Stefano Brivio <sbri...@redhat.com> wrote: > Hi, > > On Sun, 12 Jul 2020 22:07:03 +0200 > Florian Westphal <f...@strlen.de> wrote: > > > vxlan and geneve take the to-be-transmitted skb, prepend the > > encapsulation header and send the result. > > > > Neither vxlan nor geneve can do anything about a lowered path mtu > > except notifying the peer/upper dst entry. > > It could, and I think it should, update its MTU, though. I didn't > include this in the original implementation of PMTU discovery for UDP > tunnels as it worked just fine for locally generated and routed > traffic, but here we go.
I don't think its a good idea to muck with network config in response to untrusted entity. > As PMTU discovery happens, we have a route exception on the lower > layer for the given path, and we know that VXLAN will use that path, > so we also know there's no point in having a higher MTU on the VXLAN > device, it's really the maximum packet size we can use. No, in the setup that prompted this series the route exception is wrong. The current "fix" is a shell script that flushes the exception as soon as its added to keep the tunnel working... > > Some setups, however, will use vxlan as a bridge port (or openvs vport). > > And, on top of that, I think what we're missing on the bridge is to > update the MTU when a port lowers its MTU. The MTU is changed only as > interfaces are added, which feels like a bug. We could use the lower > layer notifier to fix this. I will defer to someone who knows bridges better but I think that in bridge case we 100% depend on a human to set everything. bridge might be forwarding frames of non-ip protocol and I worry that this is a self-induced DoS when we start to alter configuration behind sysadmins back. > I tried to represent the issue you're hitting with a new test case in > the pmtu.sh selftest, also included in the diff. Would that work for > Open vSwitch? No idea, I don't understand how it can work at all, we can't 'chop up'/mangle l2 frame in arbitrary fashion to somehow make them pass to the output port. We also can't influence MTU config of the links peer. > If OVS queries the MTU of VXLAN devices, I guess that should be enough. What should it be doing...?
