EDIT: Removed references to segment routing as this can be reproduced without it. Added steps to reproduce.

reproduced in versions:

v19.04.2-rc0~20-g4df8ab00d

v19.08-rc0~711-g198cba8be~b3278

Two machines running VPP, connected by dpdk interface. Setup one tap interface on each VPP, test end-to-end communication. Enabling GSO on the taps causes the source IPv6 addresses of TCP and UDP packets to be randomly modified in-transit.

-node 1:
vppctl create tap host-ip6-addr 2001::1:1/120 host-if-name vpp1
vppctl set int ip address tap0 2001::1:2/120
vppctl set int state tap0 up
vppctl set int mtu 1450 tap0

vppctl set int ip address GigabitEthernet0/9/0 2001::3:1/120
vppctl set int state GigabitEthernet0/9/0 up
vppctl ip route add 2001::2:0/120 via 2001::3:2 GigabitEthernet0/9/0

ip link set dev vpp1 up
ip route add 2001::2:0/120 via 2001::1:2

-node 2:
vppctl create tap host-ip6-addr 2001::2:1/120 host-if-name vpp1
vppctl set int ip address tap0 2001::2:2/120
vppctl set int state tap0 up
vppctl set int mtu 1450 tap0

vppctl set int ip address GigabitEthernet0/9/0 2001::3:2/120
vppctl set int state GigabitEthernet0/9/0 up
vppctl ip route add 2001::1:0/120 via 2001::3:1 GigabitEthernet0/9/0

ip link set dev vpp1 up
ip route add 2001::1:0/120 via 2001::2:2

-ping end-to-end, wait for neighbor discovery
root@node2: ping 2001::1:1 -I vpp1 -6 -c 20 -W 3

-test TCP connectivity
root@node2: nc -6dl 12345
root@node1: echo "test" | nc -6q 1 2001::2:1 12345

-repeat test with GSO enabled
root@node2: vppctl set tap gso tap0 enable
root@node2: nc -6dl 12345

root@node1: vppctl set tap gso tap0 enable
root@node1: echo "test" | nc -6q 1 2001::2:1 12345

Here the connection fails. Packet trace shows the source IPv6 address of packets received on node2 dpdk-input is not the same as when the packets left node1 dpdk-output. Specifically the second segment, eg. 2001::1:1 becomes 2001:x::1:1 though the actual value of x seems to be different for each packet.