Ok, I'll give this a look. I have a theory of what's going on.
Best regards, Byron Campen Thanks, and yes this is happening in repro. For comparison, I’ve attached two logs, one when the callee returns an OK and the other when the callee does not. In both cases the caller is sending a single BYE. BR, Jay Denkberg It is up to the proxy to make this happen. When processClientNonInvite() is called, this will cause a 408 to be sent up to the TU (transaction user, in this case the proxy). When the proxy gets this 408, it should be turning around and forwarding it back down to the stack on the server transaction. This will cause the server transaction to be cleaned up. (Keep in mind; when a request comes up to the TU from the stack, the stack will not clean up until the TU gives it a response to the request. Forwarding the 408 accomplishes this.) I should ask, is this happening with repro? If so, there is a bug in repro, and I would like to see a trace. Byron Campen
I noticed that when the callee responds both processServerNonInvite() and processClientNonInvite() are called with causes the transaction count go to 0 and therefore it is freed. However when the callee does not respond with a 200OK processClientNonInvite() is eventually called after all the retransmits have been exhausted but processServerNonInvite() is never called so the transaction count never reaches 0 and is never released. I also see that when the OK is received both TimerK (client) and TimerJ (server) are invoked. However when no OK is ever received neither timers are set up. I’m thinking (but I’m not at all sure) that if Timer J would be added when the retransmits are exhausted then the server side would also release decrement the transaction count and the memory would be released. Is this analysis correct? What is the correct way to decrement the transaction count a second time, when the retransmits have been exhausted? Thx, Jay
From: Brocha Strous
Sent: Monday, May 21, 2007 8:49 PM
To: Robert Sparks; Byron Campen
Cc: resiprocate-devel@xxxxxxxxxxxxxxxxxxxx; Jay Denkberg
Subject: RE: [reSIProcate] Leaking transactions Byron+Robert, The situation Jay is describing involves repro functioning as a proxy. The caller sends a BYE which repro responds to with 200OK and then forwards to the callee. The callee happens to not respond. Repro keeps retransmitting for a while then stops but the server-side transaction is not cleaned up. (And after many of these the memory is also growing). (We are not using DUM). Thanks, Brocha To push Byron's point a little further - Your text talks about a client transaction (sending the BYE and watching it being retransmitted). The log you point to is talking about a server transaction. Are you running the A-side and B-side (as you called them) in the same instance of resip? If so, the statistic you highlight is due to the B-side BYE handler not handing a response back to the stack. On May 21, 2007, at 8:36 AM, Byron Campen wrote: For server transactions, the TU is responsible for sending a response to the stack no matter what. Failure to do so will cause that transaction to hang around indefinitely. The stack ensures that the request is well-formed enough to at least forge a failure response before passing the request to the TU, meaning that it is possible to give some sort of response to everything that is passed up. This is, as far as I can tell, the number one cause of leaks in resip-based applications. Would the same thing be true on the server side. Namely when resip send tries to pass on a BYE to the B-side and the BSide does not reponsd. I see that resip is re-transmitting (that’s a good thing) After the internal timeout it stops retransmitting (another good thing) however the transaction count is not decremented (isn’t this a bad thing?). Even after 15 