> -----Original Message-----
> From: Marc Zyngier <maz@xxxxxxxxxx>
> Sent: Friday, April 3, 2020 1:32 PM
> To: George Cherian <gcherian@xxxxxxxxxxx>
> Cc: Dave.Martin@xxxxxxx; alexandru.elisei@xxxxxxx;
> andre.przywara@xxxxxxx; christoffer.dall@xxxxxxx;
> james.morse@xxxxxxx; jintack@xxxxxxxxxxxxxxx;
> julien.thierry.kdev@xxxxxxxxx; kvm@xxxxxxxxxxxxxxx;
> kvmarm@xxxxxxxxxxxxxxxxxxxxx; linux-arm-kernel@xxxxxxxxxxxxxxxxxxx;
> suzuki.poulose@xxxxxxx; Anil Kumar Reddy H <areddy3@xxxxxxxxxxx>;
> Ganapatrao Kulkarni <gkulkarni@xxxxxxxxxxx>
> Subject: Re: [PATCH v2 00/94] KVM: arm64: ARMv8.3/8.4 Nested
> Virtualization support
>
>
> ----------------------------------------------------------------------
> Hi George,
>
> On 2020-04-03 08:27, George Cherian wrote:
> > Hi Marc,
> >
> > On 2/11/20 9:48 AM, Marc Zyngier wrote:
> >> This is a major rework of the NV series that I posted over 6 months
> >> ago[1], and a lot has changed since then:
> >>
> >> - Early ARMv8.4-NV support
> >> - ARMv8.4-TTL support in host and guest
> >> - ARMv8.5-GTG support in host and guest
> >> - Lots of comments addressed after the review
> >> - Rebased on v5.6-rc1
> >> - Way too many patches
> >>
> >> In my defence, the whole of the NV code is still smaller that the
> >> 32bit KVM/arm code I'm about to remove, so I feel less bad inflicting
> >> this on everyone! ;-)
> >>
> >> >From a functionality perspective, you can expect a L2 guest to work,
> >> but don't even think of L3, as we only partially emulate the
> >> ARMv8.{3,4}-NV extensions themselves. Same thing for vgic, debug,
> >> PMU, as well as anything that would require a Stage-1 PTW. What we
> >> want to achieve is that with NV disabled, there is no performance
> >> overhead and no regression.
> >>
> >> The series is roughly divided in 5 parts: exception handling, memory
> >> virtualization, interrupts and timers for ARMv8.3, followed by the
> >> ARMv8.4 support. There are of course some dependencies, but you'll
> >> hopefully get the gist of it.
> >>
> >> For the most courageous of you, I've put out a branch[2]. Of course,
> >> you'll need some userspace. Andre maintains a hacked version of
> >> kvmtool[3] that takes a --nested option, allowing the guest to be
> >> started at EL2. You can run the whole stack in the Foundation model.
> >> Don't be in a hurry ;-).
> >>
> > The full series was tested on both Foundation model as well as Marvell
> > ThunderX3
> > Emulation Platform.
> > Basic boot testing done for Guest Hypervisor and Guest Guest.
> >
> > Tested-by: George Cherian <george.cherian@xxxxxxxxxxx>
>
> Thanks for having given this a go.
>
> However, without more details, it is pretty hard to find out what you have
> tested.
> What sort of guest have you booted, with what configuration, what
> workloads did you run in the L2 guests and what are the architectural
> features that TX3 implements?
>
We have tried the following configurations and tests (GH - Guest Hypervisor GG- Guest Guest).
1 - configuration: Host:8cpus/4GB Mem, GH:4vcpus/3GB, GG: 2vcpus/2GB
Ran hackbench and Large Malloc tests (1GB allocations) across HOST,GH and GG.
2 - configuration: Host:8cpus/4GB Mem, GH:1vcpu/3GB, GG: 1vcpu/2GB
Ran hackbench and Large Malloc tests across HOST,GH and GG. Host:
We used QEMU for all these testing.
TX3 implements v8.4 Enhanced Nested Virtualization Support.
> The last point is specially important, as the NV architecture spans two major
> revisions of the architecture and affects tons of other extensions that are
> themselves optional. Without any detail on that front, I have no idea what
> the coverage of your testing is.
>
> Thanks,
>
> M.
> --
> Jazz is not dead. It just smells funny...
