cgroups

Control groups (or cgroups as they are commonly known) are a feature provided by the Linux kernel to manage, restrict, and audit groups of processes. Compared to other approaches like the nice(1) command or /etc/security/limits.conf, cgroups are more flexible as they can operate on (sub)sets of processes (possibly with different system users).

Control groups can be accessed with various tools:

  • using directives in systemd unit files to specify limits for services and slices;
  • by accessing the cgroup filesystem directly;
  • via tools like , and (part of the and packages);
  • using the "rules engine daemon" to automatically move certain users/groups/commands to groups (/etc/cgrules.conf and cgconfig.service) (part of the and packages); and
  • through other software such as Linux Containers (LXC) virtualization.

For Arch Linux, systemd is the preferred and easiest method of invoking and configuring cgroups as it is a part of the default installation.

Installing

Make sure you have one of these packages installed for automated cgroup handling:

  • - for controlling resources of a systemd service.
  • , - set of standalone tools (, , persistence via ).

With systemd

Hierarchy

Current cgroup hierarchy can be seen with or command.

$ systemctl status
● myarchlinux
    State: running
     Jobs: 0 queued
   Failed: 0 units
    Since: Wed 2019-12-04 22:16:28 UTC; 1 day 4h ago
   CGroup: /
           ├─user.slice 
           │ └─user-1000.slice 
           │   ├─user@1000.service 
           │   │ ├─gnome-shell-wayland.service 
           │   │ │ ├─ 1129 /usr/bin/gnome-shell
           │   │ ├─gnome-terminal-server.service 
           │   │ │ ├─33519 /usr/lib/gnome-terminal-server
           │   │ │ ├─37298 fish
           │   │ │ └─39239 systemctl status
           │   │ ├─init.scope 
           │   │ │ ├─1066 /usr/lib/systemd/systemd --user
           │   │ │ └─1067 (sd-pam)
           │   └─session-2.scope 
           │     ├─1053 gdm-session-worker [pam/gdm-password]
           │     ├─1078 /usr/bin/gnome-keyring-daemon --daemonize --login
           │     ├─1082 /usr/lib/gdm-wayland-session /usr/bin/gnome-session
           │     ├─1086 /usr/lib/gnome-session-binary
           │     └─3514 /usr/bin/ssh-agent -D -a /run/user/1000/keyring/.ssh
           ├─init.scope 
           │ └─1 /sbin/init
           └─system.slice 
             ├─systemd-udevd.service 
             │ └─285 /usr/lib/systemd/systemd-udevd
             ├─systemd-journald.service 
             │ └─272 /usr/lib/systemd/systemd-journald
             ├─NetworkManager.service 
             │ └─656 /usr/bin/NetworkManager --no-daemon
             ├─gdm.service 
             │ └─668 /usr/bin/gdm
             └─systemd-logind.service 
               └─654 /usr/lib/systemd/systemd-logind

Find cgroup of a process

The cgroup name of a process can be found in .

For example, the cgroup of the shell:

cgroup resource usage

The command can be used to see the resource usage:

Custom cgroups

systemd unit files can be used to define a custom cgroup configuration.  They must be placed in a systemd directory, such as /etc/systemd/system/.  The resource control options that can be assigned are documented in .

This is an example slice unit that only allows 30% of one CPU to be used:

Remember to do a daemon-reload to pick up any new or changed files.

Service unit file

Resources can be directly specified in service definition or as a drop-in file:

[Service]
MemoryMax=1G 

This example limits the service to 1 gigabyte.

Grouping unit under a slice

Service can be specified what slice to run in:

As root

can be used to run a command in a specific slice. 
# systemd-run --slice=my.slice command
option can be used to spawn the command as specific user.
# systemd-run --uid=username --slice=my.slice command

The option can be used to spawn a command shell inside the slice.

As unprivileged user

Unprivileged users can divide the resources provided to them into new cgroups, if some conditions are met.

Cgroups v2 must be utilized for a non-root user to be allowed managing cgroup resources.

Controller types

Not all resources can be controlled by user.

ControllerCan be controlled by userOptions
cpuRequires delegationCPUAccounting, CPUWeight, CPUQuota, AllowedCPUs, AllowedMemoryNodes
ioRequires delegationIOWeight, IOReadBandwidthMax, IOWriteBandwidthMax, IODeviceLatencyTargetSec
memoryMemoryLow, MemoryHigh, MemoryMax, MemorySwapMax
pidsTasksMax
rdma?
eBPFIPAddressDeny, DeviceAllow, DevicePolicy
Note: eBPF is technically not a controller but those systemd options implemented using it and only root is allowed to set them.

User delegation

For user to control cpu and io resources, the resources need to be delegated. This can be done with a drop-in file.

For example if your user id is 1000:

Reboot and verify that the slice your user session is under has cpu and io controller:

User-defined slices

The user slice files can be placed in .

To run the command under certain slice:

$ systemd-run --user --slice=my.slice command

You can also run your login shell inside the slice:

$ systemd-run --user --slice=my.slice --shell

Run-time adjustment

cgroups resources can be adjusted at run-time using command. Option syntax is the same as in .

