operational stoppers to > 1500 MTU deployment in real life (Re: Path MTU Hop-by-Hop Option)

[Mikael Abrahamsson <[e-mail]>]

On Wed, 17 Oct 2018, George Michaelson wrote:

> dreaming of a larger packet future...

It's completely undeployable in real life without something like 
https://tools.ietf.org/html/draft-van-beijnum-multi-mtu-05

It's not realistic to expect a whole lan of managed and unmanaged devices 
to show up at a LAN and communicate with each other on a large MTU LAN.

Right now my only tools are to crank up the L2 switch to 9000 MTU and 
advertise MTU=9000 in RA. Ok, I show up with my device that only has 1500 
MTU/MRU, I see that RA and I'm like "well, I can't do higher than 1500 
anyway, so I'll continue doing that". I need to send a UDP packet to the 
other guy on the same LAN, that's fine, he has 9000 MRU/MTU. He then needs 
to send me 2000 bytes back, so he does that and because my MRU is only 
1500, my NIC now silently drops that packet. No PMTUD, just silent drop. 
*poof* gone.

There are LOTS of similar failure modes that can happen. We need 
per-destination PMTU both on/off-link. I need to be able to send RIO and 
say "ok, ::/0 has 1492 PMTU" while at the same time send PIO with 
"MTU 9000". Devices then need to understand if their neighbours actually 
can do 9000 or not.

I've even seen hardware/software that will have interface MTU of 1500, 
take the RA MTU=9000, set it's IPv6 interface MTU to 9000, set MSS to 
89xx, send a large packet (because it can), then it borks when a response 
comes back because it only has ~2000 in MRU.

Yes, real life is that broken.

-- 
Mikael Abrahamsson    email: [e-mail]