Re: [Qemu-devel] [PATCH RFC 0/5] Xen: introduce Xen PV target

[Paolo Bonzini]

Il 24/01/2014 15:23, Wei Liu ha scritto:
> On Thu, Jan 23, 2014 at 10:30:16PM +0000, Peter Maydell wrote:
> > On 23 January 2014 22:16, Wei Liu <address@hidden> wrote:
> > > As promised I hacked a prototype based on Paolo's disable TCG series.
> > > However I coded some stubs for TCG anyway. So this series in principle
> > > should work with / without Paolo's series.
> >
> > I'm afraid I still think this is a terrible idea. "Xen" isn't a CPU, and
>
> Thanks for being blunt. ;-)
>
> > "the binary is smaller" isn't IMHO sufficient justification for breaking
> > QEMU's basic structure of "target-* define target CPUs and we have
> > a lot of compile time constants which are specific to a CPU which
> > get defined there". How would you support a bigendian Xen CPU,
> > just to pick one example of where this falls down?
>
> I think about this deeper. From Xen's (and I speculate this applies to
> other hardware assisted virtulization solution as well) PoV only the
> native endianess is supported, does it make sense to have a
> target-native thing?

I think this is wrong, for a few reasons:

(1) xenpv is not hardware assisted virtualization

(2) supporting only native endianness leads to complications when 
systems are bi-endian, as is the case for PPC.  For example, virtio 1.0 
will always be little endian.

(3) there's a precedent for supporting different guests between the 
guest and the host in blkback, you can do the same for endianness.

Paolo