Re: Xenner design and kvm msr handling

[Gerd Hoffmann <kraxel@redhat.com>]

On 04/21/09 12:14, Avi Kivity wrote:
> Gerd Hoffmann wrote:
>> xenner & pv-on-hvm
>> ==================
>>
>> Once we have all this in qemu it is just a small step to also support
>> xenish pv-on-hvm drivers in qemu using the xenner emulation bits.
>> Hypercalls are handled by a small pic binary loaded into the hypercall
>> pages. Loading of the binary is triggered by the msr writes discussed.
>> Size of the binary is only two pages: one hypercall entry points, one
>> code. Communication path is the very same ioport interface also used
>> by emu, i.e. it does *not* use vmcall and thus no opcode changes are
>> needed on migration.
>>
> This gives a good case for exporting MSRs to userspace.
>
> Can you explain the protocol used by this MSR? How does the guest know
> how many pages to load? How does the kernel know which type of page to
> put where?

Sure.

   (1) cpuid 0x40000000, check vmm signature
   (2) cpuid 0x40000002 -> returns # of pages (eax) and msr (ebx)
   (3) allocate pages (normal ram)
   (4) foreach page (wrmsr "guest physical address | pageno")

Xen uses msr 0x40000000.  Due to the msr being queried via cpuid it 
should be possible to use another one.  Modulo guest bugs of course ...

> Note that it would still be interesting to have the guest call the
> kernel, so it can kick the host kernel Xen netback driver directly
> instead of going through qemu (and the userspace netback + tap).

With the current codebase netback isn't involved at all.  Backend lives 
in qemu like virtio-net.  It is a very simple one (no GSO support, ...), 
there are tons of opportunities for optimizations.  Don't feel like 
tackeling that right now though.  My patch queue is already deep enougth 
as-is.  Also as time passes the qemu network layer tweaks for virtio-net 
should make that job easier ;)

cheers,
   Gerd