[Serdev] SIP UDP / dont fragment issue

[Martin Koenig]

Hello Valentin,

Valentin Nechayev wrote:

> Wed, Jun 01, 2005 at 13:28:31, martin.koenig wrote about "Re: [Serdev] SIP UDP / dont fragment issue": 
>
>  
>
>>it is implemented via ICMP type 3 (dest. unreachable) code 4 
>>(fragmentation needed), and the MTU of the next hop specified in this 
>>ICMP message.
>>    
>>
>
>Yes, it is defined on IP level (regardless of fact that MTU isn't
>obligatory).  But setting DF on UDP packets raises a bunch of
>problems. First, one should add reliable, portable and well-known
>mechanism to pass MTU limitation from kernel to userland. BSD
>sockets interface doesn't provide that. Second, even if one knows
>maximal available data size ont the path, how can it be applied to
>real protocols which doesn't allow arbitrary splitting? Almost any
>UDP-based protocol (including SIP over UDP) doesn't allow splitting
>of a message to a few packets and (for efficiency) gathering parts
>of a few messages to one packet. TCP allows this due to its concept
>of established flow and possibility to cut the flow on any octet
>boundary. SIP doesn't have such mechanism.
>  
>
It works very well for SIP and UDP with Ser. It works with all ATAs and 
Phones that we use in our platform.

>  
>
>>It is very correct to set the dont fragment bit here, because the 
>>message is already fragmented and an endpoint would not be able to 
>>aggregate these further fragmented messages.
>>    
>>
>
>NO! IP fragmentation isn't very efficient, but it is reliable to
>multiple fragmentation stages: fragment can be re-splitted and such
>parts can be again gathered in correct way. And we again return to
>the previous paragraph: unless you have reliable way to notify
>stream splitter for a minimum of full-way-through maximal segment
>size, as in TCP implementation, you can't detect correct MSS in
>other way than test-and-fail technique.
>  
>
I've been told differend by a collague of our networking department. 
However I will make sure he gets your message to send me feedback.

>>Thats why the packet will be dropped at the router and ICMP dest. 
>>unreachable must be responded.
>>    
>>
>
>HOW it will be responded? Please name the mechanism how simple
>program without root privileges and with one UDP socket can get the
>MTU returned to the kernel. Mechanism must be well-portable and
>widely known.
>  
>
Valentin, I'm not a programmer, nor do I plan to become one. It simply 
works, please talk to the responsible Ser developer about details how 
they did do this job.

Regards,
Martin