On Wed, 2019-06-05 at 18:42 -0700, Cong Wang wrote:

On Tue, Jun 4, 2019 at 11:19 AM Eli Britstein [off-list ref] wrote:

hello Cong and Eli,

and thanks for all the thoughts.

quoted

On 6/4/2019 8:55 PM, Cong Wang wrote:

quoted

On Sat, Jun 1, 2019 at 9:22 PM Eli Britstein [off-list ref] wrote:

quoted

I think that's because QinQ, or VLAN is not an encapsulation. There is
no outer/inner packets, and if you want to mangle fields in the packet
you can do it and the result is well-defined.

Sort of, perhaps VLAN tags are too short to be called as an
encapsulation, my point is that it still needs some endpoints to push
or pop the tags, in a similar way we do encap/decap.

quoted

BTW, the motivation for my fix was a use case were 2 VGT VMs
communicating by OVS failed. Since OVS sees the same VLAN tag, it
doesn't add explicit VLAN pop/push actions (i.e pop, mangle, push). If
you force explicit pop/mangle/push you will break such applications.

 From what you said, it seems act_csum is in the middle of packet
receive/transmit path. So, which is the one pops the VLAN tags in
this scenario? If the VM's are the endpoints, why not use act_csum
there?

In a switchdev mode, we can passthru the VFs to VMs, and have their
representors in the host, enabling us to manipulate the HW eswitch
without knowledge of the VMs.

To simplify it, consider the following setup:

v1a <-> v1b and v2a <-> v2b are veth pairs.

Now, we configure v1a.20 and v2a.20 as VLAN devices over v1a/v2a
respectively (and put the "a" devs in separate namespaces).

The TC rules are on the "b" devs, for example:

tc filter add dev v1b ... action pedit ... action csum ... action
redirect dev v2b

Now, ping from v1a.20 to v1b.20. The namespaces transmit/receive tagged
packets, and are not aware of the packet manipulation (and the required
act_csum).

This is what I said, v1b is not the endpoint which pops the vlan tag,
v1b.20 is. So, why not simply move at least the csum action to
v1b.20? With that, you can still filter and redirect packets on v1b,
you still even modify it too, just defer the checksum fixup to the
endpoint.

And to be fair, if this case is a valid concern, so is VXLAN case,
just replace v1a.20 and v2a.20 with VXLAN tunnels. If you modify
the inner header, you have to fixup the checksum in the outer
UDP header.

at least with single "accelerated" vlan tags, I think that users of
tc_skb_protocol() that explicitly access the IP header should be fixed the
same way that Eli did to 'csum'.

(note that 'pedit' is *not* among them, and that's why Eli only needed to
fix 'csum' in his setup :) )
 
Now I'm no more sure about what to do with the QinQ case, whether we
should:

a) fix missing skb_may_pull() in csum, and fix (a couple of) other actions
in the same way, 

or

b) rework act_csum when it wants to read the IP header, in a way that only
ip header is mangled only when there are only zero or a single
"accelerated" tag, and do the same for (a couple of) other actions.

The pop/push approach built in the action ( option a) ) fixes my
environment - but like Cong says, it only would work with VLANs, and not
with other encapsulations. Probably it really makes sense to see if it's
possible to extend act_vlan in a way that it reconstructs the packet after
an iteration of 'vlan pop'. This might not be easy, because even the above
command assumes that inner and outer tag have the same ethertype (which is
not generally true for QinQ).

But until then, we should assume that read/writes of the IP header are not
feasible for QinQ traffic, just like it's not possible for tunneled
packets, and adjust configuration accordingly.

WDYT?

-- 
davide