PingTimeout

[Cédric Lemarchand]

Le 24/07/13 22:02, Etienne Dechamps a écrit :
> (Hey Cédric, it appears we're both following tinc *and* ZFS On Linux. 
> What are the odds...)
(Hey Etienne, they are both great pieces of software, and it seems we 
are, both again, using them! )
>
> On 24/07/2013 20:29, Cédric Lemarchand wrote:
>> I use tinc with QOS, and since some weeks I got problem with tunnel
>> disconnection because the default "Pingtimeout" of 5 sec is reach.
>> I think it happens because of the QOS shappe the tinc traffic (data
>> here) to let some BP for VOIP, but whate is very strange is that the
>> ICMP protocol is not include in the shapping, so tinc would never 'see'
>> latency on the link, even if there is latency caused by the QOS.
>>
>> So the question is does tinc really use ICMP to check remote hosts or
>> use it's own UDP packets ?
>
> tinc doesn't use ICMP. It uses PING/PONG messages on the TCP 
> metaconnections to ensure the control graph is healthy, and UDP 
> messages (MTU probes) to opportunistically check for potential UDP 
> connectivity.
>
> If your nodes are getting disconnected because of timeouts, it's 
> because tinc's PING/PONG messages are not getting through on your TCP 
> metaconnections.
Hum very interesting, so if I *only* shape the tinc's udp flows (aka 
data channels) between my nodes, tinc daemons will not get disconnected 
by the PingTimeout, even if the data channel become very busy, thus get 
hight latency (except of course in the case of a real links congestion).
Sounds good on the paper, but what happens with UDP probes ? are they 
only used for MTU path discovery or there is a similar PING/PONG checks ?