veth.sh from ubuntu_kselftests_net failed on J-5.15 (with xdp attached - gro flag)

[Po-Hsu Lin]

Issue found with Jammy 5.15.0-111.121 in sru-20240429

Reproduce rate is 100% across different arches on openstack cloud.

Test log:
ubuntu@kt-j-l-gen-5-15-bc2r4d20-u-kselftests-net-amd64:~/autotest/client/tmp/ubuntu_kselftests_net/src/linux/tools/testing/selftests/net$ sudo ./veth.sh
default - gro flag ok
        - peer gro flag ok
        - tso flag ok
        - peer tso flag ok
        - aggregation ok
        - aggregation with TSO off ok
with gro on - gro flag ok
        - peer gro flag ok
        - tso flag ok
        - peer tso flag ok
        - aggregation with TSO off ok
default channels ok
with gro enabled on link down - gro flag ok
        - peer gro flag ok
        - tso flag ok
        - peer tso flag ok
        - aggregation with TSO off ok
setting tx channels ok
bad setting: combined channels ok
setting invalid channels nr ok
bad setting: XDP with RX nr less than TX ok
bad setting: reducing RX nr below peer TX with XDP set ok
with xdp attached - gro flag fail - expected on found off
        - peer gro flag ok
        - tso flag ok
        - peer tso flag ok
        - aggregation fail - got 10 packets, expected 1
        - after dev off, flag fail - expected on found off
        - peer flag ok
        - after gro on xdp off, gro flag ok
        - peer gro flag ok
        - tso flag ok
        - peer tso flag ok
decreasing tx channels with device down ok
        - aggregation ok
increasing tx channels with device down ok
aggregation again with default and TSO off ok

This failure is different than our known issue of this test (LP: #1949569 with gro on/aggregation with TSO off) And we don't have this failure on openstack cloud in the previous cycles.

I have also verified the following combinations:
* 105 kernel + 106 source code - GOOD
* 106 kernel + 106 source code - GOOD
* 111 kernel + 106 source code - BAD
* 111 kernel + 111 source code - BAD
* 106 kernel + 111 source code - GOOD

This appears to be a possible regression in the kernel to me.

[Roxana Nicolescu (roxanan)]

This happens in noble too.
Commit "net: veth: do not manipulate GRO when using XDP" introduces the problem. Reverting it makes the test pass again.
I haven't seen anything upstream related to this, I'll first understand what it does and then talk to them to solve it.

[Roxana Nicolescu (roxanan)]

I think the test should be updated, they do not enable GRO when XDP is enabled by default, therefore it must be requested in the userspace I believe

[Roxana Nicolescu (roxanan)]

Adding
ip netns exec $NS_DST ethtool -K veth$DST gro on

Just before the test works. I'll send this to upstream tomorrow morning