SYSTEMD.RESOURCE-CONTROL(5)

📦

version 262.1 · 17 versions

👑 systemd.resource-control(5) 👑

okay bestie, this is systemd.resource-control — the page that tells you how to give your units limits and priorities instead of letting them eat the whole machine 🍓. Let's get into it.

✨ NAME ✨

systemd.resource-control — resource control unit settings 💅

🌸 SYNOPSIS 🌸

slice.slice, scope.scope, service.service, socket.socket, mount.mount,
swap.swap

💖 DESCRIPTION 💖

So. Unit config files for services, slices, scopes, sockets, mount points, and swap devices all share a subset of configuration options for resource-controlling the processes they spawn. Under the hood this leans on the Linux Control Groups (cgroups) kernel concept — organizing processes into a hierarchical tree of named groups so you can actually manage resources. tree of vibes, but make it kernel 👑

This man page lists the config options shared by those six unit types. For the common options that every unit config file has, go read systemd.unit(5). And for the specifics of each individual type, see systemd.slice(5), systemd.scope(5), systemd.service(5), systemd.socket(5), systemd.mount(5), and systemd.swap(5). The resource control options live in the [Slice], [Scope], [Service], [Socket], [Mount], or [Swap] sections — whichever matches your unit type, obviously 💁‍♀️

Also, options that control resources available to programs executed by systemd are listed over in systemd.exec(5). Those options complement the ones here. two pages, one mission 🫶

🎛️ Enabling and disabling controllers 🎛️

Controllers in the cgroup hierarchy are hierarchical, and resource control happens by distributing resource assignments between siblings in branches of the cgroup hierarchy. Good news: you do not need to explicitly enable a cgroup controller for a unit. systemd will tell the kernel to enable a controller for a given unit once that unit has configuration for that controller. For example, when CPUWeight= is set, the cpu controller gets enabled, and when TasksMax= are set, the pids controller gets enabled. On top of that, various controllers may also be enabled explicitly via the MemoryAccounting=/TasksAccounting=/IOAccounting= settings. Because of how the cgroup hierarchy works, controllers get automatically enabled for all parent units and for any sibling units, starting with the lowest level at which a controller is enabled. Units that have a controller enabled may be subject to resource control even if they have zero explicit configuration. no cap 😳

Setting Delegate= enables any delegated controllers for that unit (more below). The delegatee can then enable controllers for its children as it sees fit. In particular, if the delegatee is systemd (in the user@.service unit), it'll repeat the exact same logic as the system instance and enable controllers for user units that have resource limits configured — plus their siblings, their parents, and their parents' siblings. it's giving recursion, and honestly? iconic 🌈

Controllers may be disabled for parts of the cgroup hierarchy with DisableControllers= (see below).

✨ Example 1. Enabling and disabling controllers ✨

                                 -.slice
                                /       \
                         /-----/         \--------------\
                        /                                \
                 system.slice                       user.slice
                   /       \                          /      \
                  /         \                        /        \
                 /           \              user@42.service  user@1000.service
                /             \             Delegate=        Delegate=yes
           a.service       b.slice                             /        \
           CPUWeight=20   DisableControllers=cpu              /          \
                            /  \                      app.slice      session.slice
                           /    \                     CPUWeight=100  CPUWeight=100
                          /      \
                  b1.service   b2.service
                               CPUWeight=1000

In this hierarchy, the cpu controller is enabled for all units shown except b1.service and b2.service. Because there's no explicit configuration for system.slice and user.slice, CPU resources will be split equally between them. Similarly, resources are allocated equally between the children of user.slice and between the child slices beneath user@1000.service. Assuming there's no further configuration of resources or delegation below slices app.slice or session.slice, the cpu controller would not be enabled for units in those slices, and CPU resources would be further allocated using other mechanisms — e.g. based on nice levels. The manager for user 42 has delegation enabled without any controllers, i.e. it can manipulate its subtree of the cgroup hierarchy, but with no resource control. 💀

In the slice system.slice, CPU resources are split 1:6 for service a.service, and 5:6 for slice b.slice, because slice b.slice gets the default value of 100 for cpu.weight when CPUWeight= is not set.

The CPUWeight= setting in service b2.service is neutralized by DisableControllers= in slice b.slice, so the cpu controller would not be enabled for services b1.service and b2.service, and CPU resources would be further allocated using other mechanisms, e.g. based on nice levels.

As described in systemd.unit(5), the settings listed here may be set through the main file of a unit and through drop-in snippets in *.d/ directories. The list of directories searched for drop-ins includes names formed by repeatedly truncating the unit name after all dashes. This is particularly convenient for setting resource limits for a group of units with similar names. chef's kiss 👨‍🍳💋

For example, every user gets their own slice user-nnn.slice. Drop-ins with local configuration that affect user 1000 may be placed in /etc/systemd/system/user-1000.slice, /etc/systemd/system/user-1000.slice.d/*.conf, but also /etc/systemd/system/user-.slice.d/*.conf. This last directory applies to all user slices. 🌈

See the New Control Group Interfaces[1] for an introduction on how to make use of resource control APIs from programs.

🦋 IMPLICIT DEPENDENCIES 🦋

The following dependencies are implicitly added:

Units with the Slice= setting set automatically acquire Requires= and After= dependencies on the specified slice unit.

💅 OPTIONS 💅

Units of the types listed above can have settings for resource control configuration:

🔥 CPU Control 🔥

CPUWeight=weight, StartupCPUWeight=weight — These settings control the cpu controller in the unified hierarchy.

These options accept an integer value or the special string "idle":

If set to an integer value, assign the specified CPU time weight to the processes executed, if the unified control group hierarchy is used on the system. These options control the "cpu.weight" control group attribute. The allowed range is 1 to 10000. Defaults to unset, but the kernel default is 100. For details about this control group attribute, see Control Groups v2[2] and CFS Scheduler[3]. The available CPU time is split up among all units within one slice relative to their CPU time weight. A higher weight means more CPU time, a lower weight means less. more weight = more time, simple as 💯
If set to the special string "idle", mark the cgroup for "idle scheduling", which means it will get CPU resources only when there are no processes not marked in this way to execute in this cgroup or its siblings. This setting corresponds to the "cpu.idle" cgroup attribute.

