[sr-dev] sr closes tcp connection before register expires

[Andrei Pelinescu-Onciul]

On Dec 08, 2009 at 07:33, Juha Heinanen <jh at tutpro.com> wrote:
> Klaus Darilion writes:
> 
>  > Isn't it possible to implement tcp_persistent_flag on sr core? IMO it is 
>  > a nice feature.
> 
> i agree, lets wait what andrei says about it.  if not possible, i'll
> remove the flag from registrar module.

It is possible, but do we really need it? (in theory is useful, but I'm
not sure about the practice :-)).

The side effect is that it would use more bytes per tcp connection (4 or
8) and that global runtime timeout changes
(e.g. sercmd cfg.set_now_int tcp connection_lifetime 120 )
will not affect anymore existing connection, only new ones.

I think the change would be quite trivial, but still I would need some
volunteer testers and it's also pretty late for sr3.0.

Note also that after a quick look I don't believe it ever worked
correctly in k. I might be wrong but to me it looks like in k it has the
following problems:
 - timeout is changed only once, when the REGISTER is received. After
   that the first receive on the same connection (be it another sip
   message or a simple CRLF ping) or the first send on the connection
   will _reset_ the timeout back to tcp_con_lifetime. So it works only
   if the connection is completely idle after the first REGISTER.
 - timeout is changed after tcp reader exits => error of at least 2s and
   with extra 10s added by registrar (less important).


Andrei