On 11/20/20 6:39 PM, Marcel Apfelbaum wrote:
+netdev
[Sorry for the second email, I failed to set the text-only mode]
On Fri, Nov 20, 2020 at 7:30 PM Marcel Apfelbaum <mapfe...@redhat.com> wrote:
[...]
---------- Forwarded message ----------
From: Jakub Kicinski <k...@kernel.org>
To: Paolo Abeni <pab...@redhat.com>
Cc: Saeed Mahameed <sa...@kernel.org>, netdev@vger.kernel.org, Jonathan Corbet <cor...@lwn.net>, "David S.
Miller" <da...@davemloft.net>, Shuah Khan <sh...@kernel.org>, linux-...@vger.kernel.org,
linux-kselft...@vger.kernel.org, Marcelo Tosatti <mtosa...@redhat.com>, Daniel Borkmann <dan...@iogearbox.net>
Bcc:
Date: Wed, 4 Nov 2020 11:42:26 -0800
Subject: Re: [PATCH net-next v2 0/3] net: introduce rps_default_mask
On Wed, 04 Nov 2020 18:36:08 +0100 Paolo Abeni wrote:
On Tue, 2020-11-03 at 08:52 -0800, Jakub Kicinski wrote:
On Tue, 03 Nov 2020 16:22:07 +0100 Paolo Abeni wrote:
The relevant use case is an host running containers (with the related
orchestration tools) in a RT environment. Virtual devices (veths, ovs
ports, etc.) are created by the orchestration tools at run-time.
Critical processes are allowed to send packets/generate outgoing, it gets a
network-interface
upstart job just as it does on a real host.
network traffic - but any interrupt is moved away from the related
cores, so that usual incoming network traffic processing does not
happen there.
Still an xmit operation on a virtual devices may be transmitted via ovs
or veth, with the relevant forwarding operation happening in a softirq
on the same CPU originating the packet.
RPS is configured (even) on such virtual devices to move away the
forwarding from the relevant CPUs.
As Saeed noted, such configuration could be possibly performed via some
user-space daemon monitoring network devices and network namespaces
creation. That will be anyway prone to some race: the orchestation tool
may create and enable the netns and virtual devices before the daemon
has properly set the RPS mask.
In the latter scenario some packet forwarding could still slip in the
relevant CPU, causing measurable latency. In all non RT scenarios the
above will be likely irrelevant, but in the RT context that is not
acceptable - e.g. it causes in real environments latency above the
defined limits, while the proposed patches avoid the issue.
Do you see any other simple way to avoid the above race?
Please let me know if the above answers your doubts,
Thanks, that makes it clearer now.
Depending on how RT-aware your container management is it may or may not
be the right place to configure this, as it creates the veth interface.
Presumably it's the container management which does the placement of
the tasks to cores, why is it not setting other attributes, like RPS?
The CPU isolation is done statically at system boot by setting Linux kernel
parameters,
So the container management component, in this case the Machine Configuration
Operator (for Openshift)
or the K8s counterpart can't really help. (actually they would help if a global
RPS mask would exist)
I tried to tweak the rps_cpus mask using the container management stack, but
there
is no sane way to do it, please let me get a little into the details.
The k8s orchestration component that deals with injecting the network device(s)
into the
container is CNI, which is interface based and implemented by a lot of plugins,
making it
hardly feasible to go over all the existing plugins and change them. Also what
about
the 3rd party ones?
Writing a new CNI plugin and chain it into the existing one is also not an
option AFAIK,
they work at the network level and do not have access to sysfs (they handle the
network namespaces).
Even if it would be possible (I don't have a deep CNI understanding), it will
require a cluster global configuration
that is actually needed only for some of the cluster nodes.
CNI chaining would be ugly, agree, but in a typical setting you'd have the CNI
plugin
itself which is responsible to set up the Pod for communication to the outside
world;
part of it would be creation of devices and moving them into the target netns,
and
then you also typically have an agent running in kube-system namespace in the
hostns
to which the CNI plugin talks to via IPC e.g. to set up IPAM and other state.
Such
agent usually sets up all sort of knobs for the networking layer upon bootstrap.
Assuming you have a cluster where only some of the nodes have RT kernel, these
would
likely have special node annotations in K8s so you could select them to run
certain
workloads on it.. couldn't such agent be taught to be RT-aware and set up all
the
needed knobs? Agree it's a bit ugly to change the relevant CNI plugins to be
RT-aware,
but what if you also need other settings in future aside from RPS mask for RT?
At some
point you'd likely end up having to extend these anyway, no?
Another approach is to set the RPS configuration from the inside(of the
container),
but the /sys mount is read only for unprivileged containers, so we lose again.
That leaves us with a host daemon hack:
Since the virtual network devices are created in the host namespace and
then "moved" into the container, we can listen to some udev event
and write to the rps_cpus file after the virtual netdev is created and before
it is moved (as stated above, the work is done by a CNI plugin implementation).
That is of course extremely racy and not a valid solution.
The container orchestration is quite complex, and I'm unsure isolation
and networking configuration are performed (or can be performed) by the
same precess (without an heavy refactor).
On the flip hand, the global rps mask knob looked quite
straightforward to me.
I understand, but I can't shake the feeling this is a hack.
Whatever sets the CPU isolation should take care of the RPS settings.
Sadly it can't be done, please see above.
Possibly I can reduce the amount of new code introduced by this
patchset removing some code duplication
between rps_default_mask_sysctl() and flow_limit_cpu_sysctl(). Would
that make this change more acceptable? Or should I drop this
altogether?
I'm leaning towards drop altogether, unless you can get some
support/review tags from other netdev developers. So far it
appears we only got a down vote from Saeed.
Any solution that will allow the user space to avoid the
network soft IRQs on specific CPUs would be welcome.
The proposed global mask is a solution, maybe there other ways?
Thanks,
Marcel