Note that this value only has an effect on cgroup-v2; for cgroup-v1 it is equivalent to the minimum weight.

While StartupCPUWeight= applies to the startup and shutdown phases of the system, CPUWeight= applies to the normal runtime of the system, and — if the former is not set — also to the startup and shutdown phases. Using StartupCPUWeight= lets you prioritize specific services at boot-up and shutdown differently than during normal runtime. boot szn vs runtime szn 🥾✨

In addition to the resource allocation performed by the cpu controller, the kernel may automatically divide resources based on session-id grouping, see "The autogroup feature" in sched(7). The effect of this feature is similar to the cpu controller with no explicit configuration, so users should be careful not to mistake one for the other. 🙈

Added in version 232.

CPUQuota= — This setting controls the cpu controller in the unified hierarchy.

Assign the specified CPU time quota to the processes executed. Takes a percentage value, suffixed with "%". The percentage specifies how much CPU time the unit shall get at maximum, relative to the total CPU time available on one CPU. Use values > 100% for allotting CPU time on more than one CPU. This controls the "cpu.max" attribute on the unified control group hierarchy and "cpu.cfs_quota_us" on legacy. For details about these control group attributes, see Control Groups v2[2] and CFS Bandwidth Control[4]. Setting CPUQuota= to an empty value unsets the quota.

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

Added in version 213.

CPUQuotaPeriodSec= — This setting controls the cpu controller in the unified hierarchy.

Assign the duration over which the CPU time quota specified by CPUQuota= is measured. Takes a time duration value in seconds, with an optional suffix such as "ms" for milliseconds (or "s" for seconds.) The default setting is 100ms. The period is clamped to the range supported by the kernel, which is [1ms, 1000ms]. Additionally, the period is adjusted up so that the quota interval is also at least 1ms. Setting CPUQuotaPeriodSec= to an empty value resets it to the default. 🌀

This controls the second field of the "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.

Added in version 242.

AllowedCPUs=, StartupAllowedCPUs= — This setting controls the cpuset controller in the unified hierarchy.

Restrict processes to be executed on specific CPUs. Takes a list of CPU indices or ranges separated by either whitespace or commas. CPU ranges are specified by the lower and upper CPU indices separated by a dash. 🔥

Setting AllowedCPUs= or StartupAllowedCPUs= does not 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=.

...and that's where part 1 wraps, bestie — more coming your way 💖

Quick note bestie: while <code>StartupAllowedCPUs=</code> is for the boot-up and shutdown eras of the system, <code>AllowedCPUs=</code> is for the normal runtime era — and if the startup one isn't set, it lowkey covers startup and shutdown too. So like… setting <code>StartupAllowedCPUs=</code> is your lil' glow-up for giving specific services different CPU priority at boot and shutdown than they get during everyday runtime. ✨

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

Added in version 244. 🎀

🌸 Memory Accounting and Control 🌸

💖 `MemoryAccounting=`

This one is the queen of the memory controller in the unified hierarchy. 👑

Turns on process and kernel memory accounting for this unit. Takes a boolean argument. And here's the tea: turning on memory accounting for one unit will also implicitly turn it on for every unit inside the same slice, plus all its parent slices and the units hanging out in there too. No cap, it just spreads. 🫶

The system default for this setting can be steered with DefaultMemoryAccounting= over in systemd-system.conf(5).

Added in version 208. ✨

💜 `MemoryMin=`bytes, `MemoryLow=`bytes, `StartupMemoryLow=`bytes

These settings control the memory controller in the unified hierarchy.

They set up memory usage protection for the processes running in this unit. When the system is reclaiming memory, this unit gets treated like it's using less memory than it actually is — so memory gets preferentially reclaimed from the unprotected units instead. Using MemoryLow= gives you a weaker protection, where memory can still be reclaimed from you to avoid summoning the OOM killer when there's literally nothing else reclaimable left. 😭

For a protection to actually do something, you generally need to set a matching allocation on all the ancestors — which then gets distributed between the children (the root slice is the exception, she doesn't play by these rules). Any MemoryMin= or MemoryLow= allocation that isn't explicitly handed to specific children becomes a shared protection for all of them. And since it's shared, the children just freely compete for the memory. Sibling rivalry fr fr. 💀

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 is genuinely useful if you want to always inherit all of the protection afforded by ancestors. 👑

This controls the "memory.min" or "memory.low" control group attribute. For the deets on that attribute, see Memory Interface Files[5].

While StartupMemoryLow= applies to the startup and shutdown phases of the system, MemoryMin= applies to normal runtime of the system, and if the former is not set also to the startup and shutdown phases. Using StartupMemoryLow= lets you prioritize specific services at boot-up and shutdown differently than during normal runtime. ✨

Added in version 240. 🎀

🦋 `MemoryHigh=`bytes, `StartupMemoryHigh=`bytes

These settings control the memory controller in the unified hierarchy.

They set the throttling limit on memory usage of the executed processes in this unit. Memory usage may go above the limit if it's genuinely unavoidable, but in that case the processes get heavily slowed down and memory is taken away aggressively. This is the main mechanism for controlling a unit's memory usage, bestie — it's the moment. 💅

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]. The effective configuration is reported as EffectiveMemoryHigh= (see also EffectiveMemoryMax=).

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

Added in version 231. ✨

⭐ `MemoryMax=`bytes, `StartupMemoryMax=`bytes

These settings control the memory controller in the unified hierarchy.

They set the absolute limit on memory usage of the executed processes in this unit. If memory usage can't be contained under the limit, the out-of-memory killer is invoked inside the unit. 💀 It is recommended to use MemoryHigh= as the main control mechanism and keep MemoryMax= as the last line of defense — she's the bouncer, not the vibe curator.

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]. The effective configuration is reported as EffectiveMemoryMax= (the value is the most stringent limit of the unit and parent slices and it is capped by physical memory).

