Hello again,

I took a quick look at the code in dset.c:append_branch().  This appears to update the global variable nr_branches, but the check in t_suspend.c:t_continue() is against t->nr_of_outgoings.  There are no calls into the tm module in the PRESENCE_DISTRIBUTE and PRESENCE_ENQUEUE routes except for t_suspend(), so I suspect that the call to append_branch() didn't work because the updated nr_branches value was not used to update t->nr_of_outgoings.  Should this be what happens?

Does this make sense?  Is there a simple work-around for this?

Thanks,

Peter

On Wed, 2012-03-28 at 15:30 +0100, Peter Dunkley wrote:
Hi,

The append_branch() hasn't helped.

Having looked a bit more closely, I think what I said at the end of my last email was not quite right.

The t_continue() that kills the transaction is actually the first one.  However, it only kills the transaction (causing a 500 to be sent) after it has been successfully suspended.  What this means is that the second t_continue() seems to manage to resume the transaction (despite it having previously been killed), but at this point the presence APIs go wrong because the transaction cannot be statefully replied to.  Here is some log content:
1: Mar 28 12:02:28 pd-laptop-linux kamailio[16424]: INFO: <script>: presence: Found queued transaction [57508:831770136]
2: Mar 28 12:02:28 pd-laptop-linux kamailio[16424]: INFO: <script>: PUBLISH: route[PRESENCE_DISTRIBUTE]
3: Mar 28 12:02:28 pd-laptop-linux kamailio[16424]: INFO: <script>: Adding to queue: presenceWorker4
4: Mar 28 12:02:28 pd-laptop-linux kamailio[16424]: WARNING: <script>: PUBLISH: route[PRESENCE_ENQUEUE]
5: Mar 28 12:02:28 pd-laptop-linux kamailio[16424]: INFO: <script>: Suspended transaction for PUBLISH [57508:831770136]
6: Mar 28 12:02:28 pd-laptop-linux kamailio[16424]: ERROR: tm [t_suspend.c:223]: branch == t->nr_of_outgoings
7: Mar 28 12:02:28 pd-laptop-linux kamailio[16470]: ERROR: pua [pua_db.c:905]: no rows deleted
8: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: WARNING: <script>: presenceWorker4: found queued transaction [57508:831770136]
9: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: WARNING: <script>: PUBLISH: route[PRESENCE]: presenceWorker4
10: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: tm [t_reply.c:591]: ERROR: _reply_light: can't generate 200 reply when a final 500 was sent out
11: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: sl [sl.c:270]: failed to reply stateful (tm)
12: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: presence [presentity.c:154]: sending reply
13: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: presence [presentity.c:400]: sending 200OK
14: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: presence [publish.c:487]: when updating presentity
15: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: tm [t_reply.c:591]: ERROR: _reply_light: can't generate 500 reply when a final 500 was sent out
16: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: sl [sl.c:270]: failed to reply stateful (tm)
17: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: presence [utils_func.c:146]: sending 500 Server Internal Error reply
18: Mar 28 12:02:28 pd-laptop-linux kamailio[16412]: ERROR: presence [publish.c:517]: failed to send error reply

The scenario here is that pua_usrloc sends a PUBLISH to the local Kamailio instance.  Line 7 is the pua module handling the 500 response sent by t_continue().  Lines 10 through 18 are presence getting the PUBLISH, handling it properly (updating the DB and so on), but then failing to respond, because a 500 has already been sent for the transaction.

Thanks,

Peter

On Wed, 2012-03-28 at 14:44 +0100, Peter Dunkley wrote:
Hi,

I am not relaying or replying to messages directly here - except in the error case.  I am using the t_suspend()/t_continue() along with the presence and RLS APIs.  So what I have is the following:

