On Fri, Nov 04, 2016 at 04:05:51PM -0700, John Fastabend wrote: > On 16-11-03 05:34 PM, Michael S. Tsirkin wrote: > > On Thu, Nov 03, 2016 at 04:29:22PM -0700, John Fastabend wrote: > >> [...] > >> > >>>>> - when XDP is attached disable all LRO using > >>>>> VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET > >>>>> (not used by driver so far, designed to allow dynamic LRO control with > >>>>> ethtool) > >>>> > >>>> I see there is a UAPI bit for this but I guess we also need to add > >>>> support to vhost as well? Seems otherwise we may just drop a bunch > >>>> of packets on the floor out of handle_rx() when recvmsg returns larger > >>>> than a page size. Or did I read this wrong... > >>> > >>> It's already supported host side. However you might > >>> get some packets that were in flight when you attached. > >>> > >> > >> Really I must have missed it I don't see any *GUEST_FEATURES* flag in > >> ./drivers/vhost/? > > > > It's all done by QEMU catching these commands and calling > > ioctls on the tun/macvtap/packet socket. > > > > Well at least for the tap vhost backend in linux that I found here, > > ./qemu/net/tap-linux.c > > there is no LRO feature flag but that is OK I can get it working next > week looks fairly straight forward. > > [...]
This is because tun/tap is the reverse of virtio. LRO in virtio maps to TSO in tun. The relevant function is tap_fd_set_offload in QEMU. > >> And if I try to merge the last email I sent out here. In mergeable and > >> big_packets modes if LRO is off and MTU < PAGE_SIZE it seems we should > >> always get physically contiguous data on a single page correct? > > > > Unfortunately not in the mergeable buffer case according to spec, even > > though > > linux hosts will do that, so it's fine to optimize for that > > but need to somehow work in other cases e.g. by doing a data copy. > > > > ah OK this makes sense I was looking at vhost implementation in Linux. > > > > >> It > >> may be at some offset in a page however. But the offset should not > >> matter to XDP. If I read this right we wouldn't need to add a new > >> XDP mode and could just use the existing merge or big modes. This would > >> seem cleaner to me than adding a new mode and requiring a qemu option. > >> > >> Thanks for all the pointers by the way its very helpful. > > > > So for mergeable we spend cycles trying to make buffers as small > > as possible and I have a patch to avoid copies for that too, > > I'll post it next week hopefully. > > > > Good to know. I'll get the XDP stuff wrapped up next week or see > if Srijeet wants to do it. > > Thanks, > John
