This would break things for me. We're doing a lot of L2 over L3 site to
site stuff and we are using wireguard as the outer layer. Inner layer is
vxlan or l2tpv3.
In particular, people connect lots of stuff with no regard for MTU. For
some things it's also very hard to change so we just assume people
don't. Since the L3 network typically has the same MTU as the inner L2
network, we need fragmentation. There is no practical way to be able to
tell hosts on the L2 network about the limited mtu, for all we know they
don't even run IP....
It really does work without a hassle, it is not very very slow at all.
Performance is down perhaps by a factor of 3 compared to setting a
smaller MTU/MSS, but we can still push 350mbit/s with an atom 2ghz cpu,
and around 800mbit/s with a xeon cpu, with fragmentation for most
packets. This is one case where wireguard really works well!
IMHO, having wireguard generating fragmentable packets adds a lot to its
usefulness. With that said, it's not the end of the world for me as I
can just compile my own but I'd rather not :-)
On 2021-06-06 11:13, Jason A. Donenfeld wrote:
Hi,
WireGuard is an encrypted point-to-multipoint tunnel, where onion
layering of packets via a single interface or multiple is a useful
feature. This makes handling routing loops very hard to manage and
detect. I'm considering changing and simplifying loop mitigation to a
different strategy, but not without some discussion of its
implications.
Specifically the change would be to not allow IP fragmentation of the
encrypted UDP packets. This way, in the case of a loop, eventually the
packet size exceeds MTU, and it gets dropped: dumb and effective.
Depending on how this discussion goes, a compromise would be to not
allow fragmentation, but only for forwarded and kernel-generated
packets, not not for locally generated userspace packets. That's more
complex and I don't like it as much as just disallowing IP
fragmentation all together.
Pros:
- It solves the routing loop problem very simply.
- Usually when people are fragmenting packets like that, things become
very, very slow anyway, and it'd be better to just stop working
entirely, so that people adjust their MTU.
- Is anybody actually relying on this?
Cons:
- Maybe people are running
wireguard-over-gre-over-vxlan-over-l2tp-over-pppoe-over-god-knows-what-else,
and this reduces the MTU to below 1280, yet they still want to put
IPv6 through wireguard, and are willing to accept the performance
implications.
- Some people don't know how to fix their MTUs, and breaking rather
than just becoming really slow isn't the best outcome there, maybe.
- Maybe people are relying on this?
Before anybody asks: we're not going to add a knob for this, simply by
virtue of this being a decision with pros and cons. Please don't bring
that up.
I'd be very interested in opinions about this. Are there additional
pros and cons? I know the matter has come up a few times on the list,
mostly with people _wanting_ fragmentation (I've CCd a few people from
those threads - Roman, I expect you to vigorously argue the
pro-fragmentation stance ;-). but I'm not convinced the outcome of
those threads was correct, other than, "yea, that's easy enough to
enable." But on the other hand, maybe the cons are real enough we
should rethink this.
Please let me know thoughts and ideas.
Thanks,
Jason