#!substdef "!PRESENCE_PROCESS_SLEEP!100000!g"
modparam("mqueue", "mqueue", "name=presence")
modparam("rtimer", "timer", "name=presenceMaster;interval=PRESENCE_PROCESS_SLEEPu;mode=1;")
modparam("rtimer", "exec", "timer=presenceMaster;route=PRESENCE_MASTER_PROCESS")
modparam("mqueue", "mqueue", "name=presenceWorker0")
modparam("rtimer", "timer", "name=presenceWorker0;interval=1u;mode=1;")
modparam("rtimer", "exec", "timer=presenceWorker0;route=PRESENCE_WORKER_PROCESS")
modparam("mqueue", "mqueue", "name=presenceWorker1")
modparam("rtimer", "timer", "name=presenceWorker1;interval=1u;mode=1;")
modparam("rtimer", "exec", "timer=presenceWorker1;route=PRESENCE_WORKER_PROCESS")
...
modparam("mqueue", "mqueue", "name=presenceWorkern")
modparam("rtimer", "timer", "name=presenceWorkern;interval=1u;mode=1;")
modparam("rtimer", "exec", "timer=presenceWorkern;route=PRESENCE_WORKER_PROCESS")
...
route {
	...
	# Some logic to determine this is a presence request (within or without dialog)
	$var(queue) = "presence";
	route(PRESENCE_ENQUEUE);
	...
}
...
route[PRESENCE_ENQUEUE] {
        xlog("L_WARN", "$rm: route[PRESENCE_ENQUEUE]\n");

        if (!t_suspend()) {
                t_reply("500", "Server Internal Error");
                xlog("L_ERR", "Failed to suspend transaction for $rm\n");
                exit;
        }

        xlog("L_INFO", "Suspended transaction for $rm [$T(id_index):$T(id_label)]\n");

        if (!mq_add("$var(queue)", "$T(id_index)", "$T(id_label)")) {
                t_reply("500", "Server Internal Error");
                xlog("L_ERR", "Failed to queue transaction for $rm [$T(id_index):$T(id_label)] on $var(queue)\n");
                exit;
        }

        exit;
}

route[PRESENCE_MASTER_PROCESS] {
        xlog("L_INFO", "Running PRESENCE_MASTER_PROCESS\n");

        while (mq_fetch("presence")) {
                $var(id_index) = (int) $mqk(presence);
                $var(id_label) = (int) $mqv(presence);
                xlog("L_INFO", "presence: Found queued transaction [$var(id_index):$var(id_label)]\n");
                t_continue("$var(id_index)", "$var(id_label)", "PRESENCE_DISTRIBUTE");
        }
}

route[PRESENCE_DISTRIBUTE] {
        xlog("L_WARN", "$rm: route[PRESENCE_DISTRIBUTE]\n");

        # Some algorithm to distribute traffic across queues...
	# Perhaps on request type (so we have a NOTIFIER that
	# sends NOTIFY requests in order), perhaps on some form
	# of hash...  I am experimenting with this...
        $var(queue) = "presenceWorker" + $var(queue_number);
        xlog("L_INFO", "Adding to queue: $var(queue)\n");
        route(PRESENCE_ENQUEUE);
}

route[PRESENCE_WORKER_PROCESS] {
        lock("pres");
        $var(pres) = $shv(pres);
        $shv(pres) = $shv(pres) + 1;
        unlock("pres");

        $var(queue) = "presenceWorker" + $var(pres);
        xlog("L_WARN", "Starting process: $var(queue) (pid: $pp)\n");

        while (1) {
                while (mq_fetch($var(queue))) {
                        $var(id_index) = (int) $mqk($var(queue));
                        $var(id_label) = (int) $mqv($var(queue));
                        xlog("L_WARN", "$var(queue): found queued transaction [$var(id_index):$var(id_label)]\n");
                        t_continue("$var(id_index)", "$var(id_label)", "PRESENCE");
                }
                usleep(PRESENCE_PROCESS_SLEEP);
        }
}

route[PRESENCE] {
        xlog("L_WARN", "$rm: route[PRESENCE]: $var(queue)\n");

        if (is_method("NOTIFY")) {
                xlog("L_INFO", "Sending NOTIFY to RLS\n");
                rls_handle_notify();
        } else if (is_method("PUBLISH")) {
                xlog("L_INFO", "Sending PUBLISH to Presence\n");
                handle_publish();
        } else if (is_method("SUBSCRIBE")) {
                xlog("L_INFO", "Sending SUBSCRIBE to RLS\n");
                $var(ret_code) = rls_handle_subscribe();
                if ($var(ret_code) == 10) {
                        xlog("L_INFO", "  SUBSCRIBE not for RLS - sending to Presence\n");
                        handle_subscribe();
                }
        } else {
                xlog("L_ERR", "Received non-(NOTIFY|PUBLISH|SUBSCRIBE) request from presence queue\n");
                t_reply("500", "Server Internal Error");
        }
	exit;
}
...

