BGP adjacency not established on linux-image-5.15.0-1019-aws

[Domonkos Tomcsanyi]

We have a simple eBGP setup with a peer over IPsec using a VTI interface using frr (default BGP daemon in Ubuntu 22.04). Everything worked fine until we upgraded to the latest kernel available for our machine: linux-image-5.15.0-1019-aws. After that the adjacency would simply not come up. I have investigated for some time, and found no issues with the configuration (obviously, it was not modified since it was working fine) and also no other issues. There are no firewall rules on this machine. Also there is no issue with the IPsec and the VTI interface, traffic flows through it, the BGP neighbor is pingable etc.
We have confirmed that downgrading to linux-image-5.15.0-1017-aws fixes the issue.

According to tcpdump it seems like both ends initiate the connection, but then there are just a lot of TCP retransmissions and Duplicate ACKs after the OPEN messages. The other side sends a KEEPALIVE but FRR seems to be stuck getting it.
Basically it looks like according to the tcpdump that after the initial exchange of messages the ACKs are not recognized properly by the kernel (?) and therefore continous retransmissions occur until the timer of BGP expires.
Since other TCP applications (e.g. SSH) work fine I'm assuming this might be related to:
- VTI interfaces
- FRR itself

To drive initiatives in parallel I have opened a bug at FRR as well:
https://github.com/FRRouting/frr/issues/11943

[Domonkos Tomcsanyi (tomcsanyid)]

I have added a tcpdump to the github issue, if anybody wants to have a look.

[Domonkos Tomcsanyi (tomcsanyid)]

new kernel 5.19.0-1023-aws fixed the issue.