Re: [PATCH net-next 0/6] net: reset MAC header consistently across L3 virtual devices

On Mon, Jun 28, 2021 at 12:46:24PM -0600, David Ahern wrote:

On 6/28/21 9:04 AM, Guillaume Nault wrote:

quoted

On Sun, Jun 27, 2021 at 09:56:53AM -0600, David Ahern wrote:

quoted

On 6/26/21 2:53 PM, Guillaume Nault wrote:

quoted

On Sat, Jun 26, 2021 at 11:50:19AM -0600, David Ahern wrote:

quoted

On 6/25/21 7:32 AM, Guillaume Nault wrote:

quoted

Some virtual L3 devices, like vxlan-gpe and gre (in collect_md mode),
reset the MAC header pointer after they parsed the outer headers. This
accurately reflects the fact that the decapsulated packet is pure L3
packet, as that makes the MAC header 0 bytes long (the MAC and network
header pointers are equal).

However, many L3 devices only adjust the network header after
decapsulation and leave the MAC header pointer to its original value.
This can confuse other parts of the networking stack, like TC, which
then considers the outer headers as one big MAC header.

This patch series makes the following L3 tunnels behave like VXLAN-GPE:
bareudp, ipip, sit, gre, ip6gre, ip6tnl, gtp.

The case of gre is a bit special. It already resets the MAC header
pointer in collect_md mode, so only the classical mode needs to be
adjusted. However, gre also has a special case that expects the MAC
header pointer to keep pointing to the outer header even after
decapsulation. Therefore, patch 4 keeps an exception for this case.

Ideally, we'd centralise the call to skb_reset_mac_header() in
ip_tunnel_rcv(), to avoid manual calls in ipip (patch 2),
sit (patch 3) and gre (patch 4). That's unfortunately not feasible
currently, because of the gre special case discussed above that
precludes us from resetting the MAC header unconditionally.

What about adding a flag to ip_tunnel indicating if it can be done (or
should not be done since doing it is the most common)?

That's feasible. I didn't do it here because I wanted to keep the
patch series focused on L3 tunnels. Modifying ip_tunnel_rcv()'s
prototype would also require updating erspan_rcv(), which isn't L3
(erspan carries Ethernet frames). I was feeling such consolidation
would be best done in a follow up patch series.

I was thinking a flag in 'struct ip_tunnel'. It's the private data for
those netdevices, so a per-instance setting. I haven't walked through
the details to know if it would work.

I didn't think about that. Good idea, that looks perfectly doable. But
I'd still prefer to centralise the skb_reset_mac_header() call in a
dedicated patch set. I we did it here, we'd have to not reset the mac
header by default, to guarantee that unrelated tunnels wouldn't be
affected.
However, I think that the default behaviour should be to reset the mac
header and that only special cases, like the one in ip_gre, should
explicitely turn that off. Therefore, we'd need a follow up patch
anyway, to change the way this "reset_mac" flag would be set.

IMHO, the current approach has the advantage of clearly separating the
new feature from the refactoring. But if you feel strongly about using
a flag in struct ip_tunnel, I can rework this series.

I am accustomed to doing refactoring first to add some new feature in
the simplest and clearest way.

I agree on the general statement. Anyway, I've just received the
patchwork-bot message saying the series has been applied. I'll follow
up with the struct ip_tunnel flag once net-next reopens (as I guess it's
about to close).