Re: [PATCH kvmtool 9/9] arm64: Add support for KVM_ARM_VCPU_PMU_V3_SET_PMU

[Alexandru Elisei <alexandru.elisei@arm.com>]

Hi Marc,

On Sat, Jan 08, 2022 at 01:27:17PM +0000, Marc Zyngier wrote:
> On Fri, 07 Jan 2022 12:10:05 +0000,
> Alexandru Elisei <alexandru.elisei@arm.com> wrote:
> > 
> > Hi Marc,
> > 
> > On Tue, Jan 04, 2022 at 02:39:59PM +0000, Marc Zyngier wrote:
> > > On Mon, 15 Nov 2021 16:57:05 +0000,
> > > Alexandru Elisei <alexandru.elisei@arm.com> wrote:
> > > > 
> > > > The KVM_ARM_VCPU_PMU_V3_CTRL(KVM_ARM_VCPU_PMU_V3_SET_PMU) VCPU ioctl is
> > > > used to assign a physical PMU to the events that KVM creates when emulating
> > > > the PMU for that VCPU. This is useful on heterogeneous systems, when there
> > > > is more than one hardware PMU present.
> > > > 
> > > > The assumption that is made in the implementation is that the user will
> > > > pin the kvmtool process on a set of CPUs that share the same PMU. This
> > > > allows kvmtool to set the same PMU for all VCPUs from the main thread,
> > > > instead of in the individual VCPU threads.
> > > 
> > > May I suggest a slightly different use model? Ideally, you'd be able
> > > to run the vcpu threads on the CPUs matching the PMU affinity, and
> > > leave all the other threads to roam on other CPUs.
> > 
> > Right now, the only way for userspace to make kvmtool run on a particular
> > set of CPUs in a heterogeneous configuration is to use taskset, which means
> > the entire kvmtool process ends up being pinned on a subset of CPUs which
> > have the same PMU. I would like to keep this approach, as it's simple and
> > straightforward to implement in kvmtool, and it's easy to change in the
> > future if there's an incentive to do so.
> > 
> > It's also not clear to me how your suggestion would work. Add a command
> > line argument to pin all the VCPUs to the specified cpumask?
> > 
> > > 
> > > With your implementation, the whole of kvmtool gets stuck to a given
> > > CPU type, which can be problematic.
> > 
> > Do you have a specific use case in mind? Or is it more like a general
> > concern regarding, for example, the virtio-blk-io or virtio-net-* threads
> > competing with the VCPU threads if the VM is doing lots of I/O?
> 
> Exactly that. The real requirement is that the vcpu thread affinities
> are that of the PMU, but not that of any other thread. Maybe that's
> just another parameter, independent of the PMU setup. Something like:
> 
> 	lkvm run ... --vcpu-affinity $(< /sys/devices/armv8_pmuv3_0/cpus)
> 
> for example.

That should be easy (famous last words?) to implement as a separate patch
on top of this series, I'll give that a go in the next iteration.

Thanks for the suggestion.

Alex
_______________________________________________
kvmarm mailing list
kvmarm@lists.cs.columbia.edu
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm