Re: [Qemu-devel] [Xen-devel] [RFC PATCH 0/4] Qemu: Add Xen vIOMMU support

[Paolo Bonzini <pbonzini@redhat.com>]

On 20/03/2017 15:17, Roger Pau Monné wrote:
>>> Hi Paolo:
>>> Thanks for review. For Xen side, we won't reuse Intel IOMMU device model
>>> in Qemu and create counterpart in Xen hypervisor. The reasons are
>>>  1) Avoid round trips between Qemu and Xen hypervisor
>>>  2) Ease of integration with the rest part of the hypervisor(vIOAPIC,
>>> vMSI and so on).
>> Fair enough, though I'd be worried about increasing the attack surface
>> of the hypervisor.  For KVM, for example, IOMMU emulation requires using
>> the "split irqchip" feature to move the PIC and IOAPIC out of the kernel
>> and back to QEMU.
> Yes, that's right, we are increasing the surface of attack. But Xen also needs
> it in order to support APIC IDs > 255 on PVH guests (that have a local APIC but
> no QEMU).

Not necessarily, you only need x2APIC support for that in the local APIC
emulation.  The MSI hypercalls (similar to KVM's MSI ioctls) can be
extended to accept x2APIC VCPU ids.

Paolo