Re: [PATCH] kvm tools: virtio-net mergable rx buffers

[Rusty Russell <rusty@rustcorp.com.au>]

"Michael S. Tsirkin" <mst@redhat.com> writes:
> On Wed, Apr 24, 2013 at 09:51:57AM +0300, Pekka Enberg wrote:
>> Hi,
>> 
>> On 04/23/2013 12:35 PM, Eric Northup wrote:
>> >> Do you care about guests with drivers that don't negotiate
>> >> VIRTIO_NET_F_MRG_RXBUF?
>> 
>> On Wed, Apr 24, 2013 at 5:51 AM, Sasha Levin <sasha.levin@oracle.com> wrote:
>> > We usually try to keep backward compatibility, but in this case
>> > mergable RX buffers are about 5 years old now, so it's safe to
>> > assume they'll be running in any guest.
>> >
>> > Unless there is a specific reason to allow working without them
>> > I'd rather keep the code simple in this case.
>> 
>> Are there such guests around? What's the failure scenario for them
>> after this patch?
>> 
>>                         Pekka
>
> Warning: have not looked at the patch, just a general comment.
>
> I think it's reasonable to assume embedded guests such as PXE won't
> negotiate any features.  And, running old guests is one of the reasons
> people use virtualization at all. So 5 years is not a lot.
>
> In any case, stick to the device spec please, if you want it changed
> please send a spec patch, don't deviate from it randomly.

Supporting old guests is an quality of implementation issue.  It's like
any ABI: if noone will notice, you can remove stuff.

But the case of "I can receive GSO packets but I don't support mergeable
buffers" is a trivial one: you can "support" it by pretending the guest
can't handle GSO :)

If you want to support non-Linux guests (eg. bootloaders), you probably
want to keep support for very dumb drivers with no mergable rxbufs
though.

Cheers,
Rusty.