_read_unlock();
> +	return iot;
> +}
> +
> +static struct iothrottle *get_iothrottle_from_bio(struct bio *bio)
> +{
> +	if (!bio)
> +		return NULL;
> +	return get_iothrottle_from_page(bio_page(bio));
> +}
> +
> +int iothrottle_set_page_owner(struct page *page, struct mm_struct *mm)
> +{
> +	struct iothrottle *iot;
> +	unsigned short id = 0;
> +
> +	if (iothrottle_disabled())
> +		return 0;
> +	if (!mm)
> +		goto out;
> +	rcu_read_lock();
> +	iot = task_to_iothrottle(rcu_dereference(mm->owner));
> +	if (likely(iot))
> +		id = css_id(&iot->css);
> +	rcu_read_unlock();
> +out:
> +	return page_cgroup_set_owner(page, id);
> +}
> +
> +int iothrottle_set_pagedirty_owner(struct page *page, struct mm_struct *mm)
> +{
> +	if (!page_is_file_cache(page))
> +		return 0;
> +	if (current->flags & PF_MEMALLOC)
> +		return 0;
> +	return iothrottle_set_page_owner(page, mm);
> +}
> +
> +int iothrottle_copy_page_owner(struct page *npage, struct page *opage)
> +{
> +	if (iothrottle_disabled())
> +		return 0;
> +	return page_cgroup_copy_owner(npage, opage);
> +}
> +
> +static inline int is_kthread_io(void)
> +{
> +	return current->flags & (PF_KTHREAD | PF_FLUSHER | PF_KSWAPD);
> +}
> +
> +static bool is_urgent_io(struct bio *bio)
> +{
> +	if (bio && (bio_rw_meta(bio) || bio_noidle(bio)))
> +		return true;
> +	if (has_fs_excl())
> +		return true;
> +	return false;
> +}
> +
> +static void iothrottle_force_sleep(int type, unsigned long long sleep)
> +{
> +	pr_debug("io-throttle: task %p (%s) must sleep %llu jiffies\n",
> +			current, current->comm, sleep);
> +	iothrottle_acct_task_stat(type, sleep);
> +	schedule_timeout_killable(sleep);
> +}
> +
> +/**
> + * bw_usage() - evaluate i/o bandwidth utilization in percentage
> + * @bdev:	block device involved for the i/o
> + *
> + * Evaluate the bandwidth utilization of a block device as the percentage of
> + * time during which i/o requests were issued to the device. Device saturation
> + * occurs when this value is close to 100.
> + *
> + * NOTE: the evaluated bandwidth may be inaccurate, also because we don't use a
> + * lock to protect the block device statistics. The inaccuracy can be tolerated
> + * in this case: at worst an additional throttling may be imposed even if the
> + * consumed bandwidth doesn't hit the water mark. The advantage is obviously a
> + * more scalable and lock-free implementation.
> + **/
> +static unsigned long bw_usage(struct block_device *bdev)
> +{
> +	unsigned int iotime, delta_t;
> +	unsigned long long io_bw;
> +
> +	/* cpu ticks from the last i/o request */
> +	delta_t = (long)jiffies - (long)bdev->last_access;
> +
> +	/* i/o ticks consumed from the last i/o request */
> +	iotime = part_stat_read(bdev->bd_part, io_ticks);
> +
> +	/* i/o bw = (cpu ticks / io ticks) * 100 */
> +	io_bw = ((long)iotime - (long)bdev->last_io_ticks) * 100;
> +	if (delta_t && io_bw) {
> +		io_bw = div_u64(io_bw, delta_t);
> +		/* update bdev statistics */
> +		bdev->last_io_ticks = iotime;
> +		bdev->last_access = jiffies;
> +	}
> +	return io_bw;
> +}
> +
> +/**
> + * cgroup_io_throttle() - account and throttle synchronous i/o activity
> + * @bio:	the bio structure used to retrieve the owner of the i/o
> + *		operation.
> + * @bdev:	block device involved for the i/o.
> + * @bytes:	size in bytes of the i/o operation.
> + *
> + * This is the core of the block device i/o bandwidth controller. This function
> + * must be called by any function that generates i/o activity (directly or
> + * indirectly). It provides both i/o accounting and throttling functionalities.