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

Added in version 231. 🎀

🍓 `MemorySwapMax=`bytes, `StartupMemorySwapMax=`bytes

These settings control the memory controller in the unified hierarchy.

They set the absolute limit on swap usage of the executed processes in this unit.

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. Alternatively, a percentage value may be specified, which is taken relative to the specified swap size on the system. If assigned the special value "infinity", no swap limit is applied. These settings control the "memory.swap.max" control group attribute. For details about this control group attribute, see Memory Interface Files[5].

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

Added in version 232. ✨

🧚 `MemoryZSwapMax=`bytes, `StartupMemoryZSwapMax=`bytes

These settings control the memory controller in the unified hierarchy.

They set the absolute limit on zswap usage of the processes in this unit. So like, what even is zswap? Okay so — Zswap is a lightweight compressed cache for swap pages. It grabs pages that are in the process of being swapped out and tries to compress them into a dynamically allocated RAM-based memory pool. If the limit specified is hit, no entries from this unit will be stored in the pool until existing entries are faulted back or written out to disk. See the kernel's Zswap[6] documentation for more details. 🔥

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

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

Added in version 253. 🎀

💅 `MemoryZSwapWriteback=`

This setting controls the memory controller in the unified hierarchy.

Takes a boolean argument. Defaults to true if DefaultMemoryZSwapWriteback= is not set. When true, pages stored in the Zswap cache are permitted to be written to the backing storage; false otherwise. This lets you disable writeback of swap pages for IO-intensive applications, while still retaining the ability to store compressed pages in Zswap. Iconic, honestly. See the kernel's Zswap[6] documentation for more details. 👑

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

Added in version 256. ✨

🌈 `AllowedMemoryNodes=`, `StartupAllowedMemoryNodes=`

These settings control the cpuset controller in the unified hierarchy. (Yes bestie, cpuset — we're still in the memory section, don't question it, that's just how the kernel rolls.) 🙈

Restrict processes to be executed on specific memory NUMA nodes. Takes a list of memory NUMA node indices or ranges, separated by either whitespace or commas. Memory NUMA node ranges are specified by the lower and upper NUMA node indices separated by a dash.

Heads up: setting AllowedMemoryNodes= or StartupAllowedMemoryNodes= does not guarantee that all of the memory NUMA nodes will actually be used by the processes, because 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.

Added in version 244. 🦋

CPUSetPartition= — okay so this one sets the cpuset partition type for the processes you're running, and it takes exactly one of "member", "root", or "isolated". It's the boss of the cpuset.cpus.partition cgroup attribute, no cap.

So what do the three vibes mean? When it's "member", the cpuset is just doing normal-mode things, living its life. "root" makes a partition root, which can then go and divvy up CPUs among child cgroups — very main character, handing out slices. And "isolated" gives you full CPU isolation, which is the move for real-time workloads that need dedicated CPU resources with zero interference from other processes. It defaults to whatever the kernel default is, which is "member". For more tea about this control group attribute, go peek at Control Groups v2[2].

Heads up bestie: this setting requires AllowedCPUs= to also be set. It won't work alone, fr fr.

Added in version 261.

🧮 Process Accounting and Control

TasksAccounting= — this one is in charge of the pids controller in the unified hierarchy.

It turns task accounting on for this unit. Takes a boolean argument. If you enable it, the kernel keeps track of the total number of tasks in the unit and its children — and that count includes both kernel threads and userspace processes, with every single thread counted individually. Sneaky detail: turning on tasks accounting for one unit will also implicitly turn it on for all units contained in the same slice, and for all its parent slices and the units contained therein. It just spreads, bestie. The system default for this can be controlled with DefaultTasksAccounting= over in systemd-system.conf(5).

Added in version 227.

TasksMax=N — also controls the pids controller in the unified hierarchy.

This specifies the max number of tasks that may be created in the unit, so the number of tasks accounted for the unit (see above, bestie) stays under a specific limit. You can give it either an absolute number of tasks, or a percentage value taken relative to the configured maximum number of tasks on the system. And if you hand it the special value "infinity", then no tasks limit is applied at all — go wild. This controls the "pids.max" control group attribute. For details about that attribute, see the pids controller[7]. The effective configuration gets reported as EffectiveTasksMax=.

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

Added in version 227.

💾 IO Accounting and Control

IOAccounting= — controls the io controller in the unified hierarchy.

Turns on Block I/O accounting for this unit, if the unified control group hierarchy is used on the system. Takes a boolean argument. And yep, same spread situation: turning on block I/O accounting for one unit will also implicitly turn it on for all units contained in the same slice, and all for its parent slices and the units contained therein. The system default for this setting may be controlled with DefaultIOAccounting= in systemd-system.conf(5).

Added in version 230.

IOWeight=weight, StartupIOWeight=weight — these two control the io controller in the unified hierarchy.

They set the default overall block I/O weight for the executed processes, if the unified control group hierarchy is used on the system. Takes a single weight value (between 1 and 10000) to set the default block I/O weight. This controls the "io.weight" control group attribute, which defaults to 100. For details about this control group attribute, see IO Interface Files[8]. The available I/O bandwidth is split up among all units within one slice relative to their block I/O weight — so a higher weight means more I/O bandwidth, a lower weight means less. Simple math, gorgeous.

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 lets you prioritize specific services at boot-up and shutdown differently than during runtime. Glow-up on boot, chill at runtime, we love to see it.

Added in version 230.

IODeviceWeight=device weight — controls the io controller in the unified hierarchy.

Sets the per-device overall block I/O weight for the executed processes, if the unified control group hierarchy is used on the system. It takes a space-separated pair of a file path and a weight value to specify the device-specific weight value, between 1 and 10000. (Example: "/dev/sda 1000"). The file path may be specified as a path to a block device node, or as any other file — in which case the backing block device of the file system of the file gets determined. This controls the "io.weight" control group attribute, which defaults to 100. Use this option multiple times to set weights for multiple devices. For details about this control group attribute, see IO Interface Files[8].

Okay important tea: the specified device node should reference a block device that has an I/O scheduler associated — so it 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 attempts to discover the correct originating device backing the file system of the specified path. And ngl, this only works correctly 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 not RAID and volume management storage devices.

