In cgroup v2 we have protection tunables -- currently MemoryLow and
MemoryMin (there will be more in future for other resources, too). The
design of these protection tunables requires not only intermediate
cgroups to propagate protections, but also the units at the leaf of that
resource's operation to accept it (by setting MemoryLow or MemoryMin).
This makes sense from an low-level API design perspective, but it's a
good idea to also have a higher-level abstraction that can, by default,
propagate these resources to children recursively. In this patch, this
happens by having descendants set memory.low to N if their ancestor has
DefaultMemoryLow=N -- assuming they don't set a separate MemoryLow
value.
Any affected unit can opt out of this propagation by manually setting
`MemoryLow` to some value in its unit configuration. A unit can also
stop further propagation by setting `DefaultMemoryLow=` with no
argument. This removes further propagation in the subtree, but has no
effect on the unit itself (for that, use `MemoryLow=0`).
Our use case in production is simplifying the configuration of machines
which heavily rely on memory protection tunables, but currently require
tweaking a huge number of unit files to make that a reality. This
directive makes that significantly less fragile, and decreases the risk
of misconfiguration.
After this patch is merged, I will implement DefaultMemoryMin= using the
same principles.
This was the last kind of accounting still not exposed on for each unit.
Let's fix that.
Note that this is a relatively simplistic approach: we don't expose
per-device stats, but sum them all up, much like cgtop does. This kind
of metric is probably the most interesting for most usecases, and covers
the "systemctl status" output best. If we want per-device stats one day
we can of course always add that eventually.
It's a simple wrapper for resetting both IP and CPU accounting in one
go.
This will become particularly useful when we also needs this to reset IO
accounting (to be added in a later commit).
This adds a new per-service OOMPolicy= (along with a global
DefaultOOMPolicy=) that controls what to do if a process of the service
is killed by the kernel's OOM killer. It has three different values:
"continue" (old behaviour), "stop" (terminate the service), "kill" (let
the kernel kill all the service's processes).
On top of that, track OOM killer events per unit: generate a per-unit
structured, recognizable log message when we see an OOM killer event,
and put the service in a failure state if an OOM killer event was seen
and the selected policy was not "continue". A new "result" is defined
for this case: "oom-kill".
All of this relies on new cgroupv2 kernel functionality: the
"memory.events" notification interface and the "memory.oom.group"
attribute (which makes the kernel kill all cgroup processes
automatically).
This new setting allows configuration of CFS period on the CPU cgroup, instead
of using a hardcoded default of 100ms.
Tested:
- Legacy cgroup + Unified cgroup
- systemctl set-property
- systemctl show
- Confirmed that the cgroup settings (such as cpu.cfs_period_ns) were set
appropriately, including updating the CPU quota (cpu.cfs_quota_ns) when
CPUQuotaPeriodSec= is updated.
- Checked that clamping works properly when either period or (quota * period)
are below the resolution of 1ms, or if period is above the max of 1s.
It would be very wrong if any of the specfier printf calls modified
any of the objects or data being printed. Let's mark all arguments as const
(primarily to make it easier for the reader to see where modifications cannot
occur).
Some controllers (like the CPU controller) have a performance cost that
is non-trivial on certain workloads. While this can be mitigated and
improved to an extent, there will for some controllers always be some
overheads associated with the benefits gained from the controller.
Inside Facebook, the fix applied has been to disable the CPU controller
forcibly with `cgroup_disable=cpu` on the kernel command line.
This presents a problem: to disable or reenable the controller, a reboot
is required, but this is quite cumbersome and slow to do for many
thousands of machines, especially machines where disabling/enabling a
stateful service on a machine is a matter of several minutes.
Currently systemd provides some configuration knobs for these in the
form of `[Default]CPUAccounting`, `[Default]MemoryAccounting`, and the
like. The limitation of these is that Default*Accounting is overrideable
by individual services, of which any one could decide to reenable a
controller within the hierarchy at any point just by using a controller
feature implicitly (eg. `CPUWeight`), even if the use of that CPU
feature could just be opportunistic. Since many services are provided by
the distribution, or by upstream teams at a particular organisation,
it's not a sustainable solution to simply try to find and remove
offending directives from these units.