> + *
> + * Returns the value of sleep in jiffies if it was not possible to schedule the
> + * timeout.
> + **/
> +unsigned long long
> +cgroup_io_throttle(struct bio *bio, struct block_device *bdev, ssize_t bytes)
> +{
> +	struct iothrottle *iot = NULL, *curr_iot;
> +	struct iothrottle_sleep s = {};
> +	unsigned long long sleep;
> +	int type, can_sleep = 1;
> +
> +	if (iothrottle_disabled())
> +		return 0;
> +	if (unlikely(!bdev))
> +		return 0;
> +	BUG_ON(!bdev->bd_inode || !bdev->bd_disk);
> +	/*
> +	 * Never throttle kernel threads directly, since they may completely
> +	 * block other cgroups, the i/o on other block devices or even the
> +	 * whole system.
> +	 *
> +	 * For the same reason never throttle IO that comes from tasks that are
> +	 * holding exclusive access resources (urgent IO).
> +	 *
> +	 * And never sleep if we're inside an AIO context; just account the i/o
> +	 * activity. Throttling is performed in io_submit_one() returning
> +	 * -EAGAIN when the limits are exceeded.
> +	 */
> +	if (is_kthread_io() || is_urgent_io(bio) || is_in_aio())
> +		can_sleep = 0;
> +	/*
> +	 * WARNING: in_atomic() do not know about held spinlocks in
> +	 * non-preemptible kernels, but we want to check it here to raise
> +	 * potential bugs when a preemptible kernel is used.
> +	 */
> +	WARN_ON_ONCE(can_sleep &&
> +		(irqs_disabled() || in_interrupt() || in_atomic()));
> +	/*
> +	 * Evaluate the IO context of bio.
> +	 *
> +	 * In O_DIRECT mode the context of bio always refers to the current
> +	 * task. Otherwise, to differentiate writeback IO from synchronous IO
> +	 * we compare the bio's io-throttle cgroup with the current task's
> +	 * cgroup. If they're different we're doing writeback IO and we can't
> +	 * throttle the current task directly.
> +	 */
> +	if (!is_in_dio()) {
> +		/*
> +		 * We're not doing O_DIRECT: find the source of this IO
> +		 * request.
> +		 */
> +		iot = get_iothrottle_from_bio(bio);
> +	}
> +	rcu_read_lock();
> +	curr_iot = task_to_iothrottle(current);
> +	if (iot == NULL) {
> +		/* IO occurs in the same context of the current task */
> +		iot = curr_iot;
> +		css_get(&iot->css);
> +	}
> +	/* Do not apply throttling if the water mark is not reached */
> +	if (bw_usage(bdev) < iot->watermark) {
> +		css_put(&iot->css);
> +		rcu_read_unlock();
> +		return 0;
> +	}
> +	if (iot != curr_iot) {
> +		/*
> +		 * IO occurs in a different context of the current task
> +		 * (writeback IO).
> +		 *
> +		 * Do not throttle current task directly in this case, just
> +		 * delay the submission of the IO request (that will be
> +		 * dispatched by kiothrottled).
> +		 */
> +		can_sleep = 0;
> +	}
> +	/* Apply IO throttling */
> +	iothrottle_evaluate_sleep(&s, iot, bdev, bytes);
> +	sleep = max(s.bw_sleep, s.iops_sleep);
> +	type = (s.bw_sleep < s.iops_sleep) ?
> +			IOTHROTTLE_IOPS : IOTHROTTLE_BANDWIDTH;
> +	if (unlikely(sleep && can_sleep))
> +		iothrottle_acct_stat(iot, bdev, type, sleep);
> +	css_put(&iot->css);
> +	rcu_read_unlock();
> +	if (unlikely(sleep && can_sleep)) {
> +		/* Throttle the current task directly */
> +		iothrottle_force_sleep(type, sleep);
> +		return 0;
> +	}
> +	/*
> +	 * Account, but do not throttle async filesystems' metadata IO or IO
> +	 * that is explicitly marked to not wait or being anticipated, i.e.
