While StartupCPUWeight= applies to the startup and shutdown phases of the system, CPUWeight= applies to normal runtime of the system, and if the former is not set also to the startup and shutdown phases. Using StartupCPUWeight= allows prioritizing specific services at boot-up and shutdown differently than during normal runtime.

Example: CPUQuota=20% ensures that the executed processes will never get more than 20% CPU time on one CPU.

This controls the second field of "cpu.max" attribute on the unified control group hierarchy and "cpu.cfs_period_us" on legacy. For details about these control group attributes, see Control Groups v2[2] and CFS Scheduler[3].

Example: CPUQuotaPeriodSec=10ms to request that the CPU quota is measured in periods of 10ms.

Setting AllowedCPUs= or StartupAllowedCPUs= doesn't guarantee that all of the CPUs will be used by the processes as it may be limited by parent units. The effective configuration is reported as EffectiveCPUs=.

While StartupAllowedCPUs= applies to the startup and shutdown phases of the system, AllowedCPUs= applies to normal runtime of the system, and if the former is not set also to the startup and shutdown phases. Using StartupAllowedCPUs= allows prioritizing specific services at boot-up and shutdown differently than during normal runtime.

This setting is supported only with the unified control group hierarchy.

Setting AllowedMemoryNodes= or StartupAllowedMemoryNodes= doesn't guarantee that all of the memory NUMA nodes will be used by the processes as it may be limited by parent units. The effective configuration is reported as EffectiveMemoryNodes=.

While StartupAllowedMemoryNodes= applies to the startup and shutdown phases of the system, AllowedMemoryNodes= applies to normal runtime of the system, and if the former is not set also to the startup and shutdown phases. Using StartupAllowedMemoryNodes= allows prioritizing specific services at boot-up and shutdown differently than during normal runtime.

This setting is supported only with the unified control group hierarchy.

For a protection to be effective, it is generally required to set a corresponding allocation on all ancestors, which is then distributed between children (with the exception of the root slice). Any MemoryMin= or MemoryLow= allocation that is not explicitly distributed to specific children is used to create a shared protection for all children. As this is a shared protection, the children will freely compete for the memory.

Takes a memory size in bytes. If the value is suffixed with K, M, G or T, the specified memory size is parsed as Kilobytes, Megabytes, Gigabytes, or Terabytes (with the base 1024), respectively. Alternatively, a percentage value may be specified, which is taken relative to the installed physical memory on the system. If assigned the special value "infinity", all available memory is protected, which may be useful in order to always inherit all of the protection afforded by ancestors. This controls the "memory.min" or "memory.low" control group attribute. For details about this control group attribute, see Memory Interface Files[5].

Units may have their children use a default "memory.min" or "memory.low" value by specifying DefaultMemoryMin= or DefaultMemoryLow=, which has the same semantics as MemoryMin= and MemoryLow=. This setting does not affect "memory.min" or "memory.low" in the unit itself. Using it to set a default child allocation is only useful on kernels older than 5.7, which do not support the "memory_recursiveprot" cgroup2 mount option.

Takes a memory size in bytes. If the value is suffixed with K, M, G or T, the specified memory size is parsed as Kilobytes, Megabytes, Gigabytes, or Terabytes (with the base 1024), respectively. Alternatively, a percentage value may be specified, which is taken relative to the installed physical memory on the system. If assigned the special value "infinity", no memory throttling is applied. This controls the "memory.high" control group attribute. For details about this control group attribute, see Memory Interface Files[5].

Takes a memory size in bytes. If the value is suffixed with K, M, G or T, the specified memory size is parsed as Kilobytes, Megabytes, Gigabytes, or Terabytes (with the base 1024), respectively. Alternatively, a percentage value may be specified, which is taken relative to the installed physical memory on the system. If assigned the special value "infinity", no memory limit is applied. This controls the "memory.max" control group attribute. For details about this control group attribute, see Memory Interface Files[5].

Takes a swap size in bytes. If the value is suffixed with K, M, G or T, the specified swap size is parsed as Kilobytes, Megabytes, Gigabytes, or Terabytes (with the base 1024), respectively. If assigned the special value "infinity", no swap limit is applied. This controls the "memory.swap.max" control group attribute. For details about this control group attribute, see Memory Interface Files[5].

The system default for this setting may be controlled with DefaultTasksMax= in systemd-system.conf(5).

While StartupIOWeight= applies to the startup and shutdown phases of the system, IOWeight= applies to the later runtime of the system, and if the former is not set also to the startup and shutdown phases. This allows prioritizing specific services at boot-up and shutdown differently than during runtime.

