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
cgroupfilesystem directly; - via tools like
cgcreate,cgexecandcgclassify(part of the libcgroupAUR and libcgroup-gitAUR packages); - using the "rules engine daemon" to automatically move certain users/groups/commands to groups (
/etc/cgrules.confandcgconfig.service) (part of the libcgroupAUR and libcgroup-gitAUR 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:
- systemd - for controlling resources of a systemd service.
- libcgroupAUR, libcgroup-gitAUR - set of standalone tools (
cgcreate,cgclassify, persistence viacgconfig.conf).
With systemd
Hierarchy
Current cgroup hierarchy can be seen with systemctl status or systemd-cgls 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
│ ├─[email protected]
│ │ ├─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 /proc/PID/cgroup.
For example, the cgroup of the shell:
$ cat /proc/self/cgroup
0::/user.slice/user-1000.slice/session-3.scope
cgroup resource usage
The systemd-cgtop command can be used to see the resource usage:
$ systemd-cgtop
Control Group Tasks %CPU Memory Input/s Output/s user.slice 540 152,8 3.3G - - user.slice/user-1000.slice 540 152,8 3.3G - - user.slice/u…000.slice/session-1.scope 425 149,5 3.1G - - system.slice 37 - 215.6M - -
Custom cgroups
systemd.slice(5) 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 systemd.resource-control(5).
This is an example slice unit that only allows 30% of one CPU to be used:
/etc/systemd/system/my.slice
[Slice] CPUQuota=30%
Remember to do a daemon-reload to pick up any new or changed .slice files.
As service
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:
[Service] Slice=my.slice
As root
systemd-run can be used to run a command in a specific slice.
# systemd-run --slice=my.slice command
--uid=username option can be used to spawn the command as specific user.
# systemd-run --uid=username --slice=my.slice command
The --shell 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.
| Controller | Can be controlled by user | Options |
|---|---|---|
| cpu | Requires delegation | CPUAccounting, CPUWeight, CPUQuota, AllowedCPUs, AllowedMemoryNodes |
| io | Requires delegation | IOWeight, IOReadBandwidthMax, IOWriteBandwidthMax, IODeviceLatencyTargetSec |
| memory | Yes | MemoryLow, MemoryHigh, MemoryMax, MemorySwapMax |
| pids | Yes | TasksMax |
| rdma | No | ? |
| eBPF | No | IPAddressDeny, DeviceAllow, DevicePolicy |
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:
/etc/systemd/system/[email protected]/delegate.conf
[Service] Delegate=cpu cpuset io
Reboot and verify that the slice your user session is under has cpu and io controller:
$ cat /sys/fs/cgroup/user.slice/user-1000.slice/cgroup.controllers
cpuset cpu io memory pids
User-defined slices
The user slice files can be placed in ~/.config/systemd/user/.
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 systemctl set-property command. Option syntax is the same as in systemd.resource-control(5).
--runtime option is passed. Adjustments are saved at /etc/systemd/system.control/ for system wide options and .config/systemd/user.control/ for user options.For example, cutting off internet access for all user sessions:
$ systemctl set-property user.slice IPAddressDeny=any
With libcgroup and the cgroup virtual filesystem
One layer lower than management with systemd is the cgroup virtual file system. "libcgroup" provides a library and utilities for making management easier, so we will use them here as well.
The reason for using the lower level is simple: systemd does not provide an interface for every single interface file in cgroups and nor should it be expected for it to provide them at any point in the future. It is completely harmless to read from them for additional insights on a cgroup's resource use.
Before using a non-systemd tool...
One cgroup should only have one set of programs writing to it to avoid race conditions, the "single-writer rule". This is not enforced by the kernel, but following this recommendation prevents hard-to-debug issues from happening. To set the boundary at which systemd stops managing child cgroups, see the Delegate= property. Otherwise do not be surprised if system to overwrites what you have set.