> +	 * writes with wbc->sync_mode set to WBC_SYNC_ALL - fsync() - or
> +	 * journal activity.
> +	 */
> +	if (is_urgent_io(bio))
> +		sleep = 0;
> +	return sleep;
> +}
> diff --git a/include/linux/blk-io-throttle.h b/include/linux/blk-io-throttle.h
> new file mode 100644
> index 0000000..e448130
> --- /dev/null
> +++ b/include/linux/blk-io-throttle.h
> @@ -0,0 +1,168 @@
> +#ifndef BLK_IO_THROTTLE_H
> +#define BLK_IO_THROTTLE_H
> +
> +#include <linux/fs.h>
> +#include <linux/jiffies.h>
> +#include <linux/sched.h>
> +#include <linux/cgroup.h>
> +#include <asm/atomic.h>
> +#include <asm/current.h>
> +
> +#define IOTHROTTLE_BANDWIDTH	0
> +#define IOTHROTTLE_IOPS		1
> +#define IOTHROTTLE_FAILCNT	2
> +#define IOTHROTTLE_STAT		3
> +
> +#ifdef CONFIG_CGROUP_IO_THROTTLE
> +
> +static inline bool iothrottle_disabled(void)
> +{
> +	if (iothrottle_subsys.disabled)
> +		return true;
> +	return false;
> +}
> +
> +extern unsigned long long
> +cgroup_io_throttle(struct bio *bio, struct block_device *bdev, ssize_t bytes);
> +
> +extern int iothrottle_make_request(struct bio *bio, unsigned long deadline);
> +
> +int iothrottle_set_page_owner(struct page *page, struct mm_struct *mm);
> +int iothrottle_set_pagedirty_owner(struct page *page, struct mm_struct *mm);
> +int iothrottle_copy_page_owner(struct page *npage, struct page *opage);
> +
> +extern int iothrottle_sync(void);
> +
> +static inline void set_in_aio(void)
> +{
> +	atomic_set(&current->in_aio, 1);
> +}
> +
> +static inline void unset_in_aio(void)
> +{
> +	atomic_set(&current->in_aio, 0);
> +}
> +
> +static inline int is_in_aio(void)
> +{
> +	return atomic_read(&current->in_aio);
> +}
> +
> +static inline void set_in_dio(void)
> +{
> +	atomic_set(&current->in_dio, 1);
> +}
> +
> +static inline void unset_in_dio(void)
> +{
> +	atomic_set(&current->in_dio, 0);
> +}
> +
> +static inline int is_in_dio(void)
> +{
> +	return atomic_read(&current->in_dio);
> +}
> +
> +static inline unsigned long long
> +get_io_throttle_cnt(struct task_struct *t, int type)
> +{
> +	switch (type) {
> +	case IOTHROTTLE_BANDWIDTH:
> +		return t->io_throttle_bw_cnt;
> +	case IOTHROTTLE_IOPS:
> +		return t->io_throttle_iops_cnt;
> +	}
> +	BUG();
> +}
> +
> +static inline unsigned long long
> +get_io_throttle_sleep(struct task_struct *t, int type)
> +{
> +	switch (type) {
> +	case IOTHROTTLE_BANDWIDTH:
> +		return jiffies_to_clock_t(t->io_throttle_bw_sleep);
> +	case IOTHROTTLE_IOPS:
> +		return jiffies_to_clock_t(t->io_throttle_iops_sleep);
> +	}
> +	BUG();
> +}
> +#else /* CONFIG_CGROUP_IO_THROTTLE */
> +
> +static inline bool iothrottle_disabled(void)
> +{
> +	return true;
> +}
> +
> +static inline unsigned long long
> +cgroup_io_throttle(struct bio *bio, struct block_device *bdev, ssize_t bytes)
> +{
> +	return 0;
> +}
> +
> +static inline int
> +iothrottle_make_request(struct bio *bio, unsigned long deadline)
> +{
> +	return 0;
> +}
> +
> +static inline int iothrottle_set_page_owner(struct page *page,
> +				struct mm_struct *mm)
> +{
> +	return 0;
> +}
> +
> +static inline int iothrottle_set_pagedirty_owner(struct page *page,
> +				struct mm_struct *mm)
> +{
> +	return 0;
> +}
> +
> +static inline int iothrottle_copy_page_owner(struct page *npage,
> +				struct page *opage)
> +{
> +	return 0;
> +}
> +
> +static inline int iothrottle_sync(void)
> +{
> +	return 0;
> +}
> +
> +static inline void set_in_aio(void) { }
> +
> +static inline void unset_in_aio(void) { }
> +
> +static inline int is_in_aio(void)
> +{
> +	return 0;
> +}
> +
> +static inline void set_in_dio(void) { }
> +
> +static inline void unset_in_dio(void) { }
> +
> +static inline int is_in_dio(void)
> +{
> +	return 0;
> +}
> +
> +static inline unsigned long long
> +get_io_throttle_cnt(struct task_struct *t, int type)
> +{
> +	return 0;
> +}
> +
> +static inline unsigned long long
> +get_io_throttle_sleep(struct task_struct *t, int type)
> +{
> +	return 0;
> +}
> +#endif /* CONFIG_CGROUP_IO_THROTTLE */
> +
> +static inline struct block_device *as_to_bdev(struct address_space *mapping)
> +{
> +	return (mapping->host && mapping->host->i_sb->s_bdev) ?
> +		mapping->host->i_sb->s_bdev : NULL;
> +}
> +
> +#endif /* BLK_IO_THROTTLE_H */
> diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
> index 665fa70..40cb412 100644
> --- a/include/linux/cgroup.h
> +++ b/include/linux/cgroup.h
> @@ -28,6 +28,7 @@ struct css_id;
>  extern int cgroup_init_early(void);
>  extern int cgroup_init(void);
>  extern void cgroup_lock(void);
> +extern int cgroup_is_locked(void);
>  extern bool cgroup_lock_live_group(struct cgroup *cgrp);
>  extern void cgroup_unlock(void);
>  extern void cgroup_fork(struct task_struct *p);
> diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h
> index 9c8d31b..c37cc4b 100644
> --- a/include/linux/cgroup_subsys.h
> +++ b/include/linux/cgroup_subsys.h
> @@ -43,6 +43,12 @@ SUBSYS(mem_cgroup)
> 
>  /* */
> 
> +#ifdef CONFIG_CGROUP_IO_THROTTLE
> +SUBSYS(iothrottle)
> +#endif
> +
> +/* */
> +
>  #ifdef CONFIG_CGROUP_DEVICE
>  SUBSYS(devices)
>  #endif
> diff --git a/init/Kconfig b/init/Kconfig
> index 5428ac7..d496c5f 100644
> --- a/init/Kconfig
> +++ b/init/Kconfig
> @@ -565,6 +565,18 @@ config RESOURCE_COUNTERS
>  	  infrastructure that works with cgroups.
>  	depends on CGROUPS
> 
> +config CGROUP_IO_THROTTLE
> +	bool "Enable cgroup I/O throttling"
> +	depends on CGROUPS && RESOURCE_COUNTERS && EXPERIMENTAL
> +	select MM_OWNER
> +	select PAGE_TRACKING
> +	help
> +	  This allows to limit the maximum I/O bandwidth for specific
> +	  cgroup(s).
> +	  See Documentation/cgroups/io-throttle.txt for more information.
> +
> +	  If unsure, say N.
> +
>  config CGROUP_MEM_RES_CTLR
>  	bool "Memory Resource Controller for Control Groups"
>  	depends on CGROUPS && RESOURCE_COUNTERS
> diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> index 382109b..5dbb2a7 100644
> --- a/kernel/cgroup.c
> +++ b/kernel/cgroup.c
> @@ -584,6 +584,15 @@ void cgroup_unlock(void)
>  	mutex_unlock(&cgroup_mutex);
>  }
> 
> +/**
> + * cgroup_is_locked - check if the cgroup mutex is locked
> + *
> + */
> +int cgroup_is_locked(void)
> +{
> +	return mutex_is_locked(&cgroup_mutex);
> +}
> +
>  /*
>   * A couple of forward declarations required, due to cyclic reference loop:
>   * cgroup_mkdir -> cgroup_create -> cgroup_populate_dir ->
> -- 
> 1.6.0.4
>