The specified device node should reference a block device that has an I/O scheduler associated, i.e. should not refer to partition or loopback block devices, but to the originating, physical device. When a path to a regular file or directory is specified it is attempted to discover the correct originating device backing the file system of the specified path. This works correctly only for simpler cases, where the file system is directly placed on a partition or physical block device, or where simple 1:1 encryption using dm-crypt/LUKS is used. This discovery does not cover complex storage and in particular RAID and volume management storage devices.

Similar restrictions on block device discovery as for IODeviceWeight= apply, see above.

Similar restrictions on block device discovery as for IODeviceWeight= apply, see above.

Implies "IOAccounting=yes".

These settings are supported only if the unified control group hierarchy is used.

Similar restrictions on block device discovery as for IODeviceWeight= apply, see above.

When this option is used in socket units, it applies to all IPv4 and IPv6 sockets associated with it (including both listening and connection sockets where this applies). Note that for socket-activated services, this configuration setting and the accounting data of the service unit and the socket unit are kept separate, and displayed separately. No propagation of the setting and the collected statistics is done, in either direction. Moreover, any traffic sent or received on any of the socket unit's sockets is accounted to the socket unit — and never to the service unit it might have activated, even if the socket is used by it.

The system default for this setting may be controlled with DefaultIPAccounting= in systemd-system.conf(5).

The access lists configured with this option are applied to all sockets created by processes of this unit (or in the case of socket units, associated with it). The lists are implicitly combined with any lists configured for any of the parent slice units this unit might be a member of. By default both access lists are empty. Both ingress and egress traffic is filtered by these settings. In case of ingress traffic the source IP address is checked against these access lists, in case of egress traffic the destination IP address is checked. The following rules are applied in turn:

In order to implement an allow-listing IP firewall, it is recommended to use a IPAddressDeny=any setting on an upper-level slice unit (such as the root slice -.slice or the slice containing all system services system.slice – see systemd.special(7) for details on these slice units), plus individual per-service IPAddressAllow= lines permitting network access to relevant services, and only them.

Note that for socket-activated services, the IP access list configured on the socket unit applies to all sockets associated with it directly, but not to any sockets created by the ultimately activated services for it. Conversely, the IP access list configured for the service is not applied to any sockets passed into the service via socket activation. Thus, it is usually a good idea to replicate the IP access lists on both the socket and the service unit. Nevertheless, it may make sense to maintain one list more open and the other one more restricted, depending on the usecase.

If these settings are used multiple times in the same unit the specified lists are combined. If an empty string is assigned to these settings the specific access list is reset and all previous settings undone.

In place of explicit IPv4 or IPv6 address and prefix length specifications a small set of symbolic names may be used. The following names are defined:

Table 1. Special address/network names

Note that these settings might not be supported on some systems (for example if eBPF control group support is not enabled in the underlying kernel or container manager). These settings will have no effect in that case. If compatibility with such systems is desired it is hence recommended to not exclusively rely on them for IP security.

The filters configured with this option are applied to all sockets created by processes of this unit (or in the case of socket units, associated with it). The filters are loaded in addition to filters any of the parent slice units this unit might be a member of as well as any IPAddressAllow= and IPAddressDeny= filters in any of these units. By default there are no filters specified.

If these settings are used multiple times in the same unit all the specified programs are attached. If an empty string is assigned to these settings the program list is reset and all previous specified programs ignored.

If the path BPF_FS_PROGRAM_PATH in IPIngressFilterPath= assignment is already being handled by BPFProgram= ingress hook, e.g. BPFProgram=ingress:BPF_FS_PROGRAM_PATH, the assignment will be still considered valid and the program will be attached to a cgroup. Same for IPEgressFilterPath= path and egress hook.

Note that for socket-activated services, the IP filter programs configured on the socket unit apply to all sockets associated with it directly, but not to any sockets created by the ultimately activated services for it. Conversely, the IP filter programs configured for the service are not applied to any sockets passed into the service via socket activation. Thus, it is usually a good idea, to replicate the IP filter programs on both the socket and the service unit, however it often makes sense to maintain one configuration more open and the other one more restricted, depending on the usecase.

Note that these settings might not be supported on some systems (for example if eBPF control group support is not enabled in the underlying kernel or container manager). These settings will fail the service in that case. If compatibility with such systems is desired it is hence recommended to attach your filter manually (requires Delegate=yes) instead of using this setting.

BPFProgram= allows attaching BPF hooks to the cgroup of a systemd unit. (This generalizes the functionality exposed via IPEgressFilterPath= for egress and IPIngressFilterPath= for ingress.) Cgroup-bpf hooks in the form of BPF programs loaded to the BPF filesystem are attached with cgroup-bpf attach flags determined by the unit. For details about attachment types and flags see https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/plain/include/uapi/linux/bpf.h. For general BPF documentation please refer to https://docs.kernel.org/bpf/index.html.