Added in version 230.

IOReadBandwidthMax=device bytes, IOWriteBandwidthMax=device bytes — these control the io controller in the unified hierarchy.

They set the per-device overall block I/O bandwidth maximum limit for the executed processes, if the unified control group hierarchy is used on the system. Heads up: this limit is not work-conserving — the executed processes are not allowed to use more even if the device has idle capacity. Takes a space-separated pair of a file path and a bandwidth value (in bytes per second) to specify the device-specific bandwidth. The file path may be a path to a block device node, or any other file, in which case the backing block device of the file system of the file is used. If the bandwidth is suffixed with K, M, G, or T, the specified bandwidth is parsed as Kilobytes, Megabytes, Gigabytes, or Terabytes, respectively, to the base of 1000. (Example: "/dev/disk/by-path/pci-0000:00:1f.2-scsi-0:0:0:0 5M"). This controls the "io.max" control group attributes. Use this option multiple times to set bandwidth limits for multiple devices. For details about this control group attribute, see IO Interface Files[8].

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

Added in version 230.

IOReadIOPSMax=device IOPS, IOWriteIOPSMax=device IOPS — these control the io controller in the unified hierarchy.

They set the per-device overall block I/O IOs-Per-Second maximum limit for the executed processes, if the unified control group hierarchy is used on the system. Same warning as before: this limit is not work-conserving and the executed processes are not allowed to use more even if the device has idle capacity. Takes a space-separated pair of a file path and an IOPS value to specify the device-specific IOPS. The file path may be a path to a block device node, or any other file, in which case the backing block device of the file system of the file is used. If the IOPS is suffixed with K, M, G, or T, the specified IOPS is parsed as KiloIOPS, MegaIOPS, GigaIOPS, or TeraIOPS, respectively, to the base of 1000. (Example: "/dev/disk/by-path/pci-0000:00:1f.2-scsi-0:0:0:0 1K"). This controls the "io.max" control group attributes. Use this option multiple times to set IOPS limits for multiple devices. For details about this control group attribute, see IO Interface Files[8].

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

Added in version 230.

IODeviceLatencyTargetSec=device target — controls the io controller in the unified hierarchy.

Sets the per-device average target I/O latency for the executed processes, if the unified control group hierarchy is used on the system. Takes a file path and a timespan separated by a space to specify the device-specific latency target. (Example: "/dev/sda 25ms"). The file path may be specified as a path to a block device node, or as any other file, in which case the backing block device of the file system of the file is determined. This controls the "io.latency" control group attribute. Use this option multiple times to set a latency target for multiple devices. For details about this control group attribute, see IO Interface Files[8].

Implies "IOAccounting=yes" — it just turns that on for you, bestie.

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.

Added in version 240.

🌐 Network Accounting and Control

IPAccounting= — takes a boolean argument. If true, it turns on IPv4 and IPv6 network traffic accounting for packets sent or received by the unit. When this option is on, all IPv4 and IPv6 sockets created by any process of the unit are accounted for. 💅

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). Now here's the spicy part: 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. It's giving "what happens in the socket unit stays in the socket unit."

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

Note that this functionality is currently only available for system services, not for per-user services. 😭

Added in version 235.

IPAddressAllow=ADDRESS[/PREFIXLENGTH]..., IPAddressDeny=ADDRESS[/PREFIXLENGTH]... — bestie, this is the IP firewall for your unit 🔥 It turns on network traffic filtering for IP packets sent and received over AF_INET and AF_INET6 sockets.

Both directives take a space-separated list of IPv4 or IPv6 addresses, and each one can optionally have an address prefix length in bits tacked on after a / character. No suffix? Then the address counts as a host address, meaning the filter covers the whole address (that's 32 bits for IPv4, 128 bits for IPv6).

The access lists you set up here get applied to all the sockets created by processes of this unit (or, for socket units, associated with it). And the way they work is lowkey additive: 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 gets filtered by these settings. For ingress traffic, the source IP address is checked against these access lists; for egress traffic, the destination IP address is checked. And the rules get applied in this order:

Access is granted when the checked IP address matches an entry in the IPAddressAllow= list.
Otherwise, access is denied when the checked IP address matches an entry in the IPAddressDeny= list.
Otherwise, access is granted.

So if you're trying to build an allow-listing IP firewall, here's the move 💅 the manpage recommends using a IPAddressDeny=any setting on an upper-level slice unit (like the root slice -.slice, or the slice containing all system services system.slice — peep systemd.special(7) for the details on those slice units), plus individual per-service IPAddressAllow= lines permitting network access to relevant services, and only them. Periodt.

Okay important one: 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. And conversely, the IP access list configured for the service is not applied to any sockets passed into the service via socket activation. So yeah, it's usually a good idea to replicate the IP access lists on both the socket and the service unit. That said, it may make sense to keep one list more open and the other one more restricted, depending on your use case. iykyk ✨

If you use these settings multiple times in the same unit, the specified lists are combined. And if you assign an empty string to these settings, the specific access list gets reset and all previous settings undone. Clean slate energy 🧼

In place of explicit IPv4 or IPv6 address and prefix length specifications, you can use a small set of symbolic names. Here's what's defined:

🎀 Table 1. Special address/network names

any — definition: 0.0.0.0/0 ::/0 — meaning: Any host
localhost — definition: 127.0.0.0/8 ::1/128 — meaning: All addresses on the local loopback
link-local — definition: 169.254.0.0/16 fe80::/64 — meaning: All link-local IP addresses
multicast — definition: 224.0.0.0/4 ff00::/8 — meaning: All IP multicasting addresses

Heads up bestie 😭 these settings might not be supported on some systems (for example if eBPF control group support isn't enabled in the underlying kernel or container manager). In that case these settings will have no effect. So if you want compatibility with such systems, the recommendation is to not exclusively rely on them for IP security. Don't put all your eggs in this basket fr fr.

This option cannot be bypassed by prefixing + to the executable path in the service unit, since it applies to the whole control group. No sneaking around it 🙈

Added in version 235.

