Re: [PATCH v8 20/20] KVM: ARM64: Add a new kvm ARM PMU device

[Peter Maydell <peter.maydell@linaro.org>]

On 8 January 2016 at 12:15, Shannon Zhao <shannon.zhao@linaro.org> wrote:
> Firstly, userspace will call kvm device ioctl to create the PMU device, if
> it fails it will return an error.
> Secondly, userspace will call SET_DEVICE_ATTR ioctl to configure the PMU irq
> number. If the irq is invalid, this will return an error and the userspace
> will stop if it takes care the error. If the irq is valid, the PMU is well
> initialized.
>
> If userspace doesn't call SET_ATTR ioctl to configure the PMU irq, the PMU
> is not well initialized and guest can't use the PMU.

Yes, this is my point. If you require userspace to do:
 * create PMU device
 * configure PMU device via SET_DEVICE_ATTR
 * complete init of PMU device via a KVM_DEV_ARM_PMU_CTRL_INIT attr

then any bugs on the userspace side where it isn't configuring
correctly can be easily caught, and will turn into a "VM fails
to start with helpful error message", rather than "VM starts but
doesn't work in an obscure way".

This is particularly important for devices which have userspace
visible registers, because userspace will also want to be able
to read and write state of those registers before first vcpu run.
If there is no "init is complete" ioctl then the kernel code ends
up having to guess when init has completed for various entry points
where the guest does something else (register access, vcpu start, etc).

It might happen to not be so important for the PMU, but I think it
would be very useful if we have a pattern where every device that
is created via this API works in the same way (and it means that if
we end up extending the PMU in future we will already have the API
in place and don't need to worry about legacy userspace).

Speaking of userspace access to registers, what is the intended API
for migrating the PMU register state ?

thanks
-- PMM