The specification of BPF program consists of a type followed by a program-path with ":" as the separator: type:program-path.

type is the string name of BPF attach type also used in bpftool. type can be one of egress, ingress, sock_create, sock_ops, device, bind4, bind6, connect4, connect6, post_bind4, post_bind6, sendmsg4, sendmsg6, sysctl, recvmsg4, recvmsg6, getsockopt, setsockopt.

Setting BPFProgram= to an empty value makes previous assignments ineffective.

Multiple assignments of the same type:program-path value have the same effect as a single assignment: the program with the path program-path will be attached to cgroup hook type just once.

If BPF egress pinned to program-path path is already being handled by IPEgressFilterPath=, BPFProgram= assignment will be considered valid and BPFProgram= will be attached to a cgroup. Similarly for ingress hook and IPIngressFilterPath= assignment.

BPF programs passed with BPFProgram= are attached to the cgroup of a unit with BPF attach flag multi, that allows further attachments of the same type within cgroup hierarchy topped by the unit cgroup.

Examples:

BPFProgram=egress:/sys/fs/bpf/egress-hook
BPFProgram=bind6:/sys/fs/bpf/sock-addr-hook

bind-rule describes socket properties such as address-family, transport-protocol and ip-ports.

bind-rule := { [address-family:][transport-protocol:][ip-ports] | any }

address-family := { ipv4 | ipv6 }

transport-protocol := { tcp | udp }

ip-ports := { ip-port | ip-port-range }

An optional address-family expects ipv4 or ipv6 values. If not specified, a rule will be matched for both IPv4 and IPv6 addresses and applied depending on other socket fields, e.g. transport-protocol, ip-port.

An optional transport-protocol expects tcp or udp transport protocol names. If not specified, a rule will be matched for any transport protocol.

An optional ip-port value must lie within 1...65535 interval inclusively, i.e. dynamic port 0 is not allowed. A range of sequential ports is described by ip-port-range := ip-port-low-ip-port-high, where ip-port-low is smaller than or equal to ip-port-high and both are within 1...65535 inclusively.

A special value any can be used to apply a rule to any address family, transport protocol and any port with a positive value.

To allow multiple rules assign SocketBindAllow= or SocketBindDeny= multiple times. To clear the existing assignments pass an empty SocketBindAllow= or SocketBindDeny= assignment.

For each of SocketBindAllow= and SocketBindDeny=, maximum allowed number of assignments is 128.

The feature is implemented with cgroup/bind4 and cgroup/bind6 cgroup-bpf hooks.

Examples:

...
# Allow binding IPv6 socket addresses with a port greater than or equal to 10000.
[Service]
SocketBindAllow=ipv6:10000-65535
SocketBindDeny=any
...
# Allow binding IPv4 and IPv6 socket addresses with 1234 and 4321 ports.
[Service]
SocketBindAllow=1234
SocketBindAllow=4321
SocketBindDeny=any
...
# Deny binding IPv6 socket addresses.
[Service]
SocketBindDeny=ipv6
...
# Deny binding IPv4 and IPv6 socket addresses.
[Service]
SocketBindDeny=any
...
# Allow binding only over TCP
[Service]
SocketBindAllow=tcp
SocketBindDeny=any
...
# Allow binding only over IPv6/TCP
[Service]
SocketBindAllow=ipv6:tcp
SocketBindDeny=any
...
# Allow binding ports within 10000-65535 range over IPv4/UDP.
[Service]
SocketBindAllow=ipv4:udp:10000-65535
SocketBindDeny=any
...

This option can appear multiple times, in which case the network interface names are merged. If the empty string is assigned the set is reset, all prior assignments will have not effect.

If you specify both types of this option (i.e. allow-listing and deny-listing), the first encountered will take precedence and will dictate the default action (allow vs deny). Then the next occurrences of this option will add or delete the listed network interface names from the set, depending of its type and the default action.

The loopback interface ("lo") is not treated in any special way, you have to configure it explicitly in the unit file.

Example 1: allow-list

RestrictNetworkInterfaces=eth1
RestrictNetworkInterfaces=eth2

Programs in the unit will be only able to use the eth1 and eth2 network interfaces.

Example 2: deny-list

RestrictNetworkInterfaces=~eth1 eth2

Programs in the unit will be able to use any network interface but eth1 and eth2.

Example 3: mixed

RestrictNetworkInterfaces=eth1 eth2
RestrictNetworkInterfaces=~eth1

Programs in the unit will be only able to use the eth2 network interface.

When access to all physical devices should be disallowed, PrivateDevices= may be used instead. See systemd.exec(5).

