2008-06-07 08:29:51

by Andrea Righi

[permalink] [raw]
Subject: [PATCH 2/3] i/o bandwidth controller infrastructure

This is the core io-throttle kernel infrastructure. It creates the basic
interfaces to cgroups and implements the I/O measurement and throttling
functions.

Signed-off-by: Andrea Righi <[email protected]>
---
block/Makefile | 2 +
block/blk-io-throttle.c | 405 +++++++++++++++++++++++++++++++++++++++
include/linux/blk-io-throttle.h | 12 ++
include/linux/cgroup_subsys.h | 6 +
init/Kconfig | 10 +
5 files changed, 435 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 5a43c7d..8dec69b 100644
--- a/block/Makefile
+++ b/block/Makefile
@@ -14,3 +14,5 @@ obj-$(CONFIG_IOSCHED_CFQ) += cfq-iosched.o

obj-$(CONFIG_BLK_DEV_IO_TRACE) += blktrace.o
obj-$(CONFIG_BLOCK_COMPAT) += compat_ioctl.o
+
+obj-$(CONFIG_CGROUP_IO_THROTTLE) += blk-io-throttle.o
diff --git a/block/blk-io-throttle.c b/block/blk-io-throttle.c
new file mode 100644
index 0000000..804df88
--- /dev/null
+++ b/block/blk-io-throttle.c
@@ -0,0 +1,405 @@
+/*
+ * 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 <[email protected]>
+ */
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/cgroup.h>
+#include <linux/slab.h>
+#include <linux/gfp.h>
+#include <linux/err.h>
+#include <linux/sched.h>
+#include <linux/fs.h>
+#include <linux/jiffies.h>
+#include <linux/hardirq.h>
+#include <linux/spinlock.h>
+#include <linux/uaccess.h>
+#include <linux/blk-io-throttle.h>
+
+#define iothrottle_for_each(n, r) \
+ for (n = rb_first(r); n; n = rb_next(n))
+
+struct iothrottle_node {
+ struct rb_node node;
+ dev_t dev;
+ unsigned long iorate;
+ unsigned long req;
+ unsigned long last_request;
+};
+
+struct iothrottle {
+ struct cgroup_subsys_state css;
+ spinlock_t lock; /* protects the accounting of the cgroup i/o stats */
+ struct rb_root tree;
+};
+
+static inline struct iothrottle *cgroup_to_iothrottle(struct cgroup *cont)
+{
+ return container_of(cgroup_subsys_state(cont, iothrottle_subsys_id),
+ struct iothrottle, css);
+}
+
+static inline struct iothrottle *task_to_iothrottle(struct task_struct *task)
+{
+ return container_of(task_subsys_state(task, iothrottle_subsys_id),
+ struct iothrottle, css);
+}
+
+static inline struct iothrottle_node *iothrottle_search_node(
+ const struct iothrottle *iot,
+ dev_t dev)
+{
+ struct rb_node *node = (&iot->tree)->rb_node;
+
+ while (node) {
+ struct iothrottle_node *data = container_of(node,
+ struct iothrottle_node, node);
+ if (dev < data->dev)
+ node = node->rb_left;
+ else if (dev > data->dev)
+ node = node->rb_right;
+ else
+ return data;
+ }
+ return NULL;
+}
+
+static inline int iothrottle_insert_node(struct iothrottle *iot,
+ struct iothrottle_node *data)
+{
+ struct rb_root *root = &iot->tree;
+ struct rb_node **new = &(root->rb_node), *parent = NULL;
+
+ while (*new) {
+ struct iothrottle_node *this = container_of(*new,
+ struct iothrottle_node, node);
+ parent = *new;
+ if (data->dev < this->dev)
+ new = &((*new)->rb_left);
+ else if (data->dev > this->dev)
+ new = &((*new)->rb_right);
+ else
+ return -EINVAL;
+ }
+ rb_link_node(&data->node, parent, new);
+ rb_insert_color(&data->node, root);
+
+ return 0;
+}
+
+static inline void iothrottle_delete_node(struct iothrottle *iot, dev_t dev)
+{
+ struct iothrottle_node *data = iothrottle_search_node(iot, dev);
+
+ if (likely(data)) {
+ rb_erase(&data->node, &iot->tree);
+ kfree(data);
+ }
+}
+
+/*
+ * Note: called from kernel/cgroup.c with cgroup_lock() held.
+ */
+static struct cgroup_subsys_state *iothrottle_create(
+ struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ struct iothrottle *iot;
+
+ iot = kmalloc(sizeof(*iot), GFP_KERNEL);
+ if (unlikely(!iot))
+ return ERR_PTR(-ENOMEM);
+
+ spin_lock_init(&iot->lock);
+ iot->tree = RB_ROOT;
+
+ return &iot->css;
+}
+
+/*
+ * Note: called from kernel/cgroup.c with cgroup_lock() held.
+ */
+static void iothrottle_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ struct iothrottle_node *data;
+ struct rb_node *next;
+ struct iothrottle *iot = cgroup_to_iothrottle(cont);
+
+ next = rb_first(&iot->tree);
+ while (next) {
+ data = rb_entry(next, struct iothrottle_node, node);
+ next = rb_next(&data->node);
+ rb_erase(&data->node, &iot->tree);
+ kfree(data);
+ }
+ kfree(iot);
+}
+
+static ssize_t iothrottle_read(struct cgroup *cont,
+ struct cftype *cft,
+ struct file *file,
+ char __user *userbuf,
+ size_t nbytes,
+ loff_t *ppos)
+{
+ struct iothrottle *iot;
+ char *buffer, *s;
+ struct rb_node *n;
+ ssize_t ret;
+
+ buffer = kmalloc(nbytes + 1, GFP_KERNEL);
+ if (unlikely(!buffer))
+ return -ENOMEM;
+
+ cgroup_lock();
+ if (cgroup_is_removed(cont)) {
+ ret = -ENODEV;
+ goto out;
+ }
+
+ iot = cgroup_to_iothrottle(cont);
+
+ s = buffer;
+ spin_lock_irq(&iot->lock);
+ iothrottle_for_each(n, &iot->tree) {
+ struct iothrottle_node *node =
+ rb_entry(n, struct iothrottle_node, node);
+ unsigned long delta = (long)jiffies - (long)node->last_request;
+
+ BUG_ON(!node->dev);
+ s += snprintf(s, nbytes - (s - buffer),
+ "=== device (%u,%u) ===\n"
+ "bandwidth-max: %lu KiB/sec\n"
+ " requested: %lu bytes\n"
+ " last request: %lu jiffies\n"
+ " delta: %lu jiffies\n",
+ MAJOR(node->dev), MINOR(node->dev),
+ node->iorate, node->req,
+ node->last_request, delta);
+ }
+ spin_unlock_irq(&iot->lock);
+ buffer[nbytes] = '\0';
+
+ ret = simple_read_from_buffer(userbuf, nbytes,
+ ppos, buffer, (s - buffer));
+out:
+ cgroup_unlock();
+ kfree(buffer);
+ return ret;
+}
+
+static inline dev_t devname2dev_t(const char *buf)
+{
+ struct block_device *bdev;
+ dev_t ret;
+
+ bdev = lookup_bdev(buf);
+ if (IS_ERR(bdev))
+ return 0;
+
+ BUG_ON(!bdev->bd_inode);
+ ret = bdev->bd_inode->i_rdev;
+ bdput(bdev);
+
+ return ret;
+}
+
+static inline int iothrottle_parse_args(char *buf, size_t nbytes,
+ dev_t *dev, unsigned long *val)
+{
+ char *p;
+
+ p = memchr(buf, ':', nbytes);
+ if (!p)
+ return -EINVAL;
+ *p++ = '\0';
+
+ *dev = devname2dev_t(buf);
+ if (!*dev)
+ return -ENOTBLK;
+
+ return strict_strtoul(p, 10, val);
+}
+
+static ssize_t iothrottle_write(struct cgroup *cont,
+ struct cftype *cft,
+ struct file *file,
+ const char __user *userbuf,
+ size_t nbytes, loff_t *ppos)
+{
+ struct iothrottle *iot;
+ struct iothrottle_node *node, *tmpn = NULL;
+ char *buffer, *tmpp;
+ dev_t dev;
+ unsigned long val;
+ int ret;
+
+ if (unlikely(!nbytes))
+ return -EINVAL;
+
+ buffer = kmalloc(nbytes + 1, GFP_KERNEL);
+ if (unlikely(!buffer))
+ return -ENOMEM;
+
+ if (copy_from_user(buffer, userbuf, nbytes)) {
+ ret = -EFAULT;
+ goto out1;
+ }
+
+ buffer[nbytes] = '\0';
+ tmpp = strstrip(buffer);
+
+ ret = iothrottle_parse_args(tmpp, nbytes, &dev, &val);
+ if (ret)
+ goto out1;
+
+ /*
+ * Pre-allocate a temporary node structure outside locks to use
+ * GFP_KERNEL, it will be kfree()ed later if unused.
+ */
+ tmpn = kmalloc(sizeof(*tmpn), GFP_KERNEL);
+
+ cgroup_lock();
+ if (cgroup_is_removed(cont)) {
+ ret = -ENODEV;
+ goto out2;
+ }
+
+ iot = cgroup_to_iothrottle(cont);
+
+ spin_lock_irq(&iot->lock);
+ if (!val) {
+ /* Delete a block device limiting rule */
+ iothrottle_delete_node(iot, dev);
+ ret = nbytes;
+ goto out3;
+ }
+ node = iothrottle_search_node(iot, dev);
+ if (node) {
+ /* Update a block device limiting rule */
+ node->iorate = val;
+ node->req = 0;
+ node->last_request = jiffies;
+ ret = nbytes;
+ goto out3;
+ }
+ /* Add a new block device limiting rule */
+ if (unlikely(!tmpn)) {
+ ret = -ENOMEM;
+ goto out3;
+ }
+ node = tmpn;
+ tmpn = NULL;
+
+ node->iorate = val;
+ node->req = 0;
+ node->last_request = jiffies;
+ node->dev = dev;
+ ret = iothrottle_insert_node(iot, node);
+ BUG_ON(ret);
+ ret = nbytes;
+out3:
+ spin_unlock_irq(&iot->lock);
+out2:
+ cgroup_unlock();
+ if (tmpn)
+ kfree(tmpn);
+out1:
+ kfree(buffer);
+ return ret;
+}
+
+static struct cftype files[] = {
+ {
+ .name = "bandwidth",
+ .read = iothrottle_read,
+ .write = iothrottle_write,
+ },
+};
+
+static int iothrottle_populate(struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ return cgroup_add_files(cont, 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,
+};
+
+static inline int __cant_sleep(void)
+{
+ return in_atomic() || in_interrupt() || irqs_disabled();
+}
+
+void cgroup_io_account(struct block_device *bdev, size_t bytes)
+{
+ struct iothrottle *iot;
+ struct iothrottle_node *node;
+ unsigned long delta, t;
+ long sleep;
+
+ if (unlikely(!bdev))
+ return;
+
+ BUG_ON(!bdev->bd_inode);
+
+ iot = task_to_iothrottle(current);
+ if (unlikely(!iot))
+ return;
+
+ spin_lock_irq(&iot->lock);
+
+ node = iothrottle_search_node(iot, bdev->bd_inode->i_rdev);
+ if (!node || !node->iorate)
+ goto out;
+
+ /* Account the I/O activity */
+ node->req += bytes;
+
+ /* Evaluate if we need to throttle the current process */
+ delta = (long)jiffies - (long)node->last_request;
+ if (!delta)
+ goto out;
+
+ t = msecs_to_jiffies(node->req / node->iorate);
+ if (!t)
+ goto out;
+
+ sleep = t - delta;
+ if (unlikely(sleep > 0)) {
+ spin_unlock_irq(&iot->lock);
+ if (__cant_sleep())
+ return;
+ pr_debug("io-throttle: task %p (%s) must sleep %lu jiffies\n",
+ current, current->comm, sleep);
+ schedule_timeout_killable(sleep);
+ return;
+ }
+
+ /* Reset I/O accounting */
+ node->req = 0;
+ node->last_request = jiffies;
+out:
+ spin_unlock_irq(&iot->lock);
+}
+EXPORT_SYMBOL(cgroup_io_account);
diff --git a/include/linux/blk-io-throttle.h b/include/linux/blk-io-throttle.h
new file mode 100644
index 0000000..cff0c13
--- /dev/null
+++ b/include/linux/blk-io-throttle.h
@@ -0,0 +1,12 @@
+#ifndef BLK_IO_THROTTLE_H
+#define BLK_IO_THROTTLE_H
+
+#ifdef CONFIG_CGROUP_IO_THROTTLE
+extern void cgroup_io_account(struct block_device *bdev, size_t bytes);
+#else
+static inline void cgroup_io_account(struct block_device *bdev, size_t bytes)
+{
+}
+#endif /* CONFIG_CGROUP_IO_THROTTLE */
+
+#endif /* BLK_IO_THROTTLE_H */
diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h
index e287745..0caf3c2 100644
--- a/include/linux/cgroup_subsys.h
+++ b/include/linux/cgroup_subsys.h
@@ -48,3 +48,9 @@ SUBSYS(devices)
#endif

