Reduce read buf of first stun message.
Add timeout to read first message to clean interrupted
connection earlier.
Add alive duration for gather to access connection created
from stun bind, avoid connection leak from malicious client.
Safari and iOS Chrome (WebKit) sometimes open multiple TCP
connections that end up being part of the same tcpPacketConn
object in TCPMuxDefault. After some time, the extra connection
is closed even though the main one is still operational. In this
case, the entire peer connection gets disconnected. Instead, we
probably only want to close the entire connection only once there
is no other open TCP connection (this PR).
To be honest, I am not sure why Safari and iOS Chrome do this.
However, there is probably no chance fixing it there.
This change adapts pion/ice to use a new interface for most network
related operations. The interface was formerly a simple struct vnet.Net
which was originally intended to facilicate testing. By replacing it
with an interface we have greater flexibility and allow users to hook
into the networking stack by providing their own implementation of
the interface.
Go encourages short names. And a quick
survey through the Pion code bases reveals,
that we have been using buf more often than buffer.
So this change will harmonize the different spellings.
When we have multiple host candidates and mux to a
single port, if these candidates share a same conn
(either tcp or udp), they might read other's msg
cause ice connection cost long time or failed.
Previously `(*TCPMux).RemoveConnByUfrag` would close the connections
while holding its lock, but it is possible for closing to block
and/or take a while.
If closing one of the connections blocks (which was occurring rarely
due to blocking on a channel send), then the entire `TCPMux` would
deadlock and all further operations with the ICE agent would end
up getting blocked due to tying up the task loop.
This change should help matters, but it's unclear if
`(*TCPMux).RemoveConnByUfrag` somehow needs to be made cancelable by
`Agent`'s task context, as there are no guarantees that closing
a connection will happen quickly.
A newline is automatically added by the DefaultLeveledLogger.
The current code-base used a mix of log statements w/o endlines.
This is misleading for custom loggers.
This addresses a few points issue of #245:
- Take a net.Listener instead of having global state
- Expose a net.TCPMux based API
Also, the unused closeChannel was removed from tcp_mux.go
Closes#253.