SocketBindAllow=bind-rule, SocketBindDeny=bind-rule — this one configures restrictions on the ability of unit processes to invoke bind(2) on a socket. You can define both allow and deny rules that restrict which addresses a socket may be bound to. Very gatekeeping of you, honestly 💅

bind-rule describes socket properties like 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 you don't specify it, 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 you don't specify it, a rule will be matched for any transport protocol.

An optional ip-port value must lie within the 1...65535 interval inclusively — so dynamic port 0 is not allowed, bestie, don't even try. 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=, the maximum allowed number of assignments is 128.

Binding to a socket is allowed when a socket address matches an entry in the SocketBindAllow= list.
Otherwise, binding is denied when the socket address matches an entry in the SocketBindDeny= list.
Otherwise, binding is allowed.

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

And note: these settings apply to any bind(2) system call invocation by the unit processes, regardless of which network namespace they're placed in. Or, in other words: changing the network namespace is not a suitable mechanism for escaping these restrictions on bind(). You can't outrun the vibes check on this one 😭

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 cannot be bypassed by prefixing + to the executable path in the service unit, since it applies to the whole control group. Same energy as before 🫶

Added in version 249.

RestrictNetworkInterfaces= — takes a list of space-separated network interface names, and this option restricts the network interfaces that processes of this unit can use. By default, processes can only use the network interfaces that are listed (that's an allow-list). But if the first character of the rule is ~, the effect is inverted: now the processes can only use network interfaces that are not listed (a deny-list). Flip the script ✨

This option can appear multiple times, in which case the network interface names are merged. And if the empty string is assigned, the set is reset and all prior assignments will have no effect.

If you specify both types of this option (i.e. allow-listing and deny-listing), the first one 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 on its type and the default action. So order matters, bestie 👑

The loopback interface ("lo") is not treated in any special way — you have to configure it explicitly in the unit file. No free passes 🎀

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.

This option cannot be bypassed by prefixing + to the executable path in the service unit, since it applies to the whole control group. It's giving unavoidable 💀

Added in version 250.

BindNetworkInterface= — okay so this one takes the name of a network interface, and it makes every socket created by processes of this unit get bound to that specified network interface. 🦋

It's especially useful for confining a process to a VRF when the program doesn't offer native support for it, and it's equivalent to running the program using ip vrf exec. Iconic.

On systems using nss-resolve, the interface used for DNS resolution can be chosen by using the SYSTEMD_NSS_RESOLVE_IFINDEX environment variable. ✨

The feature is implemented with cgroup/sock_create cgroup-bpf hooks.

Example:

[Service]
BindNetworkInterface=vrf-mgmt

This option cannot be bypassed by prefixing + to the executable path in the service unit, because it applies to the whole control group. No sneaky workarounds bestie, it's giving airtight. 💅

Added in version 260.

NFTSet=family:table:set — this setting is your little method for integrating dynamic cgroup, user and group IDs into firewall rules with NFT[9] sets. The whole point is being able to use those IDs as selectors in firewall rules easily, which in turn allows more fine grained filtering. NFT rules for cgroup matching use numeric cgroup IDs, which change every time a service is restarted, making them hard to use in systemd environment otherwise. And the dynamic and random IDs used by DynamicUser= can be integrated with this setting too. 🎀

This option expects a whitespace separated list of NFT set definitions. Each definition consists of a colon-separated tuple of source type (one of "cgroup", "user" or "group"), NFT address family (one of "arp", "bridge", "inet", "ip", "ip6", or "netdev"), table name and set name. The names of tables and sets must conform to lexicographic restrictions of NFT table names. The type of the element used in the NFT filter must match the type implied by the directive ("cgroup", "user" or "group") as shown in the table below. When a control group or a unit is realized, the corresponding ID will be appended to the NFT sets and it will be removed when the control group or unit is removed. systemd only inserts elements to (or removes from) the sets, so the related NFT rules, tables and sets must be prepared elsewhere in advance. Failures to manage the sets will be ignored. 🙈

Table 2. Defined source type values

Source type "cgroup" — control group ID — corresponding NFT type name "cgroupsv2"
Source type "user" — user ID — corresponding NFT type name "meta skuid"
Source type "group" — group ID — corresponding NFT type name "meta skgid"

If the firewall rules are reinstalled so that the contents of NFT sets are destroyed, command systemctl daemon-reload can be used to refill the sets. 🫶

Example:

[Service]
NFTSet=cgroup:inet:filter:my_service user:inet:filter:serviceuser

Corresponding NFT rules:

table inet filter {
        set my_service {
                type cgroupsv2
        }
        set serviceuser {
                typeof meta skuid
        }
        chain x {
                socket cgroupv2 level 2 @my_service accept
                drop
        }
        chain y {
                meta skuid @serviceuser accept
                drop
        }
}

This option is only available for system services and is not supported for services running in per-user instances of the service manager. 😭

Added in version 255.

✨ BPF Programs ✨

IPIngressFilterPath=BPF_FS_PROGRAM_PATH, IPEgressFilterPath=BPF_FS_PROGRAM_PATH — these add custom network traffic filters implemented as BPF programs, applying to all IP packets sent and received over AF_INET and AF_INET6 sockets. Takes an absolute path to a pinned BPF program in the BPF virtual filesystem (/sys/fs/bpf/). 🧚

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. Periodt.

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. So like, it's 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 use case. 🌈

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 attach your filter manually (requires Delegate=yes) instead of using this setting. 😤

Added in version 243.

BPFProgram=type:program-path — BPFProgram= allows attaching custom BPF programs to the cgroup of a unit. (This generalizes the functionality exposed via IPEgressFilterPath= and IPIngressFilterPath= for other hooks.) 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 bpf.h[10]. Also refer to the general BPF documentation[11]. 🔥

The specification of BPF program consists of a pair of BPF program type and program path in the file system, with : as the separator: type:program-path.

