Re: [Qemu-devel] KVM/QEMU on Raspberry Pi 3

[Christoffer Dall]

On Thu, Feb 2, 2017 at 3:59 PM, Marc Zyngier <address@hidden> wrote:
> [+Christoffer]
>
> Hi Pekka,
>
> On 02/02/17 14:44, Pekka Enberg wrote:
>> Hi,
>>
>> Has anyone been able to successfully run QEMU/KVM under Raspberry Pi 3?
>>
>> I have installed 64-bit Fedora 24 by Gerd Hoffmann on the hardware:
>>
>>    https://www.kraxel.org/blog/2016/04/fedora-on-raspberry-pi-updates/
>>
>> and built a VM image using virt-builder:
>>
>>    virt-builder --root-password password:root --arch aarch64 fedora-24
>>
>> I also built the latest UEFI for QEMU from sources:
>>
>>    https://wiki.linaro.org/LEG/UEFIforQEMU
>>
>> and updated to QEMU 2.8.0 from rawhide:
>>
>>    address@hidden ~]# qemu-system-aarch64 -version
>>    QEMU emulator version 2.8.0(qemu-2.8.0-1.fc26)
>>    Copyright (c) 2003-2016 Fabrice Bellard and the QEMU Project developers
>>
>> The VM image should be fine because I’m able to boot to it under CPU
>> emulation:
>>
>> qemu-system-aarch64 \
>>            -nographic \
>>            -M virt \
>>            -cpu cortex-a57 \
>>            -smp 1 \
>>            -m 512 \
>>            -bios QEMU_EFI.fd \
>>            -device virtio-blk-device,drive=image -drive
>> if=none,id=image,file=fedora-24.img \
>>            -netdev bridge,id=hn0,br=virbr0 -device
>> virtio-net-pci,netdev=hn0,romfile= \
>>            -device virtio-rng-pci
>>
>> However, when I enable KVM, keyboard stops working (interrupt delivery
>> issue?) and Fedora boot process hangs at random places before reaching
>> login:
>
> [snip]
>
> TL;DR: as it stands now, none of the RPi{2,3} can run KVM out of the
> box, as they lack a virtualization capable interrupt controller. This
> means that timer interrupts cannot be reported to the core, and things
> die a painful death.
>
> The longer story: we have a set of patches somewhere on the list that do
> enable the timer interrupts to be reported to userspace (QEMU), which
> can then inject them into its on GIC emulation and kick the vcpu. So
> far, work on this seems to have stopped (API issues? QEMU patches?)
>
> Christoffer was about to revive the kernel patches, but we need someone
> to pick up the QEMU part, and work with us and the QEMU people so that
> we agree once and for all on the ABI to signal PPIs to userspace.

Interesting timing.  I just revived the patches today, but I reworked
the ABI slightly and rebased the work on top of support for the
physical timer in the guest and ran into two issues: First, turning of
the in-kernel irqchip no longer works, and I also lost track of how
the patches should look, so indeed I need help from a QEMU person to
look at that.  Second, there is some sort of regression on the 32-bit
side using the physical timer patches.

Once I've sorted some of this out, I can send out the patch series.
At least I want to figure out the 32-bit breakage first, and then I
may simply send out the patches with a big fat UNTESTED warning in
hope that someone will work on the qemu side with me.

Thanks,
-Christoffer