Hi, Ovidiu!

I agree, the fact that no one reported it before is a bit interesting. On the other hand, not sure if >1000 buffers is expected, perhaps the flags you are using for rtpengine are increasing the number of buffers.

Best regards,

Răzvan Crainea
OpenSIPS Core Developer
http://www.opensips-solutions.com

On 6/29/20 9:43 PM, Ovidiu Sas wrote:
Hi Razvan,

I have updated the rtpproxy module and everything is backported down to 2.4.

Detecting IOV_MAX would be kind of an overkill ...
The error logs were improved and the error is now searchable in the
mailing list. If everyone stumble across it, it should be easy to
troubleshoot.
It seems that IOV_MAX will be properly defined in the future (it is
defined in some experimental headers) and the detection of it will
become obsolete.
We could move the definition of OSIP_IOV_MAX into a common header file
for both rtpengine and rtpproxy module.

I'm still puzzled as to why this was not reported before for the
rtpengine module. It is quite easy to reproduce with very low traffic.

Thanks,
Ovidiu

On Mon, Jun 29, 2020 at 7:20 AM Răzvan Crainea <raz...@opensips.org> wrote:

Hi, Ovidiu!

I agree we should match the two modules to use a common logic.
I believe we should backport this in 2.4 as well, since it doesn't work
for all scenarios, therefore it is considered a bug.

I somehow doubt this should be a module's param, because it is only
needed in certain circumstances, where IOV_MAX is not "visible" in the
code. IMO, a decent, commonly accepted limit should be enough - so I
believe the way you've done it for 3.0 and 3.1 is more than fine. Adding
a parameter will only complicate provisioning and the gain of setting a
custom value would be hard to quantify.

Probably a better approach would be to make sure the IOV_MAX is properly
defined using something described here[1]. What do you think?

[1]
https://stackoverflow.com/questions/27271801/c-the-permitted-maximum-of-the-iovcnt-argument-in-writev

Cheers,

Răzvan Crainea
OpenSIPS Core Developer
http://www.opensips-solutions.com

On 6/27/20 9:35 PM, Ovidiu Sas wrote:
Fix implemented on trunk, 3.1 and 3.0.
We should update the fix on rtpproxy to match rtpengine and move the
OSIP_IOV_MAX define somewhere upper in the code tree to be visible by
both rtpproxy and rtpengine modules.
For trunk, maybe we should add a new param.

Should we backport this fix to 2.4?

Thanks,
Ovidiu

On Fri, Jun 26, 2020 at 9:35 AM Ovidiu Sas <o...@voipembedded.com> wrote:

Hello Razvan,

On ubuntu we have a failure for more then 1024 buffers. There is something 
defined in some experimental headers (as __IOV_MAX), but the IOV_MAX define is 
not visible in the rtpengine module.

And yes, there is an extra buffer. I guess the length of that extra buffer was 
zero and that’s why it didn’t create issues. But as soon as I concatenated the 
buffer, the issue showed up and I had an extra trailing string at the end of 
the rtpengine command that was sent on the wire.
After reducing  the number of buffers by one, all was good on both scenarios: 
sending with and without concatenated buffers.
I will prepare a fix for it and push it.

-ovidiu

On Fri, Jun 26, 2020 at 03:55 Răzvan Crainea <raz...@opensips.org> wrote:

Hi, Ovidiu!

So you're saying that the IOV_MAX is not explicitly defined, but it does
fail after 1024 buffers, correct? If so, perhaps we should limit the
number of buffers to 1024, if not already defined.
I did notice the extra vcnt as well, but I though that was related to
the fact that it allocates one extra iovec (at the headr) for the
cookie, and sometime uses it, sometime doesn't. Nevertheless, it is
consistent on both cases, UNIX & UDP, both have that vcnt + 1. Now I
haven't checked whether this is correct or not, could you please confirm?

Best regards,

Răzvan Crainea
OpenSIPS Core Developer
http://www.opensips-solutions.com

On 6/26/20 2:29 AM, Ovidiu Sas wrote:
It looks like the hidden IOV_MAX is set to 1024 in debian.
I tested a patch and all looks good.
It seems that there is an error in the code: writev has the wrong
number of iovcnt (should be one less).
I tested and all looks ok (the code in rtpproxy.c has the proper iovcnt).

-ovidiu

On Thu, Jun 25, 2020 at 1:03 PM Ovidiu Sas <o...@voipembedded.com> wrote:

We will need to add a param to control the max number of buffers.

-ovidiu

On Thu, Jun 25, 2020 at 1:02 PM Ovidiu Sas <o...@voipembedded.com> wrote:

It seems that when we have more than roughly 1000 buffers, the send fails.

-ovidiu

On Wed, Jun 24, 2020 at 8:54 AM Ovidiu Sas <o...@voipembedded.com> wrote:

Hello Razvan,

The system is a debian buster one.
I patched the code:
#ifdef IOV_MAX
       LM_NOTICE("IOV_MAX=[%d]\n", IOV_MAX);
#else
       LM_NOTICE("no IOV_MAX\n");
#endif

and I get:
NOTICE:rtpengine:send_rtpe_command: no IOV