/* */
+
+#ifdef CONFIG_CGROUP_IO_THROTTLE
+SUBSYS(iothrottle)
+#endif
+
+/* */
diff --git a/init/Kconfig b/init/Kconfig
index 6199d11..3117d99 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -306,6 +306,16 @@ config CGROUP_DEVICE
Provides a cgroup implementing whitelists for devices which
a process in the cgroup can mknod or open.

+config CGROUP_IO_THROTTLE
+ bool "Enable cgroup I/O throttling (EXPERIMENTAL)"
+ depends on CGROUPS && EXPERIMENTAL
+ help
+ This allows to limit the maximum I/O bandwidth for specific
+ cgroup(s).
+ See Documentation/controllers/io-throttle.txt for more information.
+
+ If unsure, say N.
+
config CPUSETS
bool "Cpuset support"
depends on SMP && CGROUPS
--
1.5.4.3


2008-06-18 18:06:30

by Carl Henrik Lunde

[permalink] [raw]
Subject: Re: [PATCH 2/3] i/o bandwidth controller infrastructure

On Sat, Jun 07, 2008 at 12:27:29AM +0200, Andrea Righi wrote:
> This is the core io-throttle kernel infrastructure. It creates the basic
> interfaces to cgroups and implements the I/O measurement and throttling
> functions.
[...]
> +void cgroup_io_account(struct block_device *bdev, size_t bytes)
[...]
> + /* Account the I/O activity */
> + node->req += bytes;
> +
> + /* Evaluate if we need to throttle the current process */
> + delta = (long)jiffies - (long)node->last_request;
> + if (!delta)
> + goto out;
> +
> + t = msecs_to_jiffies(node->req / node->iorate);
> + if (!t)
> + goto out;
> +
> + sleep = t - delta;
> + if (unlikely(sleep > 0)) {
> + spin_unlock_irq(&iot->lock);
> + if (__cant_sleep())
> + return;
> + pr_debug("io-throttle: task %p (%s) must sleep %lu jiffies\n",
> + current, current->comm, sleep);
> + schedule_timeout_killable(sleep);
> + return;
> + }
> +
> + /* Reset I/O accounting */
> + node->req = 0;
> + node->last_request = jiffies;
[...]

Did you consider using token bucket instead of this (leaky bucket?)?

I've attached a patch which implements token bucket. Although not as
precise as the leaky bucket the performance is better at high bandwidth
streaming loads.

