Re: [PATCH net-next V2 3/3] tap: XDP support

From: Michael S. Tsirkin
Date: Tue Aug 15 2017 - 23:45:42 EST


On Tue, Aug 15, 2017 at 01:02:05PM +0800, Jason Wang wrote:
>
>
> On 2017å08æ15æ 00:01, Michael S. Tsirkin wrote:
> > On Sat, Aug 12, 2017 at 10:48:49AM +0800, Jason Wang wrote:
> > >
> > > On 2017å08æ12æ 07:12, Jakub Kicinski wrote:
> > > > On Fri, 11 Aug 2017 19:41:18 +0800, Jason Wang wrote:
> > > > > This patch tries to implement XDP for tun. The implementation was
> > > > > split into two parts:
> > > > >
> > > > > - fast path: small and no gso packet. We try to do XDP at page level
> > > > > before build_skb(). For XDP_TX, since creating/destroying queues
> > > > > were completely under control of userspace, it was implemented
> > > > > through generic XDP helper after skb has been built. This could be
> > > > > optimized in the future.
> > > > > - slow path: big or gso packet. We try to do it after skb was created
> > > > > through generic XDP helpers.
> > > > >
> > > > > Test were done through pktgen with small packets.
> > > > >
> > > > > xdp1 test shows ~41.1% improvement:
> > > > >
> > > > > Before: ~1.7Mpps
> > > > > After: ~2.3Mpps
> > > > >
> > > > > xdp_redirect to ixgbe shows ~60% improvement:
> > > > >
> > > > > Before: ~0.8Mpps
> > > > > After: ~1.38Mpps
> > > > >
> > > > > Suggested-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
> > > > > Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
> > > > Looks OK to me now :)
> > > >
> > > > Out of curiosity, you say the build_skb() is for "small packets", and it
> > > > seems you are always reserving the 256B regardless of XDP being
> > > > installed. Does this have no performance impact on non-XDP case?
> > > Have a test, only less than 1% were noticed which I think could be ignored.
> > >
> > > Thanks
> > What did you test btw?
>
> Pktgen
>
> > The biggest issue would be with something like
> > UDP with short packets.
> >
>
> Note that we do this only when sndbuf is INT_MAX. So this is probably not an
> issue.

I'd expect to see the issue for guest to host when the packets are
queued at a UDP socket in host.

> The only thing matter is more stress to page allocator, but according
> to the result of pktgen it was very small that could be ignored.
>
> Thanks

Besides guest to host, for bridging in host bigger truesize might affect
byte queue counts as well.

--
MST