This commit presents a more direct solution -- a DisableControllers=
directive that forcibly disallows a controller from being enabled within
a subtree.
Previously we tried to be smart: when a new unit appeared and it only
added controllers to the cgroup mask we'd update the cached members mask
in all parents by ORing in the controller flags in their cached values.
Unfortunately this was quite broken, as we missed some conditions when
this cache had to be reset (for example, when a unit got unloaded),
moreover the optimization doesn't work when a controller is removed
anyway (as in that case there's no other way for the parent to iterate
though all children if any other, remaining child unit still needs it).
Hence, let's simplify the logic substantially: instead of updating the
cache on the right events (which we didn't get right), let's simply
invalidate the cache, and generate it lazily when we encounter it later.
This should actually result in better behaviour as we don't have to
calculate the new members mask for a whole subtree whever we have the
suspicion something changed, but can delay it to the point where we
actually need the members mask.
This allows us to simplify things quite a bit, which is good, since
validating this cache for correctness is hard enough.
Fixes: #9512
This changes cg_enable_everywhere() to return which controllers are
enabled for the specified cgroup. This information is then used to
correctly track the enablement mask currently in effect for a unit.
Moreover, when we try to turn off a controller, and this works, then
this is indicates that the parent unit might succesfully turn it off
now, too as our unit might have kept it busy.
So far, when realizing cgroups, i.e. when syncing up the kernel
representation of relevant cgroups with our own idea we would strictly
work from the root to the leaves. This is generally a good approach, as
when controllers are enabled this has to happen in root-to-leaves order.
However, when controllers are disabled this has to happen in the
opposite order: in leaves-to-root order (this is because controllers can
only be enabled in a child if it is already enabled in the parent, and
if it shall be disabled in the parent then it has to be disabled in the
child first, otherwise it is considered busy when it is attempted to
remove it in the parent).
To make things complicated when invalidating a unit's cgroup membershup
systemd can actually turn off some controllers previously turned on at
the very same time as it turns on other controllers previously turned
off. In such a case we have to work up leaves-to-root *and*
root-to-leaves right after each other. With this patch this is
implemented: we still generally operate root-to-leaves, but as soon as
we noticed we successfully turned off a controller previously turned on
for a cgroup we'll re-enqueue the cgroup realization for all parents of
a unit, thus implementing leaves-to-root where necessary.
Let's emphasize that this function checks for the host root cgroup, i.e.
returns false for the root cgroup when we run in a container where
CLONE_NEWCGROUP is used. There has been some confusion around this
already, for example cgroup_context_apply() uses the function
incorrectly (which we'll fix in a later commit).
Just some refactoring, not change in behaviour.
The kernel added support for a new cgroup memory controller knob memory.min in
bf8d5d52ffe8 ("memcg: introduce memory.min") which was merged during v4.18
merge window.
Add MemoryMin to support memory.min.
These lines are generally out-of-date, incomplete and unnecessary. With
SPDX and git repository much more accurate and fine grained information
about licensing and authorship is available, hence let's drop the
per-file copyright notice. Of course, removing copyright lines of others
is problematic, hence this commit only removes my own lines and leaves
all others untouched. It might be nicer if sooner or later those could
go away too, making git the only and accurate source of authorship
information.
This part of the copyright blurb stems from the GPL use recommendations:
https://www.gnu.org/licenses/gpl-howto.en.html
The concept appears to originate in times where version control was per
file, instead of per tree, and was a way to glue the files together.
Ultimately, we nowadays don't live in that world anymore, and this
information is entirely useless anyway, as people are very welcome to
copy these files into any projects they like, and they shouldn't have to
change bits that are part of our copyright header for that.
hence, let's just get rid of this old cruft, and shorten our codebase a
bit.
Files which are installed as-is (any .service and other unit files, .conf
files, .policy files, etc), are left as is. My assumption is that SPDX
identifiers are not yet that well known, so it's better to retain the
extended header to avoid any doubt.
I also kept any copyright lines. We can probably remove them, but it'd nice to
obtain explicit acks from all involved authors before doing that.
