On 9/22/21 11:23 PM, Francesco Ruggeri wrote:
On Tue, Sep 21, 2021 at 9:15 AM Leonard Crestez cdleonard@gmail.com wrote:
- Sequence Number Extension not implemented so connections will flap
every ~4G of traffic.
Could you expand on this? What exactly do you mean by flap? Will the connection be terminated? I assume that depending on the initial sequence numbers the first flaps may occur well before 4G. Do you use a SNE of 0 in the hash computation, or do you just not include the SNE in it?
SNE is hardcoded to zero, with the logical consequence of incorrect signatures on sequence number wrapping. The SNE has to be included because otherwise all signatures would be invalid.
You are correct that this can break much sooner than 4G of traffic, but still in the GB range on average. I didn't test the exact behavior (not clear how) but if signatures don't validate the connection will likely timeout.
My plan is to use TCP_REPAIR to control sequence numbers and test this reliably in an isolated environment (not interop with a cisco VM or similar). I want to implement TCP_REPAIR support for TCP-AO anyway.
It was skipped because the patch series is already quite large.
-- Regards, Leonard