QEMU KVM acceleration on aarch64 is disabled by default even though it seems to be working

[malte-v]

I was wondering why everything was so slow in QEMU on my M2 Macbook Air running the asahi branch Linux kernel, then I found this in run.py:

    # FIXME: Booting with KVM on aarch64 is broken, so disable it for now.
    # FIXME: QEMU on Windows on ARM does not support WHPX yet
    if config.virtualization_support and config.qemu_kind != QEMUKind.MacOS and config.architecture == Arch.Aarch64:
        config.virtualization_support = False

Commenting out these lines, I found that the system boots up just fine and is obviously much snappier than under software emulation. Just posting this in case the issue has been resolved. In case it's still broken on other systems, I think it would be good to at least print a message that KVM acceleration is disabled on purpose.

[spholz]

Yes, booting with KVM is still broken on some arm systems. I wanted to fix the bug causing this, but it's a bit annoying to do so. QEMU always maps all memory as cacheable and therefore adds a dma-coherent property in most devicetree nodes to tell us that we should also use cacheable mappings. However, we currently don't do so and switching to cacheable mappings requires carefully tweaking each DMA-capable driver to a new DMA abstraction that I have local patches for. Mapping physical memory with incompatible attributes (so e.g. cacheable vs. non-cacheable) is not allowed by arm.

[mediouni-m]

On KVM, this is not a problem for Armv8.2 hardware onwards as KVM enables the S2FWB bit if that's present. However, Hyper-V doesn't enforce S2FWB so WHPX support would have to be gated on that being addressed.