cgroups-notes

Links

man page

Brief

Cgroups are controlled via virtual file system on /sys/fs/cgroup.

cgroup v1 example

The following commands are limiting CPU share for process with PID of 21764:

cd /sys/fs/cgroup

# Create directory for CPU controller (CPU shares limit)
sudo mkdir cpu

sudo mount -t cgroup -o cpu none /sys/fs/cgroup/cpu
sudo mkdir /sys/fs/cgroup/cpu/my_cgroup

cd my_cgroup

# Add PID to cgroup's tasks
echo "21764" | sudo tee tasks

# Set the CPU shares for the cgroup
echo "16" | sudo tee cpu.shares

If you use the test_app that has infinite loop as example app here,
note that it will still have 100% cpu consumption.
If you want to see the cgroup's CPU limit effect you have to run
multiple instances of test_app while only one is in cgroup.
The instance that is in cgroup with limited CPU will consume less
CPU than the rest.

cgroup directory content example:

pi@raspberrypi:/sys/fs/cgroup/cpu/my_cgroup $ ls -lrt
total 0
-rw-r--r-- 1 root root 0 Oct 15 01:24 notify_on_release
-r--r--r-- 1 root root 0 Oct 15 01:24 cpu.stat
-rw-r--r-- 1 root root 0 Oct 15 01:24 cpu.idle
-rw-r--r-- 1 root root 0 Oct 15 01:24 cpu.cfs_quota_us
-rw-r--r-- 1 root root 0 Oct 15 01:24 cpu.cfs_period_us
-rw-r--r-- 1 root root 0 Oct 15 01:24 cpu.cfs_burst_us
-rw-r--r-- 1 root root 0 Oct 15 01:24 cgroup.procs
-rw-r--r-- 1 root root 0 Oct 15 01:24 cgroup.clone_children
-rw-r--r-- 1 root root 0 Oct 15 02:12 tasks
-rw-r--r-- 1 root root 0 Oct 15 02:15 cpu.shares

Checking cgroups version and supported controllers

You can check this by looking at the /proc/cgroups file.
If it contains 1 next to the cgroup2 entry, it means Cgroups v2 is supported.
Each controller (resource like cpu,mem,etc.) has enabled flag.

pi@raspberrypi:/sys/fs/cgroup $ cat /proc/cgroups
#subsys_name    hierarchy       num_cgroups     enabled
cpuset  0       59      1
cpu     0       59      1
cpuacct 0       59      1
blkio   0       59      1
memory  0       59      0
devices 0       59      1
freezer 0       59      1
net_cls 0       59      1
perf_event      0       59      1
net_prio        0       59      1
pids    0       59      1

cgroup file system

cgroups kernel code is located under linux/kernel/cgroup directory.
cgroup.c has call for kernfs_create_root kernel API to create new
file system.

Above we used the command sudo mount -t cgroup -o cpu none /sys/fs/cgroup/cpu
in order to add new controller to the cgroups.
This is implemented via the cgroup file system.
The read/write and other operations are described in cgroup.c file.
cgroup.c link