Previously, I had just a single queue "presence" which all of the presence worker process took requests from.  This meant that t_suspend()/t_continue() was used just once and this worked (the presence/RLS APIs respond to the requests statefully).  The reason for doing this in the first place is that I was getting problems with back-end RLS traffic all using a single TCP connection, which meant all the back-end presence requests were being handled by the same Kamailio process, which caused a bottleneck (using UDP causes a different set of problems under load and isn't really an option).  Although the queue is a FIFO the fact that different processes could take different amounts of time means that things were happening out of order (Klaus and Anca have had a discussion about just this kind of issue with presence on the mailing list recently) and this is causing me problems.

What I have now (above) is presence requests being pulled from the TCP buffer and suspended as quickly as possible.  A presenceMaster process then dequeues the request (continues it), performs some analysis to determine which worker should deal with it, and then suspends it again queuing it for the right worker.  All of this works up until the t_continue() for the worker (in the PRESENCE_WORKER_PROCESS) is called.  At this point the transaction is killed.

What I can't understand is why the first t_suspend()/t_continue() works here, but the second fails.  My previous version of this (with the single queue and single t_suspend()/t_continue() call) worked fine, but it seems that the sequence of t_suspend(), t_continue(), t_suspend(), t_continue() - with no changes to or handling of the request in-between - fails.

Thanks,

Peter


On Wed, 2012-03-28 at 15:13 +0200, Miklos Tirpak wrote:
Peter,

t_suspend() and t_continue() should work multiple times as long as they 
are executed sequentially after each other, i.e. there cannot be two 
branches suspended at the same time.

The error you get means to me that t_continue() executed the specified 
route block, but in that route, the request was neither replied nor a 
new branch was added. Hence, the transaction is hanging in memory and 
the module sees no pending branch that could return a reply later.

Make sure that in the route block executed by t_continue() there is 
either a t_reply() or you append a new branch and forward it with 
t_relay() (or append a new branch and call t_suspend() again). I think 
you also need to handle the failure of t_relay() and explicitly call 
t_reply() when t_relay() fails in this route.

Regards,
Miklos

On 03/28/2012 02:21 PM, Daniel-Constantin Mierla wrote:
> Hello,
>
> I have been using it only once and didn't looked much deeper into the code.
>
> Maybe Miklos (cc-ed) can give faster more details, afaik he is the
> developer of that piece.
>
> Cheers,
> Daniel
>
> On 3/28/12 1:13 PM, Peter Dunkley wrote:
>> Hi,
>>
>> I am trying to use t_suspend()/t_continue() multiple times on the same
>> transaction. Calling t_suspend() more than once works, but the second
>> time I call t_continue() the transaction is killed and a 500 response
>> is sent. It is the "if (branch == t->nr_of_outgoings)" check from the
>> code fragment below (from t_suspend.c:t_continue()) that results in
>> the transaction being killed - you can see the debug/error line I
>> added to determine this in the fragment.
>>
>> Is using t_suspend()/t_continue() multiple times something that should
>> work?
>>
>> Thanks,
>>
>> Peter
>>
>> if (t->uas.status < 200) {
>> /* No final reply has been sent yet.
>> * Check whether or not there is any pending branch.
>> */
>> for ( branch = 0;
>> branch < t->nr_of_outgoings;
>> branch++
>> ) {
>> if ((t->uac[branch].request.buffer != NULL)
>> && (t->uac[branch].last_received < 200)
>> )
>> break;
>> }
>>
>> if (branch == t->nr_of_outgoings) {
>> /* There is not any open branch so there is
>> * no chance that a final response will be received. */
>> ret = 0;
>> LM_ERR("branch == t->nr_of_outgoings\n");
>> goto kill_trans;
>> }
>> }
>>
>> --
>> Peter Dunkley
>> Technical Director
>> Crocodile RCS Ltd
>>
>>
>>
>> _______________________________________________
>> sr-dev mailing list
>> sr-dev@lists.sip-router.org
>> http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-dev
>
> --
> Daniel-Constantin Mierla
> Kamailio Advanced Training, April 23-26, 2012, Berlin, Germany
> http://www.asipto.com/index.php/kamailio-advanced-training/
>

_______________________________________________
sr-dev mailing list
sr-dev@lists.sip-router.org
http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-dev

_______________________________________________
sr-dev mailing list
sr-dev@lists.sip-router.org
http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-dev

-- 
Peter Dunkley
Technical Director
Crocodile RCS Ltd