17 Oct
2014
17 Oct
'14
11:51 a.m.
Hello mesh-dev.
I think we may finally have an explanation of the vexing issue of "I can't
connect to the internet over peoplesopen.net."
Marc and I spent some time staring at wireshark dumps and thinking
about why some clients are unable to consistently connect via the
tunnel last night. I think Marc came up with a disappointing but correct
answer: it is basically an mtu issue (mtu is not being discovered
correctly), BUT there is no good fix because we are tunneling at layer 2.
When a packet arrives at a node from a client with too large an mtu,
what SHOULD happen in a normal forwarding situation (per RFC 1191)
is that the node issue a ICMP "Destination Unreachable" packet with a
"Fragmentation required" code. The client then uses this information to
reset its mtu.
This doesn't happen because we aren't really forwarding (forwarding happens
at layer 3). Instead, our interfaces (open0 and bat0) are bridged. So if a
frame
coming from open0 doesn't fit into bat0 it most likely gets silently
dropped.
So bridging open0 with bat0 is a disaster. A quick fix might be to replace
bridging with forwarding (at the IP level). I suspect this is not the right
thing
to do. It might be better to abandon the idea of meshing at layer 2; there
are numerous advantages to this.
In any case; we should discuss options this Tuesday.
Alex
12:19 p.m.
Also - just to throw something out there, this architecture is pretty cool:
http://libre-mesh.org/projects/libre-mesh/wiki/Network_Architecture?version…
I think there may be some implications here though for whether or not
clients actually have network addressable hostnames and/or how something
like mdns or an equivalent would work. It's also important to keep in
mind that we may be able to iterate towards more complicated setups, and
that there will likely always be the hot new idea that we want to
incorporate but that we have VERY constrained developer hours and may
need to make some satisficing decisions.
On 10/17/14, 12:06 PM, Max B wrote:
12:38 p.m.
Max,
We can't do what you're suggesting because the open0 interface is not
operating as a layer 3 interface. It is
bridged to br-openmesh along with bat0. This means that an over-sized
packet (e.g. one that doesn't fit in the
L2TP envelope of size ~1400bytes) arriving at open0 and headed toward bat0,
wouldn't trigger an ICMP
response, and would be unceremoniously dropped.
The ICMP response is triggered by the IP protocol layer (layer 3). That
response is also the only way
a Windows (and I think OS X) client would know that the mtu is smaller than
it thinks.
Assuming this is correct, which is still up for debate, we have two options:
(1) find a way to make the Win/OSX (Android/iOS?) client understand that it
must use a lower mtu
(DHCP is not an option).
(2) remove the bridge and forward at layer 3 (so that ICMP responses would
be triggered, and the
client can discover its mtu).
Alex
2014-10-17 12:06 GMT-07:00 Max B <maxb.personal(a)gmail.com>om>:
19 Oct
19 Oct
3:17 p.m.
Can we get a more clear understanding of how exactly we might reproduce
these issues? All of my desktop clients (mac osx, linux, windows 7) seem to
be able to connect. They're all able to view youtube videos and seem to
browse the web fairly regularly. My android phone is working for 90% of
applications, although it looks like maybe some of the apps aren't
connecting reliably. That being said, some of these apps seem to always
have some sort of connectivity issues and I'm having a hard time isolating
them.
Furthermore, isn't this MTU problem an issue that would occur on every
batman-adv network that is connected to the internet? I'm not able to
articulate this as well as I'd like, but I'm not seeing how this is
specific to our particular network structure...
Also - tried this on my picostation:
root@my:~# iptables -t mangle -A POSTROUTING -s 10.0.0.0/8 -p tcp
--tcp-flags SYN,RST SYN -j TCPMSS --set-mss 1400
root@my:~# iptables -t mangle -A POSTROUTING -d 10.0.0.0/8 -p tcp
--tcp-flags SYN,RST SYN -j TCPMSS --set-mss 1400
I'm wondering, though, that if this is also layer 3 routing, it probably
won't flag the sort of issue that you're describing...
I'm curious though, if the scenario you've described is accurate, why
wouldn't the bridge (which is layer 3, and which has an ip address and a
set mtu) respond with the ICMP response? In terms of layer 3 traffic, we
have a client with a layer 3 ip addr and then we also have a mesh node with
a layer 3 ip addr (which is the bridged interface).
Also - are these only hosts which have dhcp clients that don't respect the
MTU option?
Hopefully catch you all on Tuesday, but things have been a little crazy on
my end, so we'll see....
On Fri, Oct 17, 2014 at 12:38 PM, Alexander Papazoglou <papazoga(a)gmail.com>
wrote:
7:35 p.m.
Met with Alex today and we went over a bunch of stuff.
TL;DR
Iphones and other OS now work(! ...mostly), but UDP is probably still
broken. There may be some other hacked solutions, but we may start taking a
more serious look at babel or bmx6.
Alex explained because open0 and bat0 are bridged and because the
destination address of a packet is on the layer 2 network managed by
batman, the br-openmesh bridged interface won't return the mtu.
The first place that layer 3 routing happens is on the exit/relay server.
Which is already after the place an MTU drop would have taken place.
One partial solution that we implemented was setting up the TCPMSS Clamping
on the exit server. This commit handles that"
https://github.com/sudomesh/exitnode/commit/66a7523895053357a993a4ff61362eb…
We tested it and an iphone was able to connect to the peoplesopen.net
network and the variety of functionality that we attempted seemed to be
working just fine. So that's nice! However, this will ONLY work for TCP
connections. Everything over udp will still have the mis-matched MTU issues.
Alex and I mentioned that it would be fairly easy to unbridge the open0 and
bat0 interfaces, assign them IP addresses, and set up forwarding rules over
them. That way, traffic arriving at the node would actually be layer 3
routed and the node would have the opportunity to return the ICMP MTU
response. It might work, but it's a kind of weird hack that is maybe less
ideal.
We also talked a lot about how maybe the fundamental issue here is trying
to create a layer 2 mesh network over layer 3. We agreed that we'd take a
more serious look at what exactly an implementation of a layer 3 mesh
protocol might require/how it would be implemented and that we may start
some amount of parallel work testing out babel or bmx6. If anyone is
interested in that, let us know and we can figure out how to split up that
work.`
Max
On Sun, Oct 19, 2014 at 3:17 PM, max b <maxb.personal(a)gmail.com> wrote:
3798
days inactive
3801
days old
6 comments
3 participants
participants (3)
-
Alexander Papazoglou -
Max B -
Mo Balaa