Re: [Qemu-devel] [PATCH v3 2/3] arm64: kvm: inject SError with virtual syndrome

[James Morse]

Hi gengdongjiu,

On 05/05/17 14:19, gengdongjiu wrote:
> On 2017/5/2 23:37, James Morse wrote:
> > ... I think you expect an SError to arrive at EL2 and have its ESR recorded 
> > in
> > vcpu->arch.fault.vsesr_el2. Some time later KVM decides to inject an SError 
> > into
> > the guest, and this ESR is reused...
> > 
> > We shouldn't do this. Qemu/kvmtool may want to inject a virtual-SError that
> > never started as a physical-SError. Qemu/kvmtool may choose to notify the 
> > guest
> > of RAS events via another mechanism, or not at all.
> > 
> > KVM should not give the guest an ESR value of its choice. For SError the ESR
> > describes whether the error is corrected, correctable or fatal. Qemu/kvmtool
> > must choose this.

> Below is my previous solution:
> For the SError, CPU will firstly trap to EL3 firmware and records the 
> syndrome to ESR_EL3.
> Before jumping to El2 hypervisors, it will copy the esr_el3 to esr_el2.

(Copying the ESR value won't always be the right thing to do.)


> so in order to pass this syndrome to vsesr_el2, using the esr_el2 value to 
> assign it.

> If Qemu/kvmtool chooses the ESR value and ESR only describes whether the 
> error is corrected/correctable/fatal,
> whether the information is not enough for the guest?

So the API should specify which of these three severities to use? I think this
is too specific. The API should be useful for anything the VSE/VSESR hardware
can do.

VSESR_EL2 is described in the RAS spec: 4.4.12 [0], its a 64 bit register. I
think we should let Qemu/kvmtool specify any 64bit value here, but KVM should
reject values that try to set bits described as RES0.

This would let Qemu/kvmtool specify any SError ISS, either setting ESR_ELx.IDS
and some virtual-machine specific value, or encoding any severity in AET and
choosing the DFSC/EA bits appropriately.


>> > I think we need an API that allows Qemu/kvmtool to inject SError into a 
>> > guest,
>> > but that isn't quite what you have here.

> KVM provides APIs to inject the SError, Qemu/kvmtool call the API though 
> IOCTL, may be OK?

(just the one API call), yes.


Thanks,

James

[0]
https://static.docs.arm.com/ddi0587/a/RAS%20Extension-release%20candidate_march_29.pdf