This adds a new bus call to service and scope units called
AttachProcesses() that moves arbitrary processes into the cgroup of the
unit. The primary user for this new API is systemd itself: the systemd
--user instance uses this call of the systemd --system instance to
migrate processes if itself gets the request to migrate processes and
the kernel refuses this due to access restrictions.
The primary use-case of this is to make "systemd-run --scope --user …"
invoked from user session scopes work correctly on pure cgroupsv2
environments. There, the kernel refuses to migrate processes between two
unprivileged-owned cgroups unless the requestor as well as the ownership
of the closest parent cgroup all match. This however is not the case
between the session-XYZ.scope unit of a login session and the
user@ABC.service of the systemd --user instance.
The new logic always tries to move the processes on its own, but if
that doesn't work when being the user manager, then the system manager
is asked to do it instead.
The new operation is relatively restrictive: it will only allow to move
the processes like this if the caller is root, or the UID of the target
unit, caller and process all match. Note that this means that
unprivileged users cannot attach processes to scope units, as those do
not have "owning" users (i.e. they have now User= field).
Fixes: #3388
Currently we allowed delegation for alluntis with cgroup backing
except for slices. Let's make this a bit more strict for now, and only
allow this in service and scope units.
Let's also add a generic accessor unit_cgroup_delegate() for checking
whether a unit has delegation turned on that checks the new bool first.
Also, when doing transient units, let's explcitly refuse turning on
delegation for unit types that don#t support it. This is mostly
cosmetical as we wouldn't act on the delegation request anyway, but
certainly helpful for debugging.
The kernel exposes the necessary data in /proc anyway, let's expose it
hence by default.
With this in place "systemctl status -- -.slice" will show accounting
data out-of-the-box now.
This code is very similar in scope and service units, let's unify it in
one function. This changes little for service units, but for scope units
makes sure we go through the cgroup queue, which is something we should
do anyway.
Now that we don't kill control processes anymore, let's at least warn
about any processes left-over in the unit cgroup at the moment of
starting the unit.
Previously it was not possible to select which controllers to enable for
a unit where Delegate=yes was set, as all controllers were enabled. With
this change, this is made configurable, and thus delegation units can
pick specifically what they want to manage themselves, and what they
don't care about.
This makes sure that if we learn via inotify or another event source
that a cgroup is empty, and we checked that this is indeed the case (as
we might get spurious notifications through inotify, as the inotify
logic through the "cgroups.event" is pretty unspecific and might be
trigger for a variety of reasons), then we'll enqueue a defer event for
it, at a priority lower than SIGCHLD handling, so that we know for sure
that if there's waitid() data for a process we used it before
considering the cgroup empty notification.
Fixes: #6608
We are about to add second cgroup-related queue, called
"cgroup_empty_queue", hence let's rename "cgroup_queue" to
"cgroup_realize_queue" (as that is its purpose) to minimize confusion
about the two queues.
Just a rename, no functional changes.
Unfortunately, due to the disagreements in the kernel development community,
CPU controller cgroup v2 support has not been merged and enabling it requires
applying two small out-of-tree kernel patches. The situation is explained in
the following documentation.
https://git.kernel.org/cgit/linux/kernel/git/tj/cgroup.git/tree/Documentation/cgroup-v2-cpu.txt?h=cgroup-v2-cpu
While it isn't clear what will happen with CPU controller cgroup v2 support,
there are critical features which are possible only on cgroup v2 such as
buffered write control making cgroup v2 essential for a lot of workloads. This
commit implements systemd CPU controller support on the unified hierarchy so
that users who choose to deploy CPU controller cgroup v2 support can easily
take advantage of it.
On the unified hierarchy, "cpu.weight" knob replaces "cpu.shares" and "cpu.max"
replaces "cpu.cfs_period_us" and "cpu.cfs_quota_us". [Startup]CPUWeight config
options are added with the usual compat translation. CPU quota settings remain
unchanged and apply to both legacy and unified hierarchies.
v2: - Error in man page corrected.
- CPU config application in cgroup_context_apply() refactored.
