Re: [PATCH 4/4] tun: indicate support for USO feature

From: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
To: Yuri Benditovich <yuri.benditovich@daynix.com>
Cc: Jason Wang <jasowang@redhat.com>, Yan Vugenfirer <yan@daynix.com>,
	davem <davem@davemloft.net>, Jakub Kicinski <kuba@kernel.org>,
	mst <mst@redhat.com>, netdev <netdev@vger.kernel.org>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	virtualization <virtualization@lists.linux-foundation.org>,
	Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Subject: Re: [PATCH 4/4] tun: indicate support for USO feature
Date: Thu, 13 May 2021 16:43:16 -0400	[thread overview]
Message-ID: <CA+FuTSfr4gLwx0PaRCB1K=TUE_yawpnWx05U9yO0eQ1B+Pa+bg@mail.gmail.com> (raw)
In-Reply-To: <CAOEp5Oe7FQQFbO7KDiyBPs1=ox+6rOimOwounTHBuVki2Y3DAg@mail.gmail.com>

> > > > > So the question is what to do now:
> > > > > A)
> > > > > Finalize patches for guest TX and respective QEMU patches
> > > > > Prepare RFC patches for guest RX, get ack on them
> > > > > Change the spec
> > > > > Finalize patches for guest RX according to the spec
> > > > >
> > > > > B)
> > > > > Reject the patches for guest TX
> > > > > Prepare RFC patches for everything, get ack on them
> > > > > Change the spec
> > > > > Finalize patches for everything according to the spec
> > > > >
> > > > > I'm for A) of course :)
> > > >
> > > > I'm for B :)
> > > >
> > > > The reasons are:
> > > >
> > > > 1) keep the assumption of tun_set_offload() to simply the logic and
> > > > compatibility
> > > > 2) it's hard or tricky to touch guest TX path only (e.g the
> > > > virtio_net_hdr_from_skb() is called in both RX and TX)
> > >
> > > I suspect there is _some_ misunderstanding here.
> > > I did not touch virtio_net_hdr_from_skb at all.
> > >
> >
> > Typo, actually I meant virtio_net_hdr_to_skb().
> OK.
> 2) tun_get_user() which is guest TX - this is covered
> 3) tap_get_user() which is guest TX - this is covered
> 4) {t}packet_send() which is userspace TX - this is OK, the userspace
> does not have this feature, it will never use USO

What do you mean exactly? I can certainly imagine packet socket users
that could benefit from using udp gso.

When adding support for a new GSO type in virtio_net_hdr, it ideally
is supported by all users of that interface. Alternatively, if some
users do not support the flag, a call that sets the flag has to
(continue to) fail hard, so that we can enable it at a later time.