minutes, the statistics message displays the following: WARNING | 20070521-043519.633 | xxxxxx | repro | RESIP:STATS | 12134 | 3054312368 | StatisticsMessage.cxx:152 | RESIP:TRANSACTION TU summary: 0 TRANSPORT 0 TRANSACTION 0 CLIENTTX 0 SERVERTX 1 TIMERS 0 Transaction summary: reqi 1 reqo 11 rspi 0 rspo 1 Details: INVi 0/S0/F0 INVo 0/S0/F0 ACKi 0 ACKo 0 BYEi 1/S0/F0 BYEo 1/S0/F0 CANi 0/S0/F0 CANo 0/S0/F0 MSGi 0/S0/F0 MSGo 0/S0/F0 OPTi 0/S0/F0 OPTo 0/S0/F0 REGi 0/S0/F0 REGo 0/S0/F0 PUBi 0/S0/F0 PUBo 0/S0/F0 SUBi 0/S0/F0 SUBo 0/S0/F0 NOTi 0/S0/F0 NOTo 0/S0/F0 Retransmissions: INVx 0 BYEx 10 CANx 0 MSGx 0 OPTx 0 REGx 0 finx 0 nonx 0 PUBx 0 SUBx 0 NOTx 0 What can the A Side do? Resip already sent out the OK to the A side and it has no idea that the B-side never esponded? I agree it needs to be documented better. : ) BTW: You don’t really need to do the mMyClientInviteSessionHanlde.isValid and the isEarly check – it is OK to call AppDialogSet::end() in both cases – it will dispatch the CANCEL vs BYE appropriately. Scott Scott, thanks for that. I changed my UAC app to do the following as you suggested and the memory leak has gone away: MyAppDialogSet::EndSession() if (mMyClientInviteSessionHandle.isValid()) if (mMyClientInviteSessionHandle->isEarly()) mMyClientInviteSessionHandle->end(); I forgot to mention, previously in my leaky app (where I was calling ClientInviteSession::end() only), I was still getting the expected AppDialogSet::destroy() call back from the DUM at the end of each session. This is a bit misleading as it implied the session had successfully 'ended' which wasn't the case? This behaviour isn't at all obvious.
From: resiprocate-devel-bounces@xxxxxxxxxxxxxxxxxxxx [mailto:resiprocate-devel-bounces@xxxxxxxxxxxxxxxxxxxx] On Behalf Of Scott Godin
Sent: 03 May 2007 13:46
To: Greg Inglis; resiprocate-devel@xxxxxxxxxxxxxxxxxxxx
Subject: Re: [reSIProcate] Leaking transactions Hmm – calling ClientInviteSession::end() will cause a BYE to sent, and the original INVITE transaction should still be active, since the invite was not CANCELED. I would have thought that the transaction state would have been cleaned up after some timeout – but there may currently be a requirement for the application to cancel the transaction completely since you did receive a provisional response. You could try switching your call to AppDialogSet()->end() or DialogUsageManager::end(), which should send a CANCEL instead to see if the problem goes away. Scott I am getting a memory leak in my application when I end my UAC session while the session is in a 'UAC_Early' state i.e. I've got a '180 Ringing' from the remote UAS endpoint but I haven't received a '200 OK' as yet. I'm getting the following statistics message from the stack afterwards: TU summary: 0 TRANSPORT 0 TRANSACTION 0 CLIENTTX 1 SERVERTX 0 TIMERS 0
Transaction summary: reqi 0 reqo 6 rspi 8 rspo 0
Details: INVi 0/S0/F0 INVo 2/S0/F1 ACKi 0 ACKo 1 BYEi 0/S0/F0 BYEo 1/S1/F0 CANi 0/S0/F0 CANo 0/S0/F0 MSGi 0/S0/F0 MSGo 0/S0/F0 OPTi 0/S0/F0 OPTo 0/S0/F0 REGi 0/S0/F0 REGo 2/S1/F1 PUBi 0/S0/F0 PUBo 0/S0/F0 SUBi 0/S0/F0 SUBo 0/S0/F0 NOTi 0/S0/F0 NOTo 0/S0/F0
Retransmissions: INVx 0 BYEx 0 CANx 0 MSGx 0 OPTx 0 REGx 0 finx 0 nonx 0 PUBx 0 SUBx 0 NOTx 0 The 'CLIENTTX 1' in the TU summary doesn't go away no matter how long I wait - I assume this means a transaction has leaked somehow? I am ending the session by calling ClientInviteSession::end() and subsequently calling process() on the stack and DUM - is there some step I am missing here? I'm using reSiprocate 1.1 built on VC8 SP1 and my SIP proxy is asterisk. Any help would be appreciated _______________________________________________ resiprocate-devel mailing list _______________________________________________ resiprocate-devel mailing list
<BYE_NO_OK.txt> <BYE_WITH_OK.txt>
|