- CPU accounting now works on unified hierarchy.
cgroup_context_apply() and friends take CGroupContext and cgroup path as input
and has no way of getting back to the associated Unit and thus uses raw cgroup
path for logging. This makes the log messages difficult to track down.
There's no reason to avoid passing in Unit into these functions. Pass in Unit
and use log_unit*() instead.
While at it, make cgroup_context_apply(), which has no outside users, static.
Also, drop cgroup path from log messages where the path itself isn't too
interesting and can be easily obtained from the unit.
On the unified hierarchy, memory controller implements three control knobs -
low, high and max which enables more useable and versatile control over memory
usage. This patch implements support for the three control knobs.
* MemoryLow, MemoryHigh and MemoryMax are added for memory.low, memory.high and
memory.max, respectively.
* As all absolute limits on the unified hierarchy use "max" for no limit, make
memory limit parse functions accept "max" in addition to "infinity" and
document "max" for the new knobs.
* Implement compatibility translation between MemoryMax and MemoryLimit.
v2:
- Fixed missing else's in config_parse_memory_limit().
- Fixed missing newline when writing out drop-ins.
- Coding style updates to use "val > 0" instead of "val".
- Minor updates to documentation.
CGroupBlockIODeviceBandwith is used to keep track of IO bandwidth limits for
legacy cgroup hierarchies. Unlike the unified hierarchy counterpart
CGroupIODeviceLimit, a CGroupBlockIODeviceBandwiddth records either a read or
write limit and has a couple issues.
* There's no way to clear specific config entry.
* When configs are cleared for an IO direction of a unit, the kernel settings
aren't cleared accordingly creating discrepancies.
This patch updates CGroupBlockIODeviceBandwidth so that it behaves similarly to
CGroupIODeviceLimit - each entry records both rbps and wbps limits and is
cleared if both are at default values after kernel settings are updated.
Currently, there are two cgroup IO limits, bandwidth max for read and write,
and they are hard-coded in various places. This is fine for two limits but IO
is expected to grow more limits - low, high and max limits for bandwidth and
IOPS - and hard-coding each limit won't make sense.
This patch replaces hard-coded limits with an array indexed by
CGroupIOLimitType and accompanying string and default value tables so that new
limits can be added trivially.
On the unified hierarchy, blkio controller is renamed to io and the interface
is changed significantly.
* blkio.weight and blkio.weight_device are consolidated into io.weight which
uses the standardized weight range [1, 10000] with 100 as the default value.
* blkio.throttle.{read|write}_{bps|iops}_device are consolidated into io.max.
Expansion of throttling features is being worked on to support
work-conserving absolute limits (io.low and io.high).
* All stats are consolidated into io.stats.
This patchset adds support for the new interface. As the interface has been
revamped and new features are expected to be added, it seems best to treat it
as a separate controller rather than trying to expand the blkio settings
although we might add automatic translation if only blkio settings are
specified.
* io.weight handling is mostly identical to blkio.weight[_device] handling
except that the weight range is different.
* Both read and write bandwidth settings are consolidated into
CGroupIODeviceLimit which describes all limits applicable to the device.
This makes it less painful to add new limits.
* "max" can be used to specify the maximum limit which is equivalent to no
config for max limits and treated as such. If a given CGroupIODeviceLimit
doesn't contain any non-default configs, the config struct is discarded once
the no limit config is applied to cgroup.
* lookup_blkio_device() is renamed to lookup_block_device().
Signed-off-by: Tejun Heo <htejun@fb.com>
Support for net_cls.class_id through the NetClass= configuration directive
has been added in v227 in preparation for a per-unit packet filter mechanism.
However, it turns out the kernel people have decided to deprecate the net_cls
and net_prio controllers in v2. Tejun provides a comprehensive justification
for this in his commit, which has landed during the merge window for kernel
v4.5:
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=bd1060a1d671
As we're aiming for full support for the v2 cgroup hierarchy, we can no
longer support this feature. Userspace tool such as nftables are moving over
to setting rules that are specific to the full cgroup path of a task, which
obsoletes these controllers anyway.
This commit removes support for tweaking details in the net_cls controller,
but keeps the NetClass= directive around for legacy compatibility reasons.
Ben Boeckel