The BPF program type is equivalent to the BPF attach type used in bpftool(8). It may be one of egress, ingress, sock_create, sock_ops, device, bind4, bind6, connect4, connect6, post_bind4, post_bind6, sendmsg4, sendmsg6, sysctl, recvmsg4, recvmsg6, getsockopt, or setsockopt. 👑

The specified program path must be an absolute path referencing a BPF program inode in the bpffs file system (which generally means it must begin with /sys/fs/bpf/). If a specified program does not exist (i.e. has not been uploaded to the BPF subsystem of the kernel yet), it will not be installed but unit activation will continue (a warning will be printed to the logs). Lowkey fine, highkey noted. 📝

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

Multiple assignments of the same program type/path pair have the same effect as a single assignment: the program will be attached 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

Added in version 249.

🌸 Device Access 🌸

DeviceAllow= — this one controls access to specific device nodes by the executed processes. It takes two space-separated strings: a device node specifier followed by a combination of r, w, m to control reading, writing, or creation of the specific device nodes by the unit (mknod), respectively. This functionality is implemented using eBPF filtering. 🦋

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.

Okay so real talk bestie 💅 — allow lists you define this way should only reference device groups that are actually resolvable at the time the unit is started. Any device groups that aren't resolvable then? They just don't get added to the device allow list, no drama, they're just not there. And if you wanna work around that lil limitation, the move is to extend 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 it's missing. Like so:

...
[Unit]
Wants=modprobe@loop.service
After=modprobe@loop.service

[Service]
DeviceAllow=block-loop
DeviceAllow=/dev/loop-control
...

Heads up bestie: this option cannot be bypassed by prefixing + to the executable path in the service unit, because it applies to the whole control group. No sneaky escapes, periodt.

Added in version 208.

`DevicePolicy=auto|closed|strict`

Controls the whole policy for allowing device access, so like, the vibe check for what devices get through:

strict

Means to only allow types of access that are explicitly specified. Nothing extra, nothing implied, we're being picky ✨

Added in version 208.

closed

In addition, allows access to standard pseudo devices including /dev/null, /dev/zero, /dev/full, /dev/random, and /dev/urandom. The basics, but make it allowed 🫶

Added in version 208.

auto

In addition, allows access to all devices if no explicit DeviceAllow= is present. This is the default, so if you didn't say anything, this is what you got, no cap.

Added in version 208.

Same energy as before: this option cannot be bypassed by prefixing + to the executable path in the service unit, because it applies to the whole control group. Fr fr.

Added in version 208.

🎛️ Control Group Management 🎛️

`Slice=`

The name of the slice unit to place the unit in. Defaults to system.slice for all non-instantiated units of all unit types — except for slice units themselves, see below, they're special 👑. Instance units are by default placed in a subslice of system.slice that is named after the template name.

You can use this option to arrange systemd units in a hierarchy of slices, and each of those slices might have resource settings applied to it. Nested resource control, it's giving organizational excellence 💖

For units of type slice, the only accepted value for this setting is the parent slice. And since the name of a slice unit already implies the parent slice, it's hence redundant to ever set this parameter directly for slice units. So like, don't bother, it's already implied bestie 🙈

Special care should be taken when relying on the default slice assignment in templated service units that have DefaultDependencies=no set — go read systemd.service(5), section "Default Dependencies" for the details. Don't skip this one.

Added in version 208.

`Delegate=`

Turns on delegation of further resource control partitioning to processes of the unit. Units where this is enabled may create and manage their own private subhierarchy of control groups below the control group of the unit itself. And for unprivileged services (i.e. those using the User= setting) the unit's control group will be made accessible to the relevant user. Giving them the keys to their own lil kingdom 🔑

When enabled, the service manager will refrain from manipulating control groups or moving processes below the unit's control group, so that a clear concept of ownership is established: the control group tree at the level of the unit's control group and above (i.e. towards the root control group) is owned and managed by the service manager of the host, while the control group tree below the unit's control group is owned and managed by the unit itself. Clean boundary, no crossover, iconic ✨

Takes either a boolean argument or a (possibly empty) list of control group controller names. Let's break it down:

If true, delegation is turned on, and all supported controllers are enabled for the unit, making them available to the unit's processes for management.
If false, delegation is turned off entirely (and no additional controllers are enabled).
If set to a list of controllers, delegation is turned on, and the specified controllers are enabled for the unit.
Assigning the empty string will enable delegation, but reset the list of controllers, and all assignments prior to this will have no effect. Poof, gone 💀

Note that additional controllers other than the ones specified might be made available as well, depending on configuration of the containing slice unit or other units contained in it. Defaults to false.

Real talk: 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. Sorry bestie, that's just how it is 😭

The following controller names may be specified: cpu, cpuset, io, memory, pids, bpf-firewall, bpf-devices, bpf-foreign, bpf-socket-bind, bpf-restrict-network-interfaces, and bpf-bind-network-interface.

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 are not covered here yet, as delegation is either not supported at all for them or not defined cleanly. So like, it's a lil messy, iykyk 🌈

Note that because of the hierarchical nature of cgroup hierarchy, any controllers that are delegated will be enabled for the parent and sibling units of the unit with delegation. It spreads, bestie, that's the whole point of a hierarchy.

For further details on the delegation model consult Control Group APIs and Delegation[12].

Added in version 218.

`DelegateSubgroup=`

Place unit processes in the specified subgroup of the unit's control group. Takes a valid control group name (not a path!) as parameter, or an empty string to turn this feature off. Defaults to off. The control group name must be usable as filename and avoid conflicts with the kernel's control group attribute files — i.e. cgroup.procs is not an acceptable name, since the kernel exposes a native control group attribute file by that name. So don't do that, that's a collision 💀

This option has no effect unless control group delegation is turned on via Delegate=, see above. Note that this setting only applies to "main" processes of a unit, i.e. for services to ExecStart=, but not for ExecReload= and similar. If delegation is enabled, the latter are always placed inside a subgroup named .control. The specified subgroup is automatically created (and potentially ownership is passed to the unit's configured user/group) when a process is started in it. Autocreated, auto-owned, slay 🎀

This option is useful to avoid manually moving the invoked process into a subgroup after it has been started. Since no processes should live in inner nodes of the control group tree, it is almost always necessary to run the main ("supervising") process of a unit that has delegation turned on in a subgroup.

