30 Jan
2013
30 Jan
'13
1:37 a.m.
OK thanks Peter for the clarity. I'll revert to non-notifier for the time
being.
I will have some interest in multi-server soon so I'll take a look at the
PostgreSQL then.
Cheers
Shane
On Wed, Jan 30, 2013 at 12:20 PM, Peter Dunkley <
peter.dunkley(a)crocodile-rcs.com> wrote:
In DB mode 1 a completely different code path -
without notifier processes
- is run.
As I mentioned you only need the notifier processes for resilient
multi-server systems. For single server the default presence behaviour
without notifier is fine - and multi-server resilient presence will not
work in 3.3 anyway (you need git master with PostgreSQL DB for this).
Regards,
Peter
Shane Harrison <shane.harrison(a)paragon.co.nz> wrote:
Thanks Peter. Note I have repeated your answer below in this correctly
titled thread since I accidently hijacked a thread with my original post.
I take your point on race conditions. But this doesn't feel like a race
condition. It is pretty predictable. Basically if the "updated" column is
set to an odd number then process_dialogs() never processes that database
notify entry.
This only happens in db-mode 2 (database only). In db-mode 1 the notifies
are handled immediately and we never see this issue. At least that is my
conclusion from observations.
Cheers
Shane
Peter Dunkley wrote
--------------------------------------------
Hello,
The polled notification stuff in 3.3.3 has some issues. There are some
database related race-hazards that mean you could have problems. These have
been resolved by adding new features to the database library - but as these
are new features they are only present in git master at the moment. The
presence modules in git master have been updated to use these new features.
It is also worth noting that the PostgreSQL database driver in git master
is significantly more advanced, in terms of features, than the other
database drivers. The presence notifier stuff has only been tested with
PostgreSQL and may well be totally dependent on features that are only in
the driver for that database (it should still run without crashing when
using other databases but will probably not function correctly). When you
do your retest please make sure that you use PostgreSQL - if you must use
another database then you will need to update the Kamailio driver for that
database to include support for the following new APIs:
- init2
- start_transaction
- end_transaction
- abort_transaction
- query_lock
Please retry with git master and PostgreSQL and let me know if the problem
persists.
Regards,
Peter
----------------------------------------------------
On Wed, Jan 30, 2013 at 9:17 AM, Shane Harrison <
shane.harrison(a)paragon.co.nz> wrote:
--
Imagination NZ Ltd
Level 6
92 Queens Drive
P0 Box 30449
Lower Hutt 5040
+64 4 5703870 Extn 875
+64 21 608919 (mobile)
Hi all,
I have a situation where subscriptions do not get notified and have
tracked it down to a problem with the polled notify processing. Can you
advise if this is a bug or correct my understanding of the code.
I am using kamailio 3.3.3 and have 1 notify process. I have enhanced
the support for the ua-profile event (rfc 6080) , although I don't believe
I have made any changes that directly impact the problem I am seeing.
When a new subscription arrives it is added to the active_watchers table,
the 'updated' column is assigned a number in the range 0 - N-1, where N is
effectively the total number of polled update tasks that are called in a
round-robin fashion, distributed across the notify processes. In this case
updated = 7.
In subscribe.c:
int update_subscription_notifier(struct sip_msg* msg, subs_t* subs, int
to_tag_gen, int* sent_reply)
{
...
/* Set the notifier/update fields for the subscription */
subs->updated = core_hash(&subs->callid, &subs->from_tag, 0) %
(pres_waitn_time * pres_notifier_poll_rate
* pres_notifier_processes);
The notify process periodically calls pres_timer_send_notify(), which
calculates the round (the update task number) and does the notify update by
checking the active_watchers table for entries with updated = round. The
update is done twice, first for the event then for event.winfo.
In notify.c:
void pres_timer_send_notify(unsigned int ticks, void *param)
{
int process_num = *((int *) param);
int round = subset + (pres_waitn_time * pres_notifier_poll_rate
* process_num);
if (process_dialogs(round, 0) < 0)
{
LM_ERR("Handling non presence.winfo dialogs\n");
return;
}
if (process_dialogs(round, 1) < 0)
{
LM_ERR("Handling presence.winfo dialogs\n");
return;
}
}
In this instance process_num = 0, so round = subset. However subset is
incremented in process_dialogs() in notify.c:
if (++subset > (pres_waitn_time * pres_notifier_poll_rate) -1)
subset = 0;
This means that round is incremented twice between calls to
process_dialogs(round, 0), in my case round is always even, hence not
detecting the subscription with updated = 7.
It seems that the subset increment should be done in
pres_timer_send_notify() rather than in process_dialogs(). Is this correct?
Additionally, is there a need for the second call that only handles
presence.winfo subscriptions? The code could be simplified by only making
one call and processing all subscriptions for the round.
Kind regards
Shane Harrison
--
Imagination NZ Ltd
Level 6
92 Queens Drive
P0 Box 30449
Lower Hutt 5040
+64 4 5703870 Extn 875
+64 21 608919 (mobile)
--
Imagination NZ Ltd
Level 6
92 Queens Drive
P0 Box 30449
Lower Hutt 5040
+64 4 5703870 Extn 875
+64 21 608919 (mobile)
_______________________________________________
sr-dev mailing list
sr-dev(a)lists.sip-router.org
http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-dev