Previously, we treated all IPv6 addresses as assignable, but that causes
problems with setups that expect the first address in a subnet to be the
router anycast address.
Note that this does not fix existing innernet networks, and those
experiencing this problem are advised to revised to recreate their
network after this fix has been merged. Sorry for the annoyance.
Fixes#131
This change adds the ability for peers to report additional candidate endpoints for other peers to attempt connections with outside of the endpoint reported by the coordinating server.
While not a complete solution to the full spectrum of NAT traversal issues (TURN-esque proxying is still notably missing), it allows peers within the same NAT to connect to each other via their LAN addresses, which is a win nonetheless. In the future, more advanced candidate discovery could be used to punch through additional types of NAT cone types as well.
Co-authored-by: Matěj Laitl <matej@laitl.cz>
The past behavior of clients was to, on every fetch from the server, update each of its peer's endpoints with the one reported from the server. While this wasn't a problem on certain types of NATs to help with holepunching, in some situations it caused previously working connections to no longer work (when one peer had a port-restricted or symmetric cone type NAT).
Based on the conversation from #5 (comment) - this changes innernet's behavior on Linux from automatically falling back to the userspace, instead requiring --backend userspace to be specified.
This should help people avoid weird situations in environments like Docker.
The server now expects a UNIX timestamp after which the invitation will be expired. If a peer invite hasn't been redeemed after it expires, the server will clean up old entries and allow the IP to be re-allocated for a new invite.
Closes#24
Jake McGinty