Re: [PATCH net-next v2 1/8] udp: fixup csum for GSO receive slow path

[Paolo Abeni <pabeni@redhat.com>]

On Mon, 2021-03-29 at 08:28 -0400, Willem de Bruijn wrote:
> On Mon, Mar 29, 2021 at 7:26 AM Paolo Abeni <pabeni@redhat.com> wrote:
> > On Fri, 2021-03-26 at 14:30 -0400, Willem de Bruijn wrote:
> > > On Thu, Mar 25, 2021 at 1:24 PM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > When UDP packets generated locally by a socket with UDP_SEGMENT
> > > > traverse the following path:
> > > > 
> > > > UDP tunnel(xmit) -> veth (segmentation) -> veth (gro) ->
> > > >         UDP tunnel (rx) -> UDP socket (no UDP_GRO)
> > > > 
> > > > they are segmented as part of the rx socket receive operation, and
> > > > present a CHECKSUM_NONE after segmentation.
> > > 
> > > would be good to capture how this happens, as it was not immediately obvious.
> > 
> > The CHECKSUM_PARTIAL is propagated up to the UDP tunnel processing,
> > where we have:
> > 
> >         __iptunnel_pull_header() -> skb_pull_rcsum() ->
> > skb_postpull_rcsum() -> __skb_postpull_rcsum() and the latter do the
> > conversion.
> 
> Please capture this in the commit message.

I will do.

> > > > Additionally the segmented packets UDP CB still refers to the original
> > > > GSO packet len. Overall that causes unexpected/wrong csum validation
> > > > errors later in the UDP receive path.
> > > > 
> > > > We could possibly address the issue with some additional checks and
> > > > csum mangling in the UDP tunnel code. Since the issue affects only
> > > > this UDP receive slow path, let's set a suitable csum status there.
> > > > 
> > > > v1 -> v2:
> > > >  - restrict the csum update to the packets strictly needing them
> > > >  - hopefully clarify the commit message and code comments
> > > > 
> > > > Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> > > > +       if (skb->ip_summed == CHECKSUM_NONE && !skb->csum_valid)
> > > > +               skb->csum_valid = 1;
> > > 
> > > Not entirely obvious is that UDP packets arriving on a device with rx
> > > checksum offload off, i.e., with CHECKSUM_NONE, are not matched by
> > > this test.
> > > 
> > > I assume that such packets are not coalesced by the GRO layer in the
> > > first place. But I can't immediately spot the reason for it..
> > 
> > Packets with CHECKSUM_NONE are actually aggregated by the GRO engine.
> > 
> > Their checksum is validated by:
> > 
> > udp4_gro_receive -> skb_gro_checksum_validate_zero_check()
> >         -> __skb_gro_checksum_validate -> __skb_gro_checksum_validate_complete()
> > 
> > and skb->ip_summed is changed to CHECKSUM_UNNECESSARY by:
> > 
> > __skb_gro_checksum_validate -> skb_gro_incr_csum_unnecessary
> >         -> __skb_incr_checksum_unnecessary()
> > 
> > and finally to CHECKSUM_PARTIAL by:
> > 
> > udp4_gro_complete() -> udp_gro_complete() -> udp_gro_complete_segment()
> > 
> > Do you prefer I resubmit with some more comments, either in the commit
> > message or in the code?
> 
> That breaks the checksum-and-copy optimization when delivering to
> local sockets. I wonder if that is a regression.

The conversion to CHECKSUM_UNNECESSARY happens since
commit 573e8fca255a27e3573b51f9b183d62641c47a3d.

Even the conversion to CHECKSUM_PARTIAL happens independently from this
series, since commit 6f1c0ea133a6e4a193a7b285efe209664caeea43.

I don't see a regression here ?!?

Thanks!

Paolo