Hi Ben,
I checked the error trace and it should not leave any dangling lock (due
mishandled error). Before disabling HEP, try to disable the async
support for HEP.
If you claim that the same 100% CPU happens with HEP + UDP, send me a
trap for that too, as in the previous case, the deadlock was exclusively
HEP + TCP related.
Anyhow, as the original trap showed a deadlock, next step will be to
recompile with the DBG_LOCK option - this enables extra code to
debug/troubleshoot locking related issues - are you able to do it?
Regards,
Bogdan-Andrei Iancu
OpenSIPS Founder and Developer
http://www.opensips-solutions.com
OpenSIPS Bootcamp 2018
http://opensips.org/training/OpenSIPS_Bootcamp_2018/
On 10/26/2018 04:14 PM, Ben Newlin wrote:
Bogdan,
Actually, yes we do. Looking back I can see these errors just before
the issue occurs:
Oct 24 19:00:36 [5700] ERROR:proto_hep:send_hep_message: Cannot send
hep message!
Oct 24 19:00:36 [5700] ERROR:proto_hep:msg_send: send() to
10.32.163.211:9061 for proto hep_tcp/9 failed
Oct 24 19:00:36 [5700] ERROR:proto_hep:hep_tcp_send: failed to send
Oct 24 19:00:36 [5700] ERROR:proto_hep:async_tsend_stream: Failed
first TCP async send : (32) Broken pipe
I will try disabling HEP and see if we can reproduce.
Just for information, I have been reproducing the issue in our testing
environment which uses TCP for HEP, however the issue is occurring in
our production environment as well which is still using UDP for HEP.
Ben Newlin
*From: *Bogdan-Andrei Iancu <[email protected]>
*Date: *Friday, October 26, 2018 at 3:06 AM
*To: *Ben Newlin <[email protected]>, OpenSIPS users mailling
list <[email protected]>
*Subject: *Re: [OpenSIPS-Users] CPU 100% with TCP
Hi Ben,
Thank you for the info.
It looks like the processes get stuck into a HEP related internal lock
- do you see any HEP related errors in your logs, prior to the dead-lock ?
Also, as PoC, could you disabled HEP tracing to see if the problem
goes away ?
Thanks,
Bogdan-Andrei Iancu
OpenSIPS Founder and Developer
http://www.opensips-solutions.com
OpenSIPS Bootcamp 2018
http://opensips.org/training/OpenSIPS_Bootcamp_2018/
On 10/24/2018 10:18 PM, Ben Newlin wrote:
Bogdan,
I have run the command but the output was too large for pastebin
so I have sent it to you directly.
Ben Newlin
*From: *Bogdan-Andrei Iancu <[email protected]>
<mailto:[email protected]>
*Date: *Wednesday, October 24, 2018 at 5:17 AM
*To: *OpenSIPS users mailling list <[email protected]>
<mailto:[email protected]>, Ben Newlin
<[email protected]> <mailto:[email protected]>
*Subject: *Re: [OpenSIPS-Users] CPU 100% with TCP
Hi Ben,
Could you run "opensipsctl trap" ?
Regards,
Bogdan-Andrei Iancu
OpenSIPS Founder and Developer
http://www.opensips-solutions.com
OpenSIPS Bootcamp 2018
http://opensips.org/training/OpenSIPS_Bootcamp_2018/
On 10/24/2018 12:56 AM, Ben Newlin wrote:
Hi,
We have implemented TCP recently and are performing TCP<->UDP
translation on one of our proxy types. This proxy only exists
for that purpose; there are no DB queries, REST calls, or
anything like that. It is designed to be very fast and high
throughput.
Recently we have found that when the remote endpoint of a TCP
connection is lost, i.e. the server goes down, while under
moderate load OpenSIPS quickly reaches 100% CPU and becomes
unresponsive. When this occurs, the “top” command shows that
between 30-90% CPU is in System (kernel) space, and each
OpenSIPS TCP process shows many times the normal CPU. We are
running OpenSIPS 2.4.2 on Amazon Linux.
I obtained as much information as I could using ps, strace,
and gdb here: https://pastebin.com/JP3DnCqs
<https://pastebin.com/JP3DnCqs>. We can reproduce the failure
consistently by removing a server during call traffic.
A few things I noticed:
* The number of running threads reported by OpenSIPS doesn’t
align with our configuration, copied here:
####### Global Parameters #########
children=32
#// Allow 503 to pass back to Control
disable_503_translation=yes
#// Even though we are not receiving HEP,
#// this listener is required by OpenSIPS
#// in order to use the proto_hep module. :/
listen=hep_tcp:10.32.40.245:9061 use_children 1
#// Configure the listeners
listen=udp:10.32.40.245:5060 as XXX.XXX.XXX.XXX
listen=tcp:10.32.40.245:5060 as XXX.XXX.XXX.XXX
#// Transaction Module
loadmodule "tm.so"
modparam("tm", "restart_fr_on_each_reply", 0)
modparam("tm", "timer_partitions", 8)
modparam("tm", "onreply_avp_mode", 1)
modparam("tm", "wt_timer", 10)
According to the documentation if “tcp_children” is not set
then the value of “children” will be used [1], but we have set
“children” to 32 and only have the default 8 TCP processes.
Also we appear to only have 1 timer process, although we have
set the number of timer partitions to 8.
* The server that is terminated was using TCP connections
exclusively, but all of the CPU seems to be in the UDP
threads. The one I looked at appeared to be handling a
CANCEL to one of the calls that was active and was
attempting to send it out via TCP. I’m not sure why it
would be trying to relay the CANCEL as no 100 Trying had
been received from the server. I have noticed that in 2.x
OpenSIPS will now send CANCELs for transactions even when
100 Trying was not received. Is that intentional? RFC 3261
states that no CANCEL should be sent unless a provisional
response has been received.
Any assistance with this would be appreciated.
[1] -
http://www.opensips.org/Documentation/Script-CoreParameters-2-4#toc66
Ben Newlin
_______________________________________________
Users mailing list
[email protected] <mailto:[email protected]>
http://lists.opensips.org/cgi-bin/mailman/listinfo/users
_______________________________________________
Users mailing list
[email protected]
http://lists.opensips.org/cgi-bin/mailman/listinfo/users
