Re: Use cases for PMTUD and PLPMTUD (was: RE: 6MAN: Adoption call on draft-hinden-6man-rfc1981bis-01)

[Fernando Gont <[e-mail]>]

On 02/07/2016 01:15 PM, Michael Richardson wrote:
> 
> Fernando Gont <[e-mail]> wrote:
>     >> Templin, Fred L <[e-mail]> wrote: > On the other
>     >> hand, paths that lead to Internet destinations cannot be > counted on
>     >> to deliver the necessary ICMPs. In that case, RFC4821 > provides a
>     >> mitigation.
>     >>
>     >> > But, if we do not believe that there are paths for which traditional
>     >> > PMTUD can still be used safely, then we should be working to
>     >> deprecate > RFC1981 instead of making it a standard.
>     >>
>     >> I think that the mitigation is far more reliable than 1981.  So, I'd
>     >> rather deprecate RFC1981.
>     >>
>     >> {I noticed that 4821 is not on by default in most Linux
>     >> distros/kernels, and I actually want to know if there is some reason
>     >> for it}
> 
>     > Yes: convergence time.
> 
> Yes, I've noticed that it takes a few re-transmissions on the first
> connection to a site in the case where the MTU is limited and the ICMPs are
> lost, but when the MTU is unblocked, there is no delay.

If you employ RFC4821 as icmp-less PMTUD, it doesn't matter whether the
ICMPs get blocked or not. If you implement RFC4821 for blackhole
detection, then I agree with you.




> An implementation that wanted to avoid even the re-transmissions in the
> blocked case could limit themselves to 1280 sized packets and send some 1500
> byte probes have gotten through.
> 
> If 4821 is turned off, then the connection is mysteriously blocked, and
> always fail.
> 
> So, your answer does not make sense to me.

"My" answer is actually what I recall from discussing this with Linux
kernel developers 10 years ago -- I guess that they were referring to
"RFC4821 as icmp-less PMTUD".

That said, if the toggle in Linux is to use RFC4821 as icmp-less PMTUD,
then my (relayed) response would apply. OTOH, if the toggle is simply to
employ RFC4821 for blackhole detection then I'd agree with you.

FWIW, I'd expect RFC4821 fo blackhole detection to be "on by default",
and, if anything, to be able to turn on RFC4821 as icmp-less PMTUD (but
this one *not* "on by default", for the reasons already stated).

Thanks!

Best regards,
-- 
Fernando Gont
SI6 Networks
e-mail: [e-mail]
PGP Fingerprint: 6666 31C6 D484 63B2 8FB1 E3C4 AE25 0D55 1D4E 7492