Re: More than 255 vcpus Windows VM setup without viommu ?

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 2045 bytes --]

On Thu, 2024-07-11 at 08:26 +0100, David Woodhouse wrote:
> 
> I used identical command lines on both, and on each host I got the same
> result with all of '-cpu host', '-cpu EPYC' and -cpu Skylake-Server'.
> It's the *host* that makes the difference, not the CPUID presented to
> the guest.

Actually... it turns out QEMU isn't really advertising the CPUID we ask
it to. Leaf zero still does say 'AuthenticAMD' vs. GenuineIntel'
according to the *host* it's running on, regardless of the -cpu option.

And it is indeed *just* that which seems to trigger the Windows bug,
setting IRQ2 to point somewhere bogus:

vtd_ir_remap_msi_req addr 0xfee00004 data 0x0 sid 0xffff do_fault 0
vtd_ir_remap_msi (addr 0xfee00004, data 0x0) -> (addr 0xfee00004, data 0x0)
kvm_irqchip_update_msi_route Updating MSI route virq=2

While in the happy case it does use a remappable format MSI message.
Not direct to vector 209 on CPU0 as I said before; I think that's IRTE
entry #209 which maps to vector 209 on CPU1.

vtd_ir_remap_msi_req addr 0xfee00010 data 0xd1 sid 0xffff do_fault 0
vtd_ir_irte_get index 0 low 0x0 high 0x100d10005
vtd_ir_remap index 0 trigger 0 vector 209 deliver 0 dest 0x1 mode 1
vtd_ir_remap_type IOAPIC
vtd_ir_remap_msi (addr 0xfee00010, data 0xd1) -> (addr 0xfee01004, data 0x40d1)

So it looks like Windows doesn't actually cope with Intel IRQ remapping
when it sees and AMD CPU, which is suboptimal.

So to support >255 vCPUs on AMD without having to also do *DMA*
translation, either we need to come up with a trick like the "no
supported address widths" we use for dma-translation=off on Intel, or
we see if we can persuade Windows to use the 15-bit MSI support.


Looking at the Linux guest support, it seems to look just at the HyperV
CPUID leaves 0x40000081 and 0x40000082. QEMU knows of those only for
SYNDBG; Sandesh do you want to try setting the
HYPERV_VS_PROPERTIES_EAX_EXTENDED_IOAPIC_RTE bit that Linux looks for,
and see how that affects Windows guests (with no emulated IOMMU)?


[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 5965 bytes --]