For example, cutting off internet access for all user sessions:

$ systemctl set-property user.slice IPAddressDeny=any

With libcgroup

You can enable the service with systemd. This allows you to track any errors in more easily.

Ad-hoc groups

One of the powers of cgroups is that you can create "ad-hoc" groups on the fly. You can even grant the privileges to create custom groups to regular users. groupname is the cgroup name:

# cgcreate -a user -t user -g memory,cpu:groupname

Now all the tunables in the group groupname are writable by your user:

Cgroups are hierarchical, so you can create as many subgroups as you like. If a normal user wants to run a shell under a new subgroup called foo:

$ cgcreate -g memory,cpu:groupname/foo
$ cgexec    -g memory,cpu:groupname/foo bash

To make sure (only meaningful for legacy (v1) cgroups):

A new subdirectory was created for this group. To limit the memory usage of all processes in this group to 10 MB, run the following:

$ echo 10000000 > /sys/fs/cgroup/memory/groupname/foo/memory.limit_in_bytes

Note that the memory limit applies to RAM use only -- once tasks hit this limit, they will begin to swap. But it will not affect the performance of other processes significantly.

Similarly you can change the CPU priority ("shares") of this group. By default all groups have 1024 shares. A group with 100 shares will get a ~10% portion of the CPU time:

$ echo 100 > /sys/fs/cgroup/cpu/groupname/foo/cpu.shares

You can find more tunables or statistics by listing the cgroup directory.

You can also change the cgroup of already running processes. To move all 'bash' commands to this group:

$ pidof bash
13244 13266
$ cgclassify -g memory,cpu:groupname/foo `pidof bash`
$ cat /proc/13244/cgroup
11:memory:/groupname/foo
6:cpu:/groupname/foo

Persistent group configuration

If you want your cgroups to be created at boot, you can define them in instead. For example, the "groupname" has a permission for and users of group to manage limits and add tasks. A subgroup "groupname/foo" group definitions would look like this:

Note:
  • Comments should begin at the start of a line! The # character for comments must appear as the first character of a line. Else, cgconfigparser will have problem parsing it but will only report cgroup change of group failed as the error, unless you started cgconfig with Systemd
  • The permissions section is optional.
  • The /sys/fs/cgroup/ hierarchy directory containing all controllers sub-directories is already created and mounted at boot as a virtual file system. This gives the ability to create a new group entry with the $CONTROLLER-NAME { } command. If for any reason you want to create and mount hierachies in another place, you will then need to write a second entry in /etc/cgconfig.conf following this way :
 mount {    
   cpuset = /your/path/groupname;
 }

This is equivalent to these shell commands:

 # mkdir /your/path/groupname
 # mount -t /your/path -o cpuset groupname /your/path/groupname

With the cgroup virtual filesystem

Starting with systemd 232, the cgm method described in the next section, this section will instead describe a manual method to limit memory usage.

Create a new cgroup named groupname:

# mkdir /sys/fs/cgroup/memory/groupname

Example: set the maximum memory limit to 100MB:

# echo 100000000 > /sys/fs/cgroup/memory/groupname/memory.limit_in_bytes

Move a process to the cgroup (note: only one PID can be written at a time, repeat this for each process that must be moved):

# echo pid > /sys/fs/cgroup/memory/groupname/cgroup.procs

Examples

Matlab

Doing large calculations in MATLAB can crash your system, because Matlab does not have any protection against taking all your machine's memory or CPU. The following examples show a cgroup that constrains Matlab to first 6 CPU cores and 5 GB of memory.

With systemd

Launch Matlab like this (be sure to use the right path):

$ systemd-run --user --slice=matlab.slice /opt/MATLAB/2012b/bin/matlab -desktop

With libcgroup

Change to the user Matlab is run as.

You can also restrict the CPU share with the cpu constraint.

Launch Matlab like this (be sure to use the right path):

$ cgexec -g memory,cpuset:matlab /opt/MATLAB/2012b/bin/matlab -desktop

Documentation

  • For information on controllers and what certain switches and tunables mean, refer to kernel's documentation v1 or v2 (or install and see )
  • A detailed and complete Resource Management Guide can be found in the fedora project documentation.

For commands and configuration files, see relevant man pages, e.g. or

Tips and tricks

Enable cgroup v1

Cgroup v2 is now enabled by default. If you want to switch to cgroup v1 instead, you need to set the following kernel parameter:

systemd.unified_cgroup_hierarchy=0
gollark: The RPC thing is apparently in beta, requires the Discord app to be running as an actual program and not in the browser, and has very restrictive restrictions, so eh.
gollark: Ah, I see. Interesting.
gollark: Unless these are new special bots?
gollark: This is for *bots*, and they have to be explicitly invited to a server.
gollark: ↑

See also

This article is issued from Archlinux. The text is licensed under Creative Commons - Attribution - Sharealike. Additional terms may apply for the media files.