Creating ad hoc groups
Delegate= settings (Delegate=yes for everything). 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
cpu or \*.Now all the tunables in the group groupname are writable by your user:
$ ls -l /sys/fs/cgroup/groupname
total 0 -r--r--r-- 1 root root 0 Jun 20 19:38 cgroup.controllers -r--r--r-- 1 root root 0 Jun 20 19:38 cgroup.events -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.freeze --w------- 1 root root 0 Jun 20 19:38 cgroup.kill -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.max.depth -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.max.descendants -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.pressure -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.procs -r--r--r-- 1 root root 0 Jun 20 19:38 cgroup.stat -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.subtree_control -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.threads -rw-r--r-- 1 root root 0 Jun 20 19:38 cgroup.type -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.idle -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.max -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.max.burst -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.pressure -r--r--r-- 1 root root 0 Jun 20 19:38 cpu.stat -r--r--r-- 1 root root 0 Jun 20 19:38 cpu.stat.local -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.uclamp.max -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.uclamp.min -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.weight -rw-r--r-- 1 root root 0 Jun 20 19:38 cpu.weight.nice -rw-r--r-- 1 root root 0 Jun 20 19:38 io.pressure -rw-r--r-- 1 root root 0 Jun 20 19:38 irq.pressure -r--r--r-- 1 root root 0 Jun 20 19:38 memory.current -r--r--r-- 1 root root 0 Jun 20 19:38 memory.events -r--r--r-- 1 root root 0 Jun 20 19:38 memory.events.local -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.high -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.low -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.max -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.min -r--r--r-- 1 root root 0 Jun 20 19:38 memory.numa_stat -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.oom.group -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.peak -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.pressure --w------- 1 root root 0 Jun 20 19:38 memory.reclaim -r--r--r-- 1 root root 0 Jun 20 19:38 memory.stat -r--r--r-- 1 root root 0 Jun 20 19:38 memory.swap.current -r--r--r-- 1 root root 0 Jun 20 19:38 memory.swap.events -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.swap.high -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.swap.max -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.swap.peak -r--r--r-- 1 root root 0 Jun 20 19:38 memory.zswap.current -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.zswap.max -rw-r--r-- 1 root root 0 Jun 20 19:38 memory.zswap.writeback -r--r--r-- 1 root root 0 Jun 20 19:38 pids.current -r--r--r-- 1 root root 0 Jun 20 19:38 pids.events -r--r--r-- 1 root root 0 Jun 20 19:38 pids.events.local -rw-r--r-- 1 root root 0 Jun 20 19:38 pids.max -r--r--r-- 1 root root 0 Jun 20 19:38 pids.peak
Cgroups are hierarchical, so you can create as many subgroups as you like. If a normal user wants to make new subgroup called foo:
$ cgcreate -g cpu:groupname/foo
Using cgroups
As previously mentioned, only one thing should write to a cgroup at any point. This does not affect non-write operations including spawning new processes inside a group, moving processes to a group, or reading properties from cgroup files.
Spawning and moving processes
libcgroup contains a simple tool for running new processes inside a cgroup. If a normal user wants to run a bash shell under a our previous groupname/foo:
$ cgexec -g cpu:groupname/foo bash
Inside of the shell, we can confirm which cgroup it belongs to with:
$ cat /proc/self/cgroup
0::/groupname/foo
This makes use of /proc/$PID/cgroup, a file that exists in every process. Manually writing to the file causes the cgroup to change as well.
To move all 'bash' commands to this group:
$ pidof bash
13244 13266
$ cgclassify -g cpu:groupname/foo `pidof bash`
$ cat /proc/13244/cgroup
0::/groupname/foo
Internally (i.e. without cgclassify the kernel provides two ways to move processes between cgroups. These two are equivalent:
$ echo 0::/groupname/foo > /proc/13244/cgroup $ echo 13244 > /sys/fs/cgroup/groupname/foo/cgroup.procs
Manipulating group properties
A new subdirectory is crated for groupname/foo at its creation, located at /sys/fs/cgroup/groupname/foo. These files can be read and written to change the group's properties. (Again, writing is not recommended unless delegation is done!)