The device node specifier is either a path to a device node in the file system, starting with /dev/, or a string starting with either "char-" or "block-" followed by a device group name, as listed in /proc/devices. The latter is useful to allow-list all current and future devices belonging to a specific device group at once. The device group is matched according to filename globbing rules, you may hence use the "*" and "?" wildcards. (Note that such globbing wildcards are not available for device node path specifications!) In order to match device nodes by numeric major/minor, use device node paths in the /dev/char/ and /dev/block/ directories. However, matching devices by major/minor is generally not recommended as assignments are neither stable nor portable between systems or different kernel versions.

Examples: /dev/sda5 is a path to a device node, referring to an ATA or SCSI block device. "char-pts" and "char-alsa" are specifiers for all pseudo TTYs and all ALSA sound devices, respectively. "char-cpu/*" is a specifier matching all CPU related device groups.

Note that allow lists defined this way should only reference device groups which are resolvable at the time the unit is started. Any device groups not resolvable then are not added to the device allow list. In order to work around this limitation, consider extending service units with a pair of After=modprobe@xyz.service and Wants=modprobe@xyz.service lines that load the necessary kernel module implementing the device group if missing. Example:

...
[Unit]
Wants=modprobe@loop.service
After=modprobe@loop.service
[Service]
DeviceAllow=block-loop
DeviceAllow=/dev/loop-control
...

strict

closed

auto

This option may be used to arrange systemd units in a hierarchy of slices each of which might have resource settings applied.

For units of type slice, the only accepted value for this setting is the parent slice. Since the name of a slice unit implies the parent slice, it is hence redundant to ever set this parameter directly for slice units.

Special care should be taken when relying on the default slice assignment in templated service units that have DefaultDependencies=no set, see systemd.service(5), section "Default Dependencies" for details.

Note that controller delegation to less privileged code is only safe on the unified control group hierarchy. Accordingly, access to the specified controllers will not be granted to unprivileged services on the legacy hierarchy, even when requested.

The following controller names may be specified: cpu, cpuacct, cpuset, io, blkio, memory, devices, pids, bpf-firewall, and bpf-devices.

Not all of these controllers are available on all kernels however, and some are specific to the unified hierarchy while others are specific to the legacy hierarchy. Also note that the kernel might support further controllers, which aren't covered here yet as delegation is either not supported at all for them or not defined cleanly.

For further details on the delegation model consult Control Group APIs and Delegation[8].

It may not be possible to successfully disable a controller if the unit or any child of the unit in question delegates controllers to its children, as any delegated subtree of the cgroup hierarchy is unmanaged by systemd.

Multiple controllers may be specified, separated by spaces. You may also pass DisableControllers= multiple times, in which case each new instance adds another controller to disable. Passing DisableControllers= by itself with no controller name present resets the disabled controller list.

The following controller names may be specified: cpu, cpuacct, cpuset, io, blkio, memory, devices, pids, bpf-firewall, and bpf-devices.

When set to kill, the unit becomes a candidate for monitoring by systemd-oomd. If the cgroup passes the limits set by oomd.conf(5) or the unit configuration, systemd-oomd will select a descendant cgroup and send SIGKILL to all of the processes under it. You can find more details on candidates and kill behavior at systemd-oomd.service(8) and oomd.conf(5).

Setting either of these properties to kill will also result in After= and Wants= dependencies on systemd-oomd.service unless DefaultDependencies=no.

When set to auto, systemd-oomd will not actively use this cgroup's data for monitoring and detection. However, if an ancestor cgroup has one of these properties set to kill, a unit with auto can still be a candidate for systemd-oomd to terminate.

When calculating candidates to relieve swap usage, systemd-oomd will only respect these extended attributes if the unit's cgroup is owned by root.

When calculating candidates to relieve memory pressure, systemd-oomd will only respect these extended attributes if the unit's cgroup owner, and the owner of the monitored ancestor cgroup are the same. For example, if systemd-oomd is calculating candidates for -.slice, then extended attributes set on descendants of /user.slice/user-1000.slice/user@1000.service/ will be ignored because the descendants are owned by UID 1000, and -.slice is owned by UID 0. But, if calculating candidates for /user.slice/user-1000.slice/user@1000.service/, then extended attributes set on the descendants would be respected.

If this property is set to avoid, the service manager will convey this to systemd-oomd, which will only select this cgroup if there are no other viable candidates.

If this property is set to omit, the service manager will convey this to systemd-oomd, which will ignore this cgroup as a candidate and will not perform any actions on it.

It is recommended to use avoid and omit sparingly, as it can adversely affect systemd-oomd's kill behavior. Also note that these extended attributes are not applied recursively to cgroups under this unit's cgroup.

Defaults to none which means systemd-oomd will rank this unit's cgroup as defined in systemd-oomd.service(8) and oomd.conf(5).