On Fri, Dec 26, 2014 at 1:20 PM, Ferry Huberts <(spam-protected)> wrote:
> Maybe we should add some 'send' checks in the relevant send paths.
> In the pud plugin I had to add
>>
>> /* force the pending buffer out if
>> there's not enough space for our message */
>> if ((int)olsrMessageLength >
>> net_outbuffer_bytes_left(ifn)) {
>> net_output(ifn);
>> }
>
>
> because sometimes the buffer was already too full and I got error reports in
> the log. With these lines added those errors disappeared.
We have the same for Hello/TC messages...
see tc_set.c line 564 as an example.
Yes, the code is really ugly because its done once per message
(instead of doing it in a generic way), but it should work (I think).
> I also don't see anything explictly related to DF, I do see fiddling with
> interface MTUs. Code that - to me - looks entirely superfluous since we
> should let the network stack handle that kind of stuff.
Fragmenting the OLSR UDP messages on IP level would interfere with the
ETX calculation... I think the "read the MTU and make them small
enough so they should not fragment" is a good approach.
> I think there is not a particular good reason to not allow fragmenting.
> Once the mesh grows larger than 159 (for a normal 1500 MTU interface) we do
> need fragmentation.
We already do fragment OLSR messages when necessary... the problem
just happens because the interface reports the wrong MTU and the
default settings of the IP stack prevent the IP stack from fragmenting
our UDP packets.
Henning Rogge