Let us try to see how much memory all the processes in our group is taking up:
$ cat /sys/fs/cgroup/groupname/foo/memory.current
1536000
To limit the RAM (not swap) usage of all processes, run the following:
$ echo 10000000 > /sys/fs/cgroup/groupname/foo/memory.max
To change the CPU priority of this group (the default is 100):
$ echo 10 > /sys/fs/cgroup/groupname/foo/cpu.weight
You can find more tunables or statistics by listing the cgroup directory.
Persistent group configuration
If you want your cgroups to be created at boot, you can define them in /etc/cgconfig.conf instead. This causes a service booted at launch to configure your cgroups. See the relevant manual page about the syntax of this file; we will make no instruction on how to use a truly deprecated mechanism.
Examples
Restrict memory or CPU use of a command
The following example shows a cgroup that constrains a given command to 2GB of memory.
$ systemd-run --scope -p MemoryMax=2G --user command
The following example shows a command restricted to 20% of one CPU core.
$ systemd-run --scope -p CPUQuota="20%" --user command
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
~/.config/systemd/user/matlab.slice
[Slice] AllowedCPUs=0-5 MemoryHigh=6G
Launch Matlab like this (be sure to use the right path):
$ systemd-run --user --slice=matlab.slice /opt/MATLAB/2012b/bin/matlab -desktop
Documentation
- For information on controllers and what certain switches and tunables mean, refer to kernel's documentation v2 (or install linux-docs and see
/usr/src/linux/Documentation/cgroup) - Linux manual page: cgroups(7)
- A detailed and complete Resource Management Guide can be found in the Red Hat Enterprise Linux documentation.
For commands and configuration files, see relevant man pages, e.g. cgcreate(1) or cgrules.conf(5)
Historical note: cgroup v1
Before our current cgroup v2 there was an earlier version called v1. V1 allowed a lot of additional flexibility including a non-unified hierarchy and thread-granular management. This was, in retrospect, a bad idea (see the rationales for v2):
- Even though many hierarchies can exist and processes can be bound to several hierarchies, a controller can ever be used in one hierarchy. This makes the many hierarchies essentially pointless, with the usual setup being to bind each controller to one hierarchy (e.g.
/sys/fs/cgroup/memory/), then each process to multiple hierarchies. This in turn made tools likecgcreateessential for syncing the membership of processes across multiple hierarchies. - Thread-granular management caused cgroup to be abused as a way for a process to manage itself. The proper way to do this is through system calls, not the convoluted interfaces that have emerged to support this use. Self-managing requires clunky string management and is inherently vulnerable to race conditions.
To avoid further chaos, cgroup v2 has two key design rules on top of the removal of features:
- If a cgroup has child cgroups, it cannot have attached processes (with the exception of the root cgroup). This is enforced on v2 and helps make the single-writer rule (below) viable.
- Each cgroup should only have one process managing it at the same time (single-writer rule). This is not enforced anywhere but should be respected at most times to avoid pain from software fighting each other over what to do to a group.
- On systemd systems, the root cgroup is managed by systemd, making any change not by systemd a violation of this rule (or, as it's not enforced, recommendation), unless when
Delegate=is set on the surrounding service or scope unit to tell systemd to not meddle with what is inside.
- On systemd systems, the root cgroup is managed by systemd, making any change not by systemd a violation of this rule (or, as it's not enforced, recommendation), unless when
Before systemd v258, the kernel parameters SYSTEMD_CGROUP_ENABLE_LEGACY_FORCE=1 systemd.unified_cgroup_hierarchy=0 could be used to force booting with cgroup-v1 (the first parameter was added in v256 to make it harder to use cgroup-v1). However, this feature has now been removed. It is still worth knowing about because some software like to put systemd.unified_cgroup_hierarchy=0 in your kernel command-line without telling you, causing your entire system to break.