Re: [PATCHv7] add mergeable buffers support to vhost_net

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, May 03, 2010 at 08:39:08AM -0700, David Stevens wrote:
> "Michael S. Tsirkin" <mst@redhat.com> wrote on 05/03/2010 03:34:11 AM:
> 
> > On Wed, Apr 28, 2010 at 01:57:12PM -0700, David L Stevens wrote:
> > > This patch adds mergeable receive buffer support to vhost_net.
> > > 
> > > Signed-off-by: David L Stevens <dlstevens@us.ibm.com>
> > 
> > I've been doing some more testing before sending out a pull
> > request, and I see a drastic performance degradation in guest to host
> > traffic when this is applied but mergeable buffers are not in used
> > by userspace (existing qemu-kvm userspace).
> 
>         Actually, I wouldn't expect it to work at all;
> the qemu-kvm
> patch (particularly the feature bit setting bug fix) is required.

Which bugfix is that?

> Without it, I think the existing code will tell the guest to use
> mergeable buffers while turning it off in vhost.

It should not do that, specifically
vhost_net_get_features does:
	features &= ~(1 << VIRTIO_NET_F_MRG_RXBUF);
unconditionally. This was added with:
5751995a20e77cd9d61d00f7390401895fa172a6

I forced mergeable buffers off with -global virtio-net-pci.mrg_rxbuf=off
and it did not seem to help either.

> That was completely
> non-functional for me -- what version of qemu-kvm are you using?

992cc816c433332f2e93db033919a9ddbfcd1da4 or later should work well
AFAIK.

>         What I did to test w/o mergeable buffers is turn off the
> bit in VHOST_FEATURES.

It should be enough to force mergeable buffers to off by qemu command
line: -global virtio-net-pci.mrg_rxbuf=off 

> I'll recheck these, but qemu-kvm definitely
> must be updated; the original doesn't correctly handle feature bits.
> 
>                                                                 +-DLS

Hmm, I don't see the bug.

-- 
MST