The leaky bucket stops at around 53 MB/s while token bucket works for
up to 64 MB/s. The baseline (no cgroups) is 66 MB/s.

benchmark:
two streaming readers (fio) with block size 128k, bucket size 4 MB
90% of the bandwidth was allocated to one process, the other gets 10%

bw-limit: actual bw algorithm bw1 bw2
5 MiB/s: 5.0 MiB/s leaky_bucket 0.5 4.5
5 MiB/s: 5.2 MiB/s token_bucket 0.6 4.6
10 MiB/s: 10.0 MiB/s leaky_bucket 1.0 9.0
10 MiB/s: 10.3 MiB/s token_bucket 1.0 9.2
15 MiB/s: 15.0 MiB/s leaky_bucket 1.5 13.5
15 MiB/s: 15.4 MiB/s token_bucket 1.5 13.8
20 MiB/s: 19.9 MiB/s leaky_bucket 2.0 17.9
20 MiB/s: 20.5 MiB/s token_bucket 2.1 18.4
25 MiB/s: 24.4 MiB/s leaky_bucket 2.5 21.9
25 MiB/s: 25.6 MiB/s token_bucket 2.6 23.0
30 MiB/s: 29.2 MiB/s leaky_bucket 3.0 26.2
30 MiB/s: 30.7 MiB/s token_bucket 3.1 27.7
35 MiB/s: 34.3 MiB/s leaky_bucket 3.4 30.9
35 MiB/s: 35.9 MiB/s token_bucket 3.6 32.3
40 MiB/s: 39.7 MiB/s leaky_bucket 3.9 35.8
40 MiB/s: 41.0 MiB/s token_bucket 4.1 36.9
45 MiB/s: 44.0 MiB/s leaky_bucket 4.3 39.7
45 MiB/s: 46.1 MiB/s token_bucket 4.6 41.5
50 MiB/s: 47.9 MiB/s leaky_bucket 4.7 43.2
50 MiB/s: 51.0 MiB/s token_bucket 5.1 45.9
55 MiB/s: 50.5 MiB/s leaky_bucket 5.0 45.5
55 MiB/s: 56.2 MiB/s token_bucket 5.6 50.5
60 MiB/s: 52.9 MiB/s leaky_bucket 5.2 47.7
60 MiB/s: 61.0 MiB/s token_bucket 6.1 54.9
65 MiB/s: 53.0 MiB/s leaky_bucket 5.4 47.6
65 MiB/s: 63.7 MiB/s token_bucket 6.6 57.1
70 MiB/s: 53.8 MiB/s leaky_bucket 5.5 48.4
70 MiB/s: 64.1 MiB/s token_bucket 7.1 57.0


diff --git a/block/blk-io-throttle.c b/block/blk-io-throttle.c
index 804df88..9ed0c7c 100644
--- a/block/blk-io-throttle.c
+++ b/block/blk-io-throttle.c
@@ -40,7 +40,8 @@ struct iothrottle_node {
struct rb_node node;
dev_t dev;
unsigned long iorate;
- unsigned long req;
+ long bucket_size; /* Max value for t */
+ long t;
unsigned long last_request;
};

@@ -180,18 +181,20 @@ static ssize_t iothrottle_read(struct cgroup *cont,
iothrottle_for_each(n, &iot->tree) {
struct iothrottle_node *node =
rb_entry(n, struct iothrottle_node, node);
- unsigned long delta = (long)jiffies - (long)node->last_request;
+ unsigned long delta = (((long)jiffies - (long)node->last_request) * 1000) / HZ;

BUG_ON(!node->dev);
s += snprintf(s, nbytes - (s - buffer),
"=== device (%u,%u) ===\n"
"bandwidth-max: %lu KiB/sec\n"
- " requested: %lu bytes\n"
- " last request: %lu jiffies\n"
- " delta: %lu jiffies\n",
+ "bucket size : %ld KiB\n"
+ "bucket fill : %ld KiB (after last request)\n"
+ "last request : %lu ms ago\n",
MAJOR(node->dev), MINOR(node->dev),
- node->iorate, node->req,
- node->last_request, delta);
+ node->iorate,
+ node->bucket_size / 1024,
+ node->t / 1024,
+ delta);
}
spin_unlock_irq(&iot->lock);
buffer[nbytes] = '\0';
@@ -220,21 +223,33 @@ static inline dev_t devname2dev_t(const char *buf)
return ret;
}

-static inline int iothrottle_parse_args(char *buf, size_t nbytes,
- dev_t *dev, unsigned long *val)
+static inline int iothrottle_parse_args(char *buf, size_t nbytes, dev_t *dev,
+ unsigned long *iorate,
+ unsigned long *bucket_size)
{
- char *p;
+ char *ioratep, *bucket_sizep;

- p = memchr(buf, ':', nbytes);
- if (!p)
+ ioratep = memchr(buf, ':', nbytes);
+ if (!ioratep)
return -EINVAL;
- *p++ = '\0';
+ *ioratep++ = '\0';
+
+ bucket_sizep = memchr(ioratep, ':', nbytes + ioratep - buf);
+ if (!bucket_sizep)
+ return -EINVAL;
+ *bucket_sizep++ = '\0';

*dev = devname2dev_t(buf);
if (!*dev)
return -ENOTBLK;

- return strict_strtoul(p, 10, val);
+ if (strict_strtoul(ioratep, 10, iorate))
+ return -EINVAL;
+
+ if (strict_strtoul(bucket_sizep, 10, bucket_size))
+ return -EINVAL;
+
+ return 0;
}

static ssize_t iothrottle_write(struct cgroup *cont,
@@ -247,7 +262,7 @@ static ssize_t iothrottle_write(struct cgroup *cont,
struct iothrottle_node *node, *tmpn = NULL;
char *buffer, *tmpp;
dev_t dev;
- unsigned long val;
+ unsigned long iorate, bucket_size;
int ret;

if (unlikely(!nbytes))
@@ -265,7 +280,7 @@ static ssize_t iothrottle_write(struct cgroup *cont,
buffer[nbytes] = '\0';
tmpp = strstrip(buffer);

- ret = iothrottle_parse_args(tmpp, nbytes, &dev, &val);
+ ret = iothrottle_parse_args(tmpp, nbytes, &dev, &iorate, &bucket_size);
if (ret)
goto out1;

@@ -284,7 +299,7 @@ static ssize_t iothrottle_write(struct cgroup *cont,
iot = cgroup_to_iothrottle(cont);

spin_lock_irq(&iot->lock);
- if (!val) {
+ if (!iorate) {
/* Delete a block device limiting rule */
iothrottle_delete_node(iot, dev);
ret = nbytes;
@@ -293,8 +308,9 @@ static ssize_t iothrottle_write(struct cgroup *cont,
node = iothrottle_search_node(iot, dev);
if (node) {
/* Update a block device limiting rule */
- node->iorate = val;
- node->req = 0;
+ node->iorate = iorate;
+ node->bucket_size = bucket_size * 1024;
+ node->t = 0;
node->last_request = jiffies;
ret = nbytes;
goto out3;
@@ -307,8 +323,9 @@ static ssize_t iothrottle_write(struct cgroup *cont,
node = tmpn;
tmpn = NULL;

- node->iorate = val;
- node->req = 0;
+ node->iorate = iorate;
+ node->bucket_size = bucket_size * 1024;
+ node->t = 0;
node->last_request = jiffies;
node->dev = dev;
ret = iothrottle_insert_node(iot, node);
@@ -355,7 +372,7 @@ void cgroup_io_account(struct block_device *bdev, size_t bytes)
{
struct iothrottle *iot;
struct iothrottle_node *node;
- unsigned long delta, t;
+ unsigned long delta;
long sleep;

if (unlikely(!bdev))
@@ -370,36 +387,37 @@ void cgroup_io_account(struct block_device *bdev, size_t bytes)
spin_lock_irq(&iot->lock);

node = iothrottle_search_node(iot, bdev->bd_inode->i_rdev);
- if (!node || !node->iorate)
- goto out;
-
- /* Account the I/O activity */
- node->req += bytes;
+ if (!node || !node->iorate) {
+ spin_unlock_irq(&iot->lock);
+ return;
+ }