Added in version 254.

`DisableControllers=`

Disables controllers from being enabled for a unit's children. If a controller listed is already in use in its subtree, the controller will be removed from the subtree. This can be used to avoid configuration in child units from being able to implicitly or explicitly enable a controller. Defaults to empty.

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. So like, blank = clean slate, got it 💅

It may not be possible to disable a controller after units have been started, 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. Once it's loose, it's loose 🙈

Added in version 240.

💀 Memory Pressure Control 💀

`ManagedOOMSwap=auto|kill`, `ManagedOOMMemoryPressure=auto|kill`

Specifies how systemd-oomd.service(8) will act on this unit's cgroups. Defaults to auto.

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. Yeah bestie, that's the whole cgroup getting yeeted 💀 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. It ties itself in automatically, main character energy 👑

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. So auto is not a full shield, just so you know 😭

Added in version 247.

`ManagedOOMMemoryPressureLimit=`

Overrides the default memory pressure limit set by oomd.conf(5) for the cgroup of this unit. Takes a percentage value between 0% and 100%, inclusive. Defaults to 0%, which means to use the default set by oomd.conf(5). This property is ignored unless ManagedOOMMemoryPressure=kill. So like, if you didn't set kill, this does literally nothing, don't waste your time fr fr.

Added in version 247.

`ManagedOOMMemoryPressureDurationSec=`

Overrides the default memory pressure duration set by oomd.conf(5) for the cgroup of this unit. The specified value supports a time unit such as "ms" or "μs" — see systemd.time(7) for details on the permitted syntax. Must be set to either empty or a value of at least 1s. Defaults to empty, which means to use the default set by oomd.conf(5). This property is ignored unless ManagedOOMMemoryPressure=kill.

Added in version 257.

`ManagedOOMPreference=none|avoid|omit`

Allows deprioritizing or omitting this unit's cgroup as a candidate when systemd-oomd needs to act. Requires support for extended attributes (see xattr(7)) in order to use avoid or omit. So like, if your filesystem ain't got xattrs, those two options are just not happening bestie 🌸

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

okay so when systemd-oomd is out here calculating candidates to relieve memory pressure, it will only respect these extended attributes if the unit's cgroup is owned by root, or if the unit's cgroup owner and the owner of the monitored ancestor cgroup are the same 💅

Example time, bestie: if systemd-oomd is calculating candidates for -.slice, then extended attributes set on descendants of /user.slice/user-1000.slice/user@1000.service/ get ignored, because the descendants are owned by UID 1000 and -.slice is owned by UID 0. But if it's 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).

Added in version 248.

OOMRules= — takes a space-separated list of OOM ruleset names, no cap 🫶

The rulesets are defined in .oomrule files placed in /etc/systemd/oomd/rules.d/, /run/systemd/oomd/rules.d/, /usr/local/lib/systemd/oomd/rules.d/, or /usr/lib/systemd/oomd/rules.d/. When set, systemd-oomd.service(8) will monitor this unit's cgroup and evaluate the specified rulesets against it. Each ruleset defines conditions (such as memory pressure or swap usage thresholds) and an action to take when those conditions are met. See oomd.conf(5) for details on the available ruleset options.

Setting this property will also result in After= and Wants= dependencies on systemd-oomd.service unless DefaultDependencies=no. 🦋

Defaults to an empty list, which means no rulesets are applied. Note that each monitored cgroup incurs a per-interval walk of its descendant cgroup tree, so monitoring very large numbers of cgroups via OOMRules= may have a measurable performance impact. 😭

Added in version 261.

MemoryPressureWatch= — controls memory pressure monitoring for invoked processes. Takes a boolean or one of "auto" and "skip".

If "no", it tells the service not to watch for memory pressure events, by setting the $MEMORY_PRESSURE_WATCH environment variable to the literal string /dev/null. If "yes", it tells the service to watch for memory pressure events. This enables memory accounting for the service, and ensures the memory.pressure cgroup attribute file is accessible for reading and writing by the service's user. It then sets the $MEMORY_PRESSURE_WATCH environment variable for processes invoked by the unit to the file system path to this file. The threshold information configured with MemoryPressureThresholdSec= is encoded in the $MEMORY_PRESSURE_WRITE environment variable. If the "auto" value is set, the protocol is enabled if memory accounting is anyway enabled for the unit, and disabled otherwise. If set to "skip", the logic is neither enabled, nor disabled and the two environment variables are not set. 💅

Note that services are free to use the two environment variables, but it is unproblematic if they ignore them. Memory pressure handling must be implemented individually in each service, and usually means different things for different software. For further details on memory pressure handling see Resource Pressure Handling in systemd[13].

Services implemented using sd-event(3) may use sd_event_add_memory_pressure(3) to watch for and handle memory pressure events. ✨

If not explicit set, defaults to the DefaultMemoryPressureWatch= setting in systemd-system.conf(5).

Added in version 254.

MemoryPressureThresholdSec= — sets the memory pressure threshold time for the memory pressure monitor as configured via MemoryPressureWatch=.

Specifies the maximum allocation latency before a memory pressure event is signalled to the service, per 2s window. If not specified, defaults to the DefaultMemoryPressureThresholdSec= setting in systemd-system.conf(5) (which in turn defaults to 200ms). The specified value expects a time unit such as "ms" or "μs", see systemd.time(7) for details on the permitted syntax. 🕰️

Added in version 254.

CPUPressureWatch= — controls CPU pressure monitoring for invoked processes. Takes a boolean or one of "auto" and "skip".

If "no", it tells the service not to watch for CPU pressure events, by setting the $CPU_PRESSURE_WATCH environment variable to the literal string /dev/null. If "yes", it tells the service to watch for CPU pressure events. This ensures the cpu.pressure cgroup attribute file is accessible for reading and writing by the service's user. It then sets the $CPU_PRESSURE_WATCH environment variable for processes invoked by the unit to the file system path to this file. The threshold information configured with CPUPressureThresholdSec= is encoded in the $CPU_PRESSURE_WRITE environment variable. If the "auto" value is set, the protocol is enabled if CPU resource controls are configured for the unit (e.g. because CPUWeight= or CPUQuota= is set), and disabled otherwise. If set to "skip", the logic is neither enabled, nor disabled and the two environment variables are not set. 🌸

Note that services are free to use the two environment variables, but it is unproblematic if they ignore them. CPU pressure handling must be implemented individually in each service, and usually means different things for different software.

