[RFC PATCH v2 00/38] Nested Virtualization on KVM/ARM

[Christoffer Dall]

Hi Jintack,

On Tue, Jul 18, 2017 at 10:23:05PM -0400, Jintack Lim wrote:
> On Tue, Jul 18, 2017 at 12:58 PM, Jintack Lim <jintack.lim at linaro.org> wrote:
> > Nested virtualization is the ability to run a virtual machine inside another
> > virtual machine. In other words, it’s about running a hypervisor (the guest
> > hypervisor) on top of another hypervisor (the host hypervisor).
> >
> > Supporting nested virtualization on ARM means that the hypervisor provides not
> > only EL0/EL1 execution environment to VMs as it usually does but also the
> > virtualization extensions including EL2 execution environment. Once the host
> > hypervisor provides those execution environments to the VMs, then the guest
> > hypervisor can run its own VMs (nested VMs) naturally.
> >
> > This series supports nested virtualization on arm64. ARM recently announced an
> > extension (ARMv8.3) which has support for nested virtualization[1]. This patch
> > set is based on the ARMv8.3 specification and tested on the FastModel with
> > ARMv8.3 extension.
> >
> > The whole patch set to support nested virtualization is huge over 70
> > patches, so I categorized them into four parts: CPU, memory, VGIC, and timer
> > virtualization. This patch series is the first part.
> >
> > CPU virtualization patch series provides basic nested virtualization framework
> > and instruction emulations including v8.1 VHE feature and v8.3 nested
> > virtualization feature for VMs.
> >
> > This patch series again can be divided into four parts. Patch 1 to 5 introduces
> > nested virtualization by discovering hardware feature, adding a kernel
> > parameter and allowing the userspace to set the initial CPU mode to EL2.
> >
> > Patch 6 to 25 are to support the EL2 execution environment, the virtual EL2, to
> > a VM on v8.0 architecture. We de-privilege the guest hypervisor and emulate the
> > virtual EL2 mode in EL1 using the hardware features provided by ARMv8.3; The
> > host hypervisor manages virtual EL2 register state for the guest hypervisor
> > and shadow EL1 register state that reflects the virtual EL2 register state to
> > run the guest hypervisor in EL1.
> >
> > Patch 26 to 33 add support for the virtual EL2 with Virtualization Host
> > Extensions. These patches emulate newly defined registers and bits in v8.1 and
> > allow the virtual EL2 to access EL2 register states via EL1 register accesses
> > as in the real EL2.
> >
> > Patch 34 to 38 are to support for the virtual EL2 with nested virtualization.
> > These enable recursive nested virtualization.
> >
> > This patch set is tested on the FastModel with the v8.3 extension for arm64 and
> > a cubietruck for arm32. On the FastModel, the host and the guest kernels are
> > compiled with and without VHE, so there are four combinations. I was able to
> > boot SMP Linux in the nested VM on all four configurations and able to run
> > hackbench. I also checked that regular VMs could boot when the nested
> > virtualization kernel parameter was not set. On the cubietruck, I also verified
> > that regular VMs could boot as well.
> >
> > I'll share my experiment setup shortly.
> 
> I summarized my experiment setup here.
> 
> https://github.com/columbia/nesting-pub/wiki/Nested-virtualization-on-ARM-setup
> 

Thanks for sharing this.

> >
> > Even though this work has some limitations and TODOs, I'd appreciate early
> > feedback on this RFC. Specifically, I'm interested in:
> >
> > - Overall design to manage vcpu context for the virtual EL2
> > - Verifying correct EL2 register configurations such as HCR_EL2, CPTR_EL2
> >   (Patch 30 and 32)
> > - Patch organization and coding style
> 
> I also wonder if the hardware and/or KVM do not support nested
> virtualization but the userspace uses nested virtualization option,
> which one is better: giving an error or launching a regular VM
> silently.
> 

I think KVM should complain to userspace if userspace tries to set a
feature it does not support, and I think userspace should give as
meaningful an error message as possible to the user when that happens.

Thanks,
-Christoffer