On Monday 16 August 2010, Jasmin Schnatterbeck wrote:
The issue is visible, whenever pua_dialoginfo sends out a PUBLISH request. During that process, build_cell is called, which destroys avps.
I experimented with adding a parameter. It works - nearly. PUBLISH is sent out, avps are OK, but after additional 2-3 seconds, one child process dies.
Hello Jasmin,
i think this happens probably because of the reasons Andrei mentioned in his reply (last wednesday) - that an internal AVP list cannout be shared between several transactions.
So there are these 2 possibilities:
- AVPs get destroyed
- kamailio dies a few seconds after sending out the PUBLISH
Which scenario happens depends on these lines in build_cell() (tm h_table.c:~300):
old = set_avp_list(AVP_TRACK_FROM | AVP_CLASS_URI,[..] *old = 0;
If these lines are present, AVPS get destroyed. If not, kamailio child dies a few seconds after sending out PUBLISH.
I tried different things here, but I did not figure out yet, why this happens. I assume that somewhere (may through a timer) somewhat is called, which uses the parameters of the cell, which refer to avps. And if there is no AVP list (but e.g. a 0 pointer), the child processing this code dies.
There is a delete timer for transaction which does some cleanup, maybe in a way like you described here.
So I would like to ask Andrei, whether if it is possible to investigate this problem - because at this moment, it makes the use of pua_dialoginfo totally impossible.
There are tow two other workarounds, but both are not really good or practical:
* changing your proxy that it not keep transaction state * changing pua_dialoginfo that it not send its PUBLISH transactional
So i think the best way to proceed is doing a change in TM like Andrei suggested, changing the code that it not copy the AVP list for a local transaction and instead create one with an empty AVP list.
Cheers,
Henning