Re: [PATCH net] veth: Avoid drop packets when xdp_redirect performs

Heng Qi [off-list ref] writes:

在 2022/10/21 下午2:31, Heng Qi 写道:

quoted

在 2022/10/21 上午12:34, Toke Høiland-Jørgensen 写道:

quoted

Heng Qi [off-list ref] writes:

quoted

maybe we should consider a simpler method: when loading xdp in veth,
we can automatically enable the napi ring of peer veth, which seems to
have no performance impact and functional impact on the veth pair, and
no longer requires users to do more things for peer veth (after all,
they may be unaware of more requirements for peer veth). Do you think
this is feasible?

It could be, perhaps? One issue is what to do once the XDP program is
then unloaded? We should probably disable NAPI on the peer in this case,
but then we'd need to track whether it was enabled by loading an XDP
program; we don't want to disable GRO/NAPI if the user requested it
explicitly. This kind of state tracking gets icky fast, so I guess it'll
depend on the patch...

Regarding tracking whether we disable the napi of peer veth when 
unloading
the veth's xdp program, this can actually be handled cleanly.

We need to note that when peer veth enable GRO, the peer veth device will
update the `dev->wanted_features` with NETIF_F_GRO of peer veth (refer to
__netdev_update_features() and veth_set_features() ).

When veth loads the xdp program and the napi of peer veth is not ready
(that is, peer veth does not load the xdp program or has no enable gro),
at this time, we can choose `ethtool -K veth0 gro on` to enable the 
napi of
peer veth, this command also makes the peer veth device update their
wanted_features, or choose we automatically enable napi for peer veth.

If we want to unload the xdp program for veth, peer veth cannot directly
disable its napi, because we need to judge whether peer veth is 
gro_requested
( ref to veth_gro_requested() ) or has its priv->_xdp_prog, if so, just
clean veth's xdp environment and disable the napi of veth instead of
directly disable the napi of peer veth, because of the existence of the
gro_requested and the xdp program loading on peer veth.

But, if peer veth does not have gro_requested or xdp_program loading 
on itself,
then we can directly disable the napi of peer veth.

Hi, Toke. Do you think the above solution is effective for the problem 
of veth
xdp_rediect dropping packets? Or is there more to add? If the above solution
seems to be no problem for the time being, I'm ready to start with this idea
and try to make the corresponding patch.

I think it might? Please write a patch and we can have a look :)

-Toke