Services implemented using sd-event(3) may use sd_event_add_cpu_pressure(3) to watch for and handle CPU pressure events. 🔥

If not explicitly set, defaults to the DefaultCPUPressureWatch= setting in systemd-system.conf(5).

Added in version 261.

CPUPressureThresholdSec= — sets the CPU pressure threshold time for the CPU pressure monitor as configured via CPUPressureWatch=.

Specifies the maximum CPU stall time before a CPU pressure event is signalled to the service, per 2s window. If not specified, defaults to the DefaultCPUPressureThresholdSec= setting in systemd-system.conf(5) (which in turn defaults to 200ms). The specified value expects a time unit such as "ms" or "μs", see systemd.time(7) for details on the permitted syntax. 🕰️

Added in version 261.

IOPressureWatch= — controls IO pressure monitoring for invoked processes. Takes a boolean or one of "auto" and "skip".

If "no", it tells the service not to watch for IO pressure events, by setting the $IO_PRESSURE_WATCH environment variable to the literal string /dev/null. If "yes", it tells the service to watch for IO pressure events. This enables IO accounting for the service, and ensures the io.pressure cgroup attribute file is accessible for reading and writing by the service's user. It then sets the $IO_PRESSURE_WATCH environment variable for processes invoked by the unit to the file system path to this file. The threshold information configured with IOPressureThresholdSec= is encoded in the $IO_PRESSURE_WRITE environment variable. If the "auto" value is set, the protocol is enabled if IO accounting is anyway enabled for the unit (e.g. because IOWeight= or IODeviceWeight= is set), and disabled otherwise. If set to "skip", the logic is neither enabled, nor disabled and the two environment variables are not set. 🦋

Note that services are free to use the two environment variables, but it is unproblematic if they ignore them. IO pressure handling must be implemented individually in each service, and usually means different things for different software.

Services implemented using sd-event(3) may use sd_event_add_io_pressure(3) to watch for and handle IO pressure events. 💖

If not explicitly set, defaults to the DefaultIOPressureWatch= setting in systemd-system.conf(5).

Added in version 261.

IOPressureThresholdSec= — sets the IO pressure threshold time for the IO pressure monitor as configured via IOPressureWatch=.

Specifies the maximum IO stall time before an IO pressure event is signalled to the service, per 2s window. If not specified, defaults to the DefaultIOPressureThresholdSec= setting in systemd-system.conf(5) (which in turn defaults to 200ms). The specified value expects a time unit such as "ms" or "μs", see systemd.time(7) for details on the permitted syntax. 🕰️

Added in version 261.

Coredump Control 💀

CoredumpReceive= — takes a boolean argument. This setting is used to enable coredump forwarding for containers that belong to this unit's cgroup. Units with CoredumpReceive=yes must also be configured with Delegate=yes. Defaults to false.

When systemd-coredump is handling a coredump for a process from a container, if the container's leader process is a descendant of a cgroup with CoredumpReceive=yes and Delegate=yes, then systemd-coredump will attempt to forward the coredump to systemd-coredump within the container. See also systemd-coredump(8). 🎀

Added in version 255.

HISTORY 🕰️✨

systemd 252 — options for controlling the Legacy Control Group Hierarchy (Control Groups version 1[14]) are now fully deprecated: CPUShares=weight, StartupCPUShares=weight, MemoryLimit=bytes, BlockIOAccounting=, BlockIOWeight=weight, StartupBlockIOWeight=weight, BlockIODeviceWeight=device weight, BlockIOReadBandwidth=device bytes, BlockIOWriteBandwidth=device bytes. Please switch to the unified cgroup hierarchy. 👑
systemd 258 — CPUAccounting= setting is deprecated, because it is always available on the unified cgroup hierarchy and such setting has no effect. 💀

📎 NOTES ✨

okay so bestie, this is the receipts section 💅 — every single footnote that got referenced up there lives here. No cap, if you wanna go deeper on literally any of the control-group lore, these are the sources. Go read them, they ate. 🍓

New Control Group Interfaces — https://systemd.io/CONTROL_GROUP_INTERFACE
Control Groups v2 — https://docs.kernel.org/admin-guide/cgroup-v2.html
CFS Scheduler — https://docs.kernel.org/scheduler/sched-design-CFS.html
CFS Bandwidth Control — https://docs.kernel.org/scheduler/sched-bwc.html
Memory Interface Files — https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files
Zswap — https://docs.kernel.org/admin-guide/mm/zswap.html
pids controller — https://docs.kernel.org/admin-guide/cgroup-v2.html#pid
IO Interface Files — https://docs.kernel.org/admin-guide/cgroup-v2.html#io-interface-files
NFT — https://netfilter.org/projects/nftables/index.html
bpf.h — https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/plain/include/uapi/linux/bpf.h
BPF documentation — https://docs.kernel.org/bpf/
Control Group APIs and Delegation — https://systemd.io/CGROUP_DELEGATION
Resource Pressure Handling in systemd — https://systemd.io/PRESSURE
Control Groups version 1 — https://docs.kernel.org/admin-guide/cgroup-v1/index.html

and that's the whole page, bestie!! 🎀✨ you made it to the very end, which is honestly iconic of you. this edition was systemd 262.1, and yes we were hanging out in SYSTEMD.RESOURCE-CONTROL(5) this whole time. go forth and limit your CPU shares responsibly, mwah 💖🦋👑

📦 where this page comes from

arch rolling: systemd 262.1-1 · LGPL-2.1-or-later AND CC0-1.0 AND GPL-2.0-or-later AND MIT-0
💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash on 2026-10-19: a language model's rewrite of the original, under its terms

SYSTEMD.RESOURCE-CONTROL(5)