- /* Evaluate if we need to throttle the current process */
+ /* Add tokens for time elapsed since last read */
delta = (long)jiffies - (long)node->last_request;
- if (!delta)
- goto out;
+ if (delta) {
+ node->last_request = jiffies;
+ node->t += (node->iorate * 1024 * delta) / HZ;

- t = msecs_to_jiffies(node->req / node->iorate);
- if (!t)
- goto out;
+ if (node->t > node->bucket_size)
+ node->t = node->bucket_size;
+ }

- sleep = t - delta;
- if (unlikely(sleep > 0)) {
- spin_unlock_irq(&iot->lock);
- if (__cant_sleep())
- return;
- pr_debug("io-throttle: task %p (%s) must sleep %lu jiffies\n",
- current, current->comm, sleep);
- schedule_timeout_killable(sleep);
- return;
+ /* Account the I/O activity */
+ node->t -= bytes;
+
+ if (node->t < 0) {
+ sleep = (-node->t) * HZ / (node->iorate * 1024);
+ } else {
+ sleep = 0;
}

- /* Reset I/O accounting */
- node->req = 0;
- node->last_request = jiffies;
-out:
spin_unlock_irq(&iot->lock);
+
+ if (sleep && !__cant_sleep()) {
+ pr_debug("io-throttle: %s[%d] must sleep %ld jiffies\n",
+ current->comm, current->pid, sleep);
+
+ schedule_timeout_killable(sleep);
+ }
}
EXPORT_SYMBOL(cgroup_io_account);

--
Carl Henrik

2008-06-18 22:31:39

by Andrea Righi

[permalink] [raw]
Subject: Re: [PATCH 2/3] i/o bandwidth controller infrastructure

Carl Henrik Lunde wrote:
> On Sat, Jun 07, 2008 at 12:27:29AM +0200, Andrea Righi wrote:
>> This is the core io-throttle kernel infrastructure. It creates the basic
>> interfaces to cgroups and implements the I/O measurement and throttling
>> functions.
> [...]
>> +void cgroup_io_account(struct block_device *bdev, size_t bytes)
> [...]
>> + /* Account the I/O activity */
>> + node->req += bytes;
>> +
>> + /* Evaluate if we need to throttle the current process */
>> + delta = (long)jiffies - (long)node->last_request;
>> + if (!delta)
>> + goto out;
>> +
>> + t = msecs_to_jiffies(node->req / node->iorate);
>> + if (!t)
>> + goto out;
>> +
>> + sleep = t - delta;
>> + if (unlikely(sleep > 0)) {
>> + spin_unlock_irq(&iot->lock);
>> + if (__cant_sleep())
>> + return;
>> + pr_debug("io-throttle: task %p (%s) must sleep %lu jiffies\n",
>> + current, current->comm, sleep);
>> + schedule_timeout_killable(sleep);
>> + return;
>> + }
>> +
>> + /* Reset I/O accounting */
>> + node->req = 0;
>> + node->last_request = jiffies;
> [...]
>
> Did you consider using token bucket instead of this (leaky bucket?)?
>
> I've attached a patch which implements token bucket. Although not as
> precise as the leaky bucket the performance is better at high bandwidth
> streaming loads.

Interesting! it could be great to have both available at runtime and
just switch between leaky or token bucket, e.g. by echo-ing "leaky" or
"token" to a file in the cgroup filesystem, ummm, block.limiting-algorithm?

>
> The leaky bucket stops at around 53 MB/s while token bucket works for
> up to 64 MB/s. The baseline (no cgroups) is 66 MB/s.
>
> benchmark:
> two streaming readers (fio) with block size 128k, bucket size 4 MB
> 90% of the bandwidth was allocated to one process, the other gets 10%

Thanks for posting the results, I'll look closely at your patch, test
it as well and try merge your work.

I also did some improvements in v2 in terms of scalability, in
particular I've replaced the rbtree with a liked list, in order to
remove the spinlocks and replace them by RCU to protect the list
structure. I need to do some stress tests before, but I'll post a v3
soon.

Some minor comments below for now.

> diff --git a/block/blk-io-throttle.c b/block/blk-io-throttle.c
> index 804df88..9ed0c7c 100644
> --- a/block/blk-io-throttle.c
> +++ b/block/blk-io-throttle.c
> @@ -40,7 +40,8 @@ struct iothrottle_node {
> struct rb_node node;
> dev_t dev;
> unsigned long iorate;
> - unsigned long req;
> + long bucket_size; /* Max value for t */
> + long t;
> unsigned long last_request;
> };
>
> @@ -180,18 +181,20 @@ static ssize_t iothrottle_read(struct cgroup *cont,
> iothrottle_for_each(n, &iot->tree) {
> struct iothrottle_node *node =
> rb_entry(n, struct iothrottle_node, node);
> - unsigned long delta = (long)jiffies - (long)node->last_request;
> + unsigned long delta = (((long)jiffies - (long)node->last_request) * 1000) / HZ;

Better to use jiffies_to_msecs() here.

>
> BUG_ON(!node->dev);
> s += snprintf(s, nbytes - (s - buffer),
> "=== device (%u,%u) ===\n"
> "bandwidth-max: %lu KiB/sec\n"
> - " requested: %lu bytes\n"
> - " last request: %lu jiffies\n"
> - " delta: %lu jiffies\n",
> + "bucket size : %ld KiB\n"
> + "bucket fill : %ld KiB (after last request)\n"
> + "last request : %lu ms ago\n",
> MAJOR(node->dev), MINOR(node->dev),
> - node->iorate, node->req,
> - node->last_request, delta);
> + node->iorate,
> + node->bucket_size / 1024,
> + node->t / 1024,
> + delta);
> }
> spin_unlock_irq(&iot->lock);
> buffer[nbytes] = '\0';
> @@ -220,21 +223,33 @@ static inline dev_t devname2dev_t(const char *buf)
> return ret;
> }
>
> -static inline int iothrottle_parse_args(char *buf, size_t nbytes,
> - dev_t *dev, unsigned long *val)
> +static inline int iothrottle_parse_args(char *buf, size_t nbytes, dev_t *dev,
> + unsigned long *iorate,
> + unsigned long *bucket_size)
> {
> - char *p;
> + char *ioratep, *bucket_sizep;
>
> - p = memchr(buf, ':', nbytes);
> - if (!p)
> + ioratep = memchr(buf, ':', nbytes);
> + if (!ioratep)
> return -EINVAL;
> - *p++ = '\0';
> + *ioratep++ = '\0';
> +
> + bucket_sizep = memchr(ioratep, ':', nbytes + ioratep - buf);
> + if (!bucket_sizep)
> + return -EINVAL;
> + *bucket_sizep++ = '\0';
>
> *dev = devname2dev_t(buf);
> if (!*dev)
> return -ENOTBLK;
>
> - return strict_strtoul(p, 10, val);
> + if (strict_strtoul(ioratep, 10, iorate))
> + return -EINVAL;
> +
> + if (strict_strtoul(bucket_sizep, 10, bucket_size))
> + return -EINVAL;
> +
> + return 0;
> }
>
> static ssize_t iothrottle_write(struct cgroup *cont,
> @@ -247,7 +262,7 @@ static ssize_t iothrottle_write(struct cgroup *cont,
> struct iothrottle_node *node, *tmpn = NULL;
> char *buffer, *tmpp;
> dev_t dev;
> - unsigned long val;
> + unsigned long iorate, bucket_size;
> int ret;
>
> if (unlikely(!nbytes))
> @@ -265,7 +280,7 @@ static ssize_t iothrottle_write(struct cgroup *cont,
> buffer[nbytes] = '\0';
> tmpp = strstrip(buffer);
>
> - ret = iothrottle_parse_args(tmpp, nbytes, &dev, &val);
> + ret = iothrottle_parse_args(tmpp, nbytes, &dev, &iorate, &bucket_size);
> if (ret)
> goto out1;
>
> @@ -284,7 +299,7 @@ static ssize_t iothrottle_write(struct cgroup *cont,
> iot = cgroup_to_iothrottle(cont);
>
> spin_lock_irq(&iot->lock);
> - if (!val) {
> + if (!iorate) {
> /* Delete a block device limiting rule */
> iothrottle_delete_node(iot, dev);
> ret = nbytes;
> @@ -293,8 +308,9 @@ static ssize_t iothrottle_write(struct cgroup *cont,
> node = iothrottle_search_node(iot, dev);
> if (node) {
> /* Update a block device limiting rule */
> - node->iorate = val;
> - node->req = 0;
> + node->iorate = iorate;
> + node->bucket_size = bucket_size * 1024;
> + node->t = 0;
> node->last_request = jiffies;
> ret = nbytes;
> goto out3;
> @@ -307,8 +323,9 @@ static ssize_t iothrottle_write(struct cgroup *cont,
> node = tmpn;
> tmpn = NULL;
>
> - node->iorate = val;
> - node->req = 0;
> + node->iorate = iorate;
> + node->bucket_size = bucket_size * 1024;
> + node->t = 0;
> node->last_request = jiffies;
> node->dev = dev;
> ret = iothrottle_insert_node(iot, node);
> @@ -355,7 +372,7 @@ void cgroup_io_account(struct block_device *bdev, size_t bytes)
> {
> struct iothrottle *iot;
> struct iothrottle_node *node;
> - unsigned long delta, t;
> + unsigned long delta;
> long sleep;
>
> if (unlikely(!bdev))
> @@ -370,36 +387,37 @@ void cgroup_io_account(struct block_device *bdev, size_t bytes)
> spin_lock_irq(&iot->lock);
>
> node = iothrottle_search_node(iot, bdev->bd_inode->i_rdev);
> - if (!node || !node->iorate)
> - goto out;
> -
> - /* Account the I/O activity */
> - node->req += bytes;
> + if (!node || !node->iorate) {
> + spin_unlock_irq(&iot->lock);
> + return;
> + }
>
> - /* Evaluate if we need to throttle the current process */
> + /* Add tokens for time elapsed since last read */
> delta = (long)jiffies - (long)node->last_request;
> - if (!delta)
> - goto out;
> + if (delta) {
> + node->last_request = jiffies;
> + node->t += (node->iorate * 1024 * delta) / HZ;

The same here:
node->t += node->iorate * 1024
* jiffies_to_msec(delta) * MSEC_PER_SEC;

>
> - t = msecs_to_jiffies(node->req / node->iorate);
> - if (!t)
> - goto out;
> + if (node->t > node->bucket_size)
> + node->t = node->bucket_size;
> + }
>
> - sleep = t - delta;
> - if (unlikely(sleep > 0)) {
> - spin_unlock_irq(&iot->lock);
> - if (__cant_sleep())
> - return;
> - pr_debug("io-throttle: task %p (%s) must sleep %lu jiffies\n",
> - current, current->comm, sleep);
> - schedule_timeout_killable(sleep);
> - return;
> + /* Account the I/O activity */
> + node->t -= bytes;
> +
> + if (node->t < 0) {
> + sleep = (-node->t) * HZ / (node->iorate * 1024);

And again:
sleep = msec_to_jiffies(-node->t / (node->iorate * 1024)
* MSEC_PER_SEC);

> + } else {
> + sleep = 0;
> }
>
> - /* Reset I/O accounting */
> - node->req = 0;
> - node->last_request = jiffies;
> -out:
> spin_unlock_irq(&iot->lock);
> +
> + if (sleep && !__cant_sleep()) {
> + pr_debug("io-throttle: %s[%d] must sleep %ld jiffies\n",
> + current->comm, current->pid, sleep);
> +
> + schedule_timeout_killable(sleep);
> + }
> }
> EXPORT_SYMBOL(cgroup_io_account);

Thanks,
-Andrea

2008-06-22 13:11:31

by Andrea Righi

[permalink] [raw]
Subject: Re: [PATCH 2/3] i/o bandwidth controller infrastructure

Carl Henrik Lunde wrote:
> Did you consider using token bucket instead of this (leaky bucket?)?
>
> I've attached a patch which implements token bucket. Although not as
> precise as the leaky bucket the performance is better at high bandwidth
> streaming loads.
>
> The leaky bucket stops at around 53 MB/s while token bucket works for
> up to 64 MB/s. The baseline (no cgroups) is 66 MB/s.
>
> benchmark:
> two streaming readers (fio) with block size 128k, bucket size 4 MB
> 90% of the bandwidth was allocated to one process, the other gets 10%
>
> bw-limit: actual bw algorithm bw1 bw2
> 5 MiB/s: 5.0 MiB/s leaky_bucket 0.5 4.5
> 5 MiB/s: 5.2 MiB/s token_bucket 0.6 4.6
> 10 MiB/s: 10.0 MiB/s leaky_bucket 1.0 9.0
> 10 MiB/s: 10.3 MiB/s token_bucket 1.0 9.2
> 15 MiB/s: 15.0 MiB/s leaky_bucket 1.5 13.5
> 15 MiB/s: 15.4 MiB/s token_bucket 1.5 13.8
> 20 MiB/s: 19.9 MiB/s leaky_bucket 2.0 17.9
> 20 MiB/s: 20.5 MiB/s token_bucket 2.1 18.4
> 25 MiB/s: 24.4 MiB/s leaky_bucket 2.5 21.9
> 25 MiB/s: 25.6 MiB/s token_bucket 2.6 23.0
> 30 MiB/s: 29.2 MiB/s leaky_bucket 3.0 26.2
> 30 MiB/s: 30.7 MiB/s token_bucket 3.1 27.7
> 35 MiB/s: 34.3 MiB/s leaky_bucket 3.4 30.9
> 35 MiB/s: 35.9 MiB/s token_bucket 3.6 32.3
> 40 MiB/s: 39.7 MiB/s leaky_bucket 3.9 35.8
> 40 MiB/s: 41.0 MiB/s token_bucket 4.1 36.9
> 45 MiB/s: 44.0 MiB/s leaky_bucket 4.3 39.7
> 45 MiB/s: 46.1 MiB/s token_bucket 4.6 41.5
> 50 MiB/s: 47.9 MiB/s leaky_bucket 4.7 43.2
> 50 MiB/s: 51.0 MiB/s token_bucket 5.1 45.9
> 55 MiB/s: 50.5 MiB/s leaky_bucket 5.0 45.5
> 55 MiB/s: 56.2 MiB/s token_bucket 5.6 50.5
> 60 MiB/s: 52.9 MiB/s leaky_bucket 5.2 47.7
> 60 MiB/s: 61.0 MiB/s token_bucket 6.1 54.9
> 65 MiB/s: 53.0 MiB/s leaky_bucket 5.4 47.6
> 65 MiB/s: 63.7 MiB/s token_bucket 6.6 57.1
> 70 MiB/s: 53.8 MiB/s leaky_bucket 5.5 48.4
> 70 MiB/s: 64.1 MiB/s token_bucket 7.1 57.0

Carl,

based on your token bucket solution I've implemented a run-time leaky
bucket / token bucket switcher:

# leaky bucket #
echo 0 > /cgroups/foo/blockio.throttling_strategy
# token bucket #
echo 1 > /cgroups/foo/blockio.throttling_strategy

The -rc of the new io-throttle patch 2/3 is below, 1/3 and 3/3 are the
same as patchset version 3, even if documentation must be updated. It
would be great if you could review the patch, in particular the
token_bucket() implementation and repeat your tests.

The all-in-one patch is available here:
http://download.systemimager.org/~arighi/linux/patches/io-throttle/cgroup-io-throttle-v4-rc1.patch

I also did some quick tests similar to yours, the benchmark I've used is
available here as well:
http://download.systemimager.org/~arighi/linux/patches/io-throttle/benchmark/iobw.c

= Results =
I/O scheduler: cfq
filesystem: ext3
Command: ionice -c 1 -n 0 iobw -direct 2 4m 32m
Bucket size: 4MiB

testing 2 parallel streams, chunk_size 4096KiB, data_size 32768KiB

=== no throttling ===
testing 2 parallel streams, chunk_size 4096KiB, data_size 32768KiB
[task 2] time: 2.929, bw: 10742 KiB/s (WRITE)
[task 2] time: 2.878, bw: 10742 KiB/s (READ )
[task 1] time: 2.377, bw: 13671 KiB/s (WRITE)
[task 1] time: 3.979, bw: 7812 KiB/s (READ )
[parent 0] time: 6.397, bw: 19531 KiB/s (TOTAL)

=== bandwidth limit: 4MiB/s (leaky bucket) ===
[task 2] time: 15.880, bw: 1953 KiB/s (WRITE)
[task 2] time: 14.278, bw: 1953 KiB/s (READ )
[task 1] time: 14.711, bw: 1953 KiB/s (WRITE)
[task 1] time: 16.563, bw: 1953 KiB/s (READ )
[parent 0] time: 31.316, bw: 3906 KiB/s (TOTAL)

=== bandwidth limit: 4MiB/s (token bucket) ===
[task 2] time: 11.864, bw: 1953 KiB/s (WRITE)
[task 2] time: 15.958, bw: 1953 KiB/s (READ )
[task 1] time: 19.233, bw: 976 KiB/s (WRITE)
[task 1] time: 12.643, bw: 1953 KiB/s (READ )
[parent 0] time: 31.917, bw: 3906 KiB/s (TOTAL)

=== bandwidth limit: 8MiB/s (leaky bucket) ===
[task 2] time: 7.198, bw: 3906 KiB/s (WRITE)
[task 2] time: 8.012, bw: 3906 KiB/s (READ )
[task 1] time: 7.891, bw: 3906 KiB/s (WRITE)
[task 1] time: 7.846, bw: 3906 KiB/s (READ )
[parent 0] time: 15.780, bw: 7812 KiB/s (TOTAL)

=== bandwidth limit: 8MiB/s (token bucket) ===
[task 1] time: 6.996, bw: 3906 KiB/s (WRITE)
[task 1] time: 6.529, bw: 4882 KiB/s (READ )
[task 2] time: 10.341, bw: 2929 KiB/s (WRITE)
[task 2] time: 5.681, bw: 4882 KiB/s (READ )
[parent 0] time: 16.079, bw: 7812 KiB/s (TOTAL)

=== bandwidth limit: 12MiB/s (leaky bucket) ===
[task 2] time: 4.992, bw: 5859 KiB/s (WRITE)
[task 2] time: 5.077, bw: 5859 KiB/s (READ )
[task 1] time: 5.500, bw: 5859 KiB/s (WRITE)
[task 1] time: 5.061, bw: 5859 KiB/s (READ )
[parent 0] time: 10.603, bw: 11718 KiB/s (TOTAL)

=== bandwidth limit: 12MiB/s (token bucket) ===
[task 1] time: 5.057, bw: 5859 KiB/s (WRITE)
[task 1] time: 4.329, bw: 6835 KiB/s (READ )
[task 2] time: 5.771, bw: 4882 KiB/s (WRITE)
[task 2] time: 4.961, bw: 5859 KiB/s (READ )
[parent 0] time: 10.786, bw: 11718 KiB/s (TOTAL)

=== bandwidth limit: 16MiB/s (leaky bucket) ===
[task 1] time: 3.737, bw: 7812 KiB/s (WRITE)
[task 1] time: 3.988, bw: 7812 KiB/s (READ )
[task 2] time: 4.043, bw: 7812 KiB/s (WRITE)
[task 2] time: 3.954, bw: 7812 KiB/s (READ )
[parent 0] time: 8.040, bw: 15625 KiB/s (TOTAL)

=== bandwidth limit: 16MiB/s (token bucket) ===
[task 1] time: 3.224, bw: 9765 KiB/s (WRITE)
[task 1] time: 3.550, bw: 8789 KiB/s (READ )
[task 2] time: 5.085, bw: 5859 KiB/s (WRITE)
[task 2] time: 3.033, bw: 10742 KiB/s (READ )
[parent 0] time: 8.160, bw: 15625 KiB/s (TOTAL)

=== bandwidth limit: 20MiB/s (leaky bucket) ===
[task 1] time: 3.265, bw: 9765 KiB/s (WRITE)
[task 1] time: 3.339, bw: 9765 KiB/s (READ )
[task 2] time: 3.001, bw: 10742 KiB/s (WRITE)
[task 2] time: 3.840, bw: 7812 KiB/s (READ )
[parent 0] time: 6.884, bw: 18554 KiB/s (TOTAL)

=== bandwidth limit: 20MiB/s (token bucket) ===
[task 1] time: 2.897, bw: 10742 KiB/s (WRITE)
[task 1] time: 3.071, bw: 9765 KiB/s (READ )
[task 2] time: 3.697, bw: 8789 KiB/s (WRITE)
[task 2] time: 2.925, bw: 10742 KiB/s (READ )
[parent 0] time: 6.657, bw: 19531 KiB/s (TOTAL)

=== bandwidth limit: 24MiB/s (leaky bucket) ===
[task 1] time: 2.283, bw: 13671 KiB/s (WRITE)
[task 1] time: 3.626, bw: 8789 KiB/s (READ )
[task 2] time: 3.892, bw: 7812 KiB/s (WRITE)
[task 2] time: 2.774, bw: 11718 KiB/s (READ )
[parent 0] time: 6.724, bw: 18554 KiB/s (TOTAL)

=== bandwidth limit: 24MiB/s (token bucket) ===
[task 2] time: 3.215, bw: 9765 KiB/s (WRITE)
[task 2] time: 2.767, bw: 11718 KiB/s (READ )
[task 1] time: 2.615, bw: 11718 KiB/s (WRITE)
[task 1] time: 3.958, bw: 7812 KiB/s (READ )
[parent 0] time: 6.610, bw: 19531 KiB/s (TOTAL)

In conclusion, results seem to confirm that leaky bucket is more precise
(more smoothed) than token bucket; token bucket, instead, is better in
terms of efficiency when approaching to the disk's I/O physical limit,
as the theory claims.

It would be also interesting to test how token bucket performance
changes using different bucket size values. I'll do more accurate tests
ASAP.

Signed-off-by: Andrea Righi <[email protected]>
---
block/Makefile | 2 +
block/blk-io-throttle.c | 490 +++++++++++++++++++++++++++++++++++++++
include/linux/blk-io-throttle.h | 12 +
include/linux/cgroup_subsys.h | 6 +
init/Kconfig | 10 +
5 files changed, 520 insertions(+), 0 deletions(-)

diff --git a/block/Makefile b/block/Makefile
index 5a43c7d..8dec69b 100644
--- a/block/Makefile
+++ b/block/Makefile
@@ -14,3 +14,5 @@ obj-$(CONFIG_IOSCHED_CFQ) += cfq-iosched.o

obj-$(CONFIG_BLK_DEV_IO_TRACE) += blktrace.o
obj-$(CONFIG_BLOCK_COMPAT) += compat_ioctl.o
+
+obj-$(CONFIG_CGROUP_IO_THROTTLE) += blk-io-throttle.o
diff --git a/block/blk-io-throttle.c b/block/blk-io-throttle.c
new file mode 100644
index 0000000..c6af273
--- /dev/null
+++ b/block/blk-io-throttle.c
@@ -0,0 +1,490 @@
+/*
+ * 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 <[email protected]>
+ */
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/cgroup.h>
+#include <linux/slab.h>
+#include <linux/gfp.h>
+#include <linux/err.h>
+#include <linux/sched.h>
+#include <linux/fs.h>
+#include <linux/jiffies.h>
+#include <linux/hardirq.h>
+#include <linux/list.h>
+#include <linux/spinlock.h>
+#include <linux/uaccess.h>
+#include <linux/vmalloc.h>
+#include <linux/blk-io-throttle.h>
+
+#define ONE_SEC 1000000L /* # of microseconds in a second */
+#define KBS(x) ((x) * ONE_SEC >> 10)
+
+struct iothrottle_node {
+ struct list_head node;
+ dev_t dev;
+ unsigned long iorate;
+ unsigned long timestamp;
+ atomic_long_t stat;
+ long bucket_size;
+ atomic_long_t token;
+};
+
+struct iothrottle {
+ struct cgroup_subsys_state css;
+ /* protects the list below, not the single elements */
+ spinlock_t lock;
+ struct list_head list;
+ int strategy;
+};
+
+static inline struct iothrottle *cgroup_to_iothrottle(struct cgroup *cont)
+{
+ return container_of(cgroup_subsys_state(cont, iothrottle_subsys_id),
+ struct iothrottle, css);
+}
+
+static inline struct iothrottle *task_to_iothrottle(struct task_struct *task)
+{
+ return container_of(task_subsys_state(task, iothrottle_subsys_id),
+ struct iothrottle, css);
+}
+
+static inline struct iothrottle_node *iothrottle_search_node(
+ const struct iothrottle *iot,
+ dev_t dev)
+{
+ struct iothrottle_node *n;
+
+ list_for_each_entry_rcu(n, &iot->list, node)
+ if (n->dev == dev)
+ return n;
+ return NULL;
+}
+
+static inline void iothrottle_insert_node(struct iothrottle *iot,
+ struct iothrottle_node *n)
+{
+ list_add_rcu(&n->node, &iot->list);
+}
+
+static inline struct iothrottle_node *iothrottle_replace_node(
+ struct iothrottle *iot,
+ struct iothrottle_node *old,
+ struct iothrottle_node *new)
+{
+ list_replace_rcu(&old->node, &new->node);
+ return old;
+}
+
+static inline struct iothrottle_node *iothrottle_delete_node(
+ struct iothrottle *iot,
+ dev_t dev)
+{
+ struct iothrottle_node *n;
+
+ list_for_each_entry(n, &iot->list, node)
+ if (n->dev == dev) {
+ list_del_rcu(&n->node);
+ return n;
+ }
+ return NULL;
+}
+
+/*
+ * Note: called from kernel/cgroup.c with cgroup_lock() held.
+ */
+static struct cgroup_subsys_state *iothrottle_create(
+ struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ struct iothrottle *iot;
+
+ iot = kmalloc(sizeof(*iot), GFP_KERNEL);
+ if (unlikely(!iot))
+ return ERR_PTR(-ENOMEM);
+
+ INIT_LIST_HEAD(&iot->list);
+ spin_lock_init(&iot->lock);
+ iot->strategy = 0;
+
+ return &iot->css;
+}
+
+/*
+ * Note: called from kernel/cgroup.c with cgroup_lock() held.
+ */
+static void iothrottle_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ struct iothrottle_node *n, *p;
+ struct iothrottle *iot = cgroup_to_iothrottle(cont);
+
+ /*
+ * don't worry about locking here, at this point there must be not any
+ * reference to the list.
+ */
+ list_for_each_entry_safe(n, p, &iot->list, node)
+ kfree(n);
+ kfree(iot);
+}
+
+static ssize_t iothrottle_read(struct cgroup *cont,
+ struct cftype *cft,
+ struct file *file,
+ char __user *userbuf,
+ size_t nbytes,
+ loff_t *ppos)
+{
+ struct iothrottle *iot;
+ char *buffer;
+ int s = 0;
+ struct iothrottle_node *n;
+ ssize_t ret;
+
+ buffer = kmalloc(nbytes + 1, GFP_KERNEL);
+ if (!buffer)
+ return -ENOMEM;
+
+ cgroup_lock();
+ if (cgroup_is_removed(cont)) {
+ ret = -ENODEV;
+ goto out;
+ }
+
+ iot = cgroup_to_iothrottle(cont);
+ rcu_read_lock();
+ list_for_each_entry_rcu(n, &iot->list, node) {
+ unsigned long delta, rate;
+
+ BUG_ON(!n->dev);
+ delta = jiffies_to_usecs((long)jiffies - (long)n->timestamp);
+ rate = delta ? KBS(atomic_long_read(&n->stat) / delta) : 0;
+ s += scnprintf(buffer + s, nbytes - s,
+ "device: %u,%u\n"
+ "bandwidth: %lu KiB/sec\n"
+ "usage: %lu KiB/sec\n"
+ "bucket size: %lu KiB\n"
+ "bucket fill: %li KiB\n",
+ MAJOR(n->dev), MINOR(n->dev),
+ n->iorate, rate,
+ n->bucket_size,
+ atomic_long_read(&n->token) >> 10);
+ }
+ rcu_read_unlock();
+ ret = simple_read_from_buffer(userbuf, nbytes, ppos, buffer, s);
+out:
+ cgroup_unlock();
+ kfree(buffer);
+ return ret;
+}
+
+static inline dev_t devname2dev_t(const char *buf)
+{
+ struct block_device *bdev;
+ dev_t ret;
+
+ bdev = lookup_bdev(buf);
+ if (IS_ERR(bdev))
+ return 0;
+
+ BUG_ON(!bdev->bd_inode);
+ ret = bdev->bd_inode->i_rdev;
+ bdput(bdev);
+
+ return ret;
+}
+
+static inline int iothrottle_parse_args(char *buf, size_t nbytes,
+ dev_t *dev, unsigned long *iorate,
+ unsigned long *bucket_size)
+{
+ char *ioratep, *bucket_sizep;
+
+ ioratep = memchr(buf, ':', nbytes);
+ if (!ioratep)
+ return -EINVAL;
+ *ioratep++ = '\0';
+
+ bucket_sizep = memchr(ioratep, ':', nbytes + ioratep - buf);
+ if (!bucket_sizep)
+ return -EINVAL;
+ *bucket_sizep++ = '\0';
+
+ /* i/o bandiwth is expressed in KiB/s */
+ *iorate = ALIGN(memparse(ioratep, &ioratep), 1024) >> 10;
+ if (*ioratep)
+ return -EINVAL;
+ *bucket_size = ALIGN(memparse(bucket_sizep, &bucket_sizep), 1024) >> 10;
+ if (*bucket_sizep)
+ return -EINVAL;
+
+ *dev = devname2dev_t(buf);
+ if (!*dev)
+ return -ENOTBLK;
+
+ return 0;
+}
+
+static ssize_t iothrottle_write(struct cgroup *cont,
+ struct cftype *cft,
+ struct file *file,
+ const char __user *userbuf,
+ size_t nbytes, loff_t *ppos)
+{
+ struct iothrottle *iot;
+ struct iothrottle_node *n, *tmpn = NULL;
+ char *buffer, *tmpp;
+ dev_t dev;
+ unsigned long iorate, bucket_size;
+ int ret;
+
+ if (!nbytes)
+ return -EINVAL;
+
+ /* Upper limit on largest io-throttle rule string user might write. */
+ if (nbytes > 1024)
+ return -E2BIG;
+
+ buffer = kmalloc(nbytes + 1, GFP_KERNEL);
+ if (!buffer)
+ return -ENOMEM;
+
+ if (copy_from_user(buffer, userbuf, nbytes)) {
+ ret = -EFAULT;
+ goto out1;
+ }
+
+ buffer[nbytes] = '\0';
+ tmpp = strstrip(buffer);
+
+ ret = iothrottle_parse_args(tmpp, nbytes, &dev, &iorate, &bucket_size);
+ if (ret)
+ goto out1;
+
+ if (iorate) {
+ tmpn = kmalloc(sizeof(*tmpn), GFP_KERNEL);
+ if (!tmpn) {
+ ret = -ENOMEM;
+ goto out1;
+ }
+ atomic_long_set(&tmpn->stat, 0);
+ tmpn->timestamp = jiffies;
+ tmpn->iorate = iorate;
+ tmpn->bucket_size = bucket_size;
+ atomic_long_set(&tmpn->token, 0);
+ tmpn->dev = dev;
+ }
+
+ cgroup_lock();
+ if (cgroup_is_removed(cont)) {
+ ret = -ENODEV;
+ goto out2;
+ }
+
+ iot = cgroup_to_iothrottle(cont);
+ spin_lock(&iot->lock);
+ if (!iorate) {
+ /* Delete a block device limiting rule */
+ n = iothrottle_delete_node(iot, dev);
+ goto out3;
+ }
+ n = iothrottle_search_node(iot, dev);
+ if (n) {
+ /* Update a block device limiting rule */
+ iothrottle_replace_node(iot, n, tmpn);
+ goto out3;
+ }
+ /* Add a new block device limiting rule */
+ iothrottle_insert_node(iot, tmpn);
+out3:
+ ret = nbytes;
+ spin_unlock(&iot->lock);
+ if (n) {
+ synchronize_rcu();
+ kfree(n);
+ }
+out2:
+ cgroup_unlock();
+out1:
+ kfree(buffer);
+ return ret;
+}
+
+static s64 iothrottle_strategy_read(struct cgroup *cont, struct cftype *cft)
+{
+ struct iothrottle *iot;
+ s64 ret;
+
+ cgroup_lock();
+ if (cgroup_is_removed(cont)) {
+ cgroup_unlock();
+ return -ENODEV;
+ }
+ iot = cgroup_to_iothrottle(cont);
+ ret = iot->strategy;
+ cgroup_unlock();
+ return ret;
+}
+
+static int iothrottle_strategy_write(struct cgroup *cont,
+ struct cftype *cft, s64 val)
+{
+ struct iothrottle *iot;
+
+ cgroup_lock();
+ if (cgroup_is_removed(cont)) {
+ cgroup_unlock();
+ return -ENODEV;
+ }
+ iot = cgroup_to_iothrottle(cont);
+ iot->strategy = (int)val;
+ cgroup_unlock();
+ return 0;
+}
+
+static struct cftype files[] = {
+ {
+ .name = "bandwidth",
+ .read = iothrottle_read,
+ .write = iothrottle_write,
+ },
+ {
+ .name = "throttling_strategy",
+ .read_s64 = iothrottle_strategy_read,
+ .write_s64 = iothrottle_strategy_write,
+ },
+};
+
+static int iothrottle_populate(struct cgroup_subsys *ss, struct cgroup *cont)
+{
+ return cgroup_add_files(cont, 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,
+};
+
+static inline int __cant_sleep(void)
+{
+ return in_atomic() || in_interrupt() || irqs_disabled();
+}
+
+static long leaky_bucket(struct iothrottle_node *n, size_t bytes)
+{
+ unsigned long delta, t;
+ long sleep;
+
+ /* Account the i/o activity */
+ atomic_long_add(bytes, &n->stat);
+
+ /* Evaluate if we need to throttle the current process */
+ delta = (long)jiffies - (long)n->timestamp;
+ if (!delta)
+ return 0;
+
+ t = usecs_to_jiffies(KBS(atomic_long_read(&n->stat) / n->iorate));
+ if (!t)
+ return 0;
+
+ sleep = t - delta;
+ if (unlikely(sleep > 0))
+ return sleep;
+
+ /* Reset i/o statistics */
+ atomic_long_set(&n->stat, 0);
+ /*
+ * NOTE: be sure i/o statistics have been resetted before updating the
+ * timestamp, otherwise a very small time delta may possibly be read by
+ * another CPU w.r.t. accounted i/o statistics, generating unnecessary
+ * long sleeps.
+ */
+ smp_wmb();
+ n->timestamp = jiffies;
+ return 0;
+}
+
+/* XXX: need locking in order to evaluate a consistent sleep??? */
+static long token_bucket(struct iothrottle_node *n, size_t bytes)
+{
+ unsigned long delta;
+ long tok;
+
+ atomic_long_sub(bytes, &n->token);
+
+ delta = (long)jiffies - (long)n->timestamp;
+ if (!delta)
+ return 0;
+
+ n->timestamp = jiffies;
+ tok = atomic_long_read(&n->token) + jiffies_to_msecs(delta) * n->iorate;
+ if (tok > n->bucket_size)
+ tok = n->bucket_size;
+ atomic_long_set(&n->token, tok);
+
+ return (tok < 0) ? msecs_to_jiffies(-tok / n->iorate) : 0;
+}
+
+void cgroup_io_throttle(struct block_device *bdev, size_t bytes)
+{
+ struct iothrottle *iot;
+ struct iothrottle_node *n;
+ long sleep;
+
+ if (unlikely(!bdev || !bytes))
+ return;
+
+ iot = task_to_iothrottle(current);
+ if (unlikely(!iot))
+ return;
+
+ BUG_ON(!bdev->bd_inode);
+
+ rcu_read_lock();
+ n = iothrottle_search_node(iot, bdev->bd_inode->i_rdev);
+ if (!n || !n->iorate) {
+ rcu_read_unlock();
+ return;
+ }
+ switch (iot->strategy) {
+ case 0:
+ sleep = leaky_bucket(n, bytes);
+ break;
+ case 1:
+ sleep = token_bucket(n, bytes);
+ break;
+ default:
+ sleep = 0;
+ }
+ if (unlikely(sleep)) {
+ rcu_read_unlock();
+ if (__cant_sleep())
+ return;
+ pr_debug("io-throttle: task %p (%s) must sleep %lu jiffies\n",
+ current, current->comm, sleep);
+ schedule_timeout_killable(sleep);
+ return;
+ }
+ rcu_read_unlock();
+}
+EXPORT_SYMBOL(cgroup_io_throttle);
diff --git a/include/linux/blk-io-throttle.h b/include/linux/blk-io-throttle.h
new file mode 100644
index 0000000..3e08738
--- /dev/null
+++ b/include/linux/blk-io-throttle.h
@@ -0,0 +1,12 @@
+#ifndef BLK_IO_THROTTLE_H
+#define BLK_IO_THROTTLE_H
+
+#ifdef CONFIG_CGROUP_IO_THROTTLE
+extern void cgroup_io_throttle(struct block_device *bdev, size_t bytes);
+#else
+static inline void cgroup_io_throttle(struct block_device *bdev, size_t bytes)
+{
+}
+#endif /* CONFIG_CGROUP_IO_THROTTLE */
+
+#endif /* BLK_IO_THROTTLE_H */
diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h
index e287745..0caf3c2 100644
--- a/include/linux/cgroup_subsys.h
+++ b/include/linux/cgroup_subsys.h
@@ -48,3 +48,9 @@ SUBSYS(devices)
#endif

/* */
+
+#ifdef CONFIG_CGROUP_IO_THROTTLE
+SUBSYS(iothrottle)
+#endif
+
+/* */
diff --git a/init/Kconfig b/init/Kconfig
index 6199d11..3117d99 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -306,6 +306,16 @@ config CGROUP_DEVICE
Provides a cgroup implementing whitelists for devices which
a process in the cgroup can mknod or open.

+config CGROUP_IO_THROTTLE
+ bool "Enable cgroup I/O throttling (EXPERIMENTAL)"
+ depends on CGROUPS && EXPERIMENTAL
+ help
+ This allows to limit the maximum I/O bandwidth for specific
+ cgroup(s).
+ See Documentation/controllers/io-throttle.txt for more information.
+
+ If unsure, say N.
+
config CPUSETS
bool "Cpuset support"
depends on SMP && CGROUPS