Re: [regression] TCP_MD5SIG on established sockets

[Eric Dumazet]

On Tue, Jun 30, 2020 at 1:34 PM Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> ----- On Jun 30, 2020, at 3:52 PM, Linus Torvalds torvalds@xxxxxxxxxxxxxxxxxxxx wrote:
>
> > On Tue, Jun 30, 2020 at 12:43 PM Linus Torvalds
> > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >>
> [...]
> > So I think it's still wrong (clearly others do change passwords
> > outside of listening state), but considering that it apparently took
> > people two years to notice, at least some of the onus on figuring out
> > a better morel is on people who didn't even bother to test things in a
> > timely manner.
>
> I'm fully willing to work with Eric on finding a way forward with a
> fix which addresses the original issue Eric's patch was trying to
> fix while preserving ABI compatibility.
>
> The main thing we need to agree on at this stage is what is our goal. We
> can either choose to restore the original ABI behavior entirely, or only
> focus on what appears to be the most important use-cases.
>
> AFAIU, restoring full ABI compatibility would require to re-enable all
> the following scenarios:
>
> A) Transition of live socket from no key -> MD5 key.
> B) Transition of live socket from MD5 key -> no key.
> C) Transition of live socket from MD5 key to a different MD5 key.
>
> Scenario (C) appears to be the most important use-case, and probably the
> easiest to restore to its original behavior.
>
> AFAIU restoring scenarios A and B would require us to validate how
> much header space is needed by each SACK, TS and MD5 option enabled
> on the socket, and reject enabling any option that adds header space
> requirement exceeding the available space.
>
> I welcome advice on what should be the end goal here.
>

The (C) & (B) case are certainly doable.

A) case is more complex, I have no idea of breakages of various TCP
stacks if a flow got SACK
at some point (in 3WHS) but suddenly becomes Reno.