On Thu, Dec 08, 2016 at 02:17:02PM -0500, David Miller wrote: > From: John Fastabend <john.fastab...@gmail.com> > Date: Wed, 07 Dec 2016 12:10:47 -0800 > > > This implements virtio_net for the mergeable buffers and big_packet > > modes. I tested this with vhost_net running on qemu and did not see > > any issues. For testing num_buf > 1 I added a hack to vhost driver > > to only but 100 bytes per buffer. > ... > > So where are we with this? > > I'm not too thrilled with the idea of making XDP_TX optional or > something like that. If someone enables XDP, there is a tradeoff. > > I also have reservations about the idea to make jumbo frames work > without giving XDP access to the whole packet. If it wants to push or > pop a header, it might need to know the whole packet length. How will > you pass that to the XDP program? > > Some kinds of encapsulation require trailers, thus preclusing access > to the entire packet precludes those kinds of transformations.
+1 > This is why we want simple, linear, buffer access for XDP. > > Even the most seemingly minor exception turns into a huge complicated > mess. +1 and from the other thread: > > Can't we disable XDP_TX somehow? Many people might only want RX drop, > > and extra queues are not always there. > > > > Alexei, Daniel, any thoughts on this? I don't like it. > I know we were trying to claim some base level of feature support for > all XDP drivers. I am sympathetic to this argument though for DDOS we > do not need XDP_TX support. And virtio can become queue constrained > in some cases. especially for ddos case doing lro/gro is not helpful. I frankly don't see a use case where you'd want to steer a packet all the way into VM just to drop them there? Without XDP_TX it's too crippled. adjust_head() won't be possible, packet mangling would have to be disabled and so on. If xdp program doesn't see raw packet it can only parse the headers of this jumbo meta-packet and drop it, but for virtio it's really too late. ddos protection needs to be done at the earliest hw nic receive. I think if driver claims xdp support it needs to support drop/pass/tx and adjust_head. For metadata passing up into stack from xdp we need adjust_head, for encap/decap we need it too. And lro is in the way of such transformations. We struggled a lot with cls_bpf due to all metadata inside skb that needs to be kept correct. Feeding non-raw packets into xdp is a rat hole.