[libnice] Issues with candidate selection through wrong network interface

[Olivier Crête]

Hi,
We've seen the same thing with the libvirtd interface from KVM. And I
spend some time looking into this and I know Fabrice also did and we
couldn't find a great solution, so in commit b4abda09 he just added an
option to ignore one specific prefix. But maybe we should extend that
to a blacklist of prefixes (so we can ignore docker*, virbr*, br*. We
currently explicitly ignore localhost in the code, maybe we should
ignore more. I'm just worried that in some cases those may be
legitimate interfaces and may actually be the right choice. For example
if the peer is running inside docker!
Olivier
On Tue, 2017-08-29 at 16:19 +0200, Juan Navarro wrote:
> Good afternoon,
> 
> I have been studying a situation where libnice chooses the wrong
> network 
> interface for WebRTC connections in a public-facing server: an
> Amazon 
> Web Services instance, where all TCP and UDP ports are totally open, 
> hosting a WebRTC application. This machine also has Docker installed
> for 
> other unrelated purposes. This means that there are two network 
> interfaces: eth0 and docker0, and when ICE finishes, libnice always 
> selects the IP address of the docker0 interface as the local
> candidate!
> 
> A TURN server (coturn) is used for ICE. I guess that the STUN
> packets 
> are able to come out through the docker0 interface, during the 
> connectivity checks, which makes this a valid candidate; however
> later 
> the actual WebRTC stream doesn't work at all. And it is my
> understanding 
> that even if it worked, it would still be an error, because outbound 
> streams would be doing a totally unnecessary loop from the local 
> application to the docker0 interface, and out of Docker to the
> public 
> internet. Likewise for inbound streams.
> 
> Running "sudo ifconfig docker0 down" solves the issue: docker0 had
> been 
> brought down, so now libnice chooses the correct interface, eth0
> (ie. 
> the local IP address of that interface) for the local candidate
> which 
> gets selected.
> 
> This happens with the latest libnice version available on Ubuntu
> 16.04 
> (Xenial): 0.1.13. It also happens with the latest release of
> libnice, 
> 0.1.14.
> 
> However, this does _not_ happen with the latest development version
> from 
> Git, which is commit dbaf8f5: libnice now chooses the IP address 
> corresponding to the interface eth0, and the WebRTC connections work 
> fine (the RTP streams are sent and received), even if a docker0 
> interface is present.
> 
> Now here comes the problem: I have some other machines where there is
> no 
> docker0, but there are some other kind of virtual interfaces, such
> as 
> br-* (eg. a machine with 8 interfaces named like "br-1adb6aca2ec6", 
> which are bridged interfaces if I'm not mistaken). In those cases,
> the 
> latest libnice still behaves like the previous versions: one of
> those 
> br-* interfaces gets selected as local candidate, which makes the 
> following streams to fail reaching its destination. This even
> happens 
> for localhost connections! (where the Chrome browser and the libnice 
> application are running in the same machine). However bringing all
> of 
> the virtual interfaces down -with ifconfig- makes it work (leaving
> only 
> the actual network interface which has a more direct connection to
> the app).
> 
> 
> 
> 
> I would like to gain a deeper knowledge of how all this works in 
> libnice, and how should I act in order to solve this problem. I
> guess 
> there must be lots of people out there running into the same issue,
> just 
> by having Docker installed, or any other software which creates
> virtual 
> network interfaces. So, how do you deal with this issue?
> 
> Kind regards,
> Juan
> 
> 
> _______________________________________________
> nice mailing list
> nice at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/nice
-- 
Olivier Crête
olivier.crete at collabora.com

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/nice/attachments/20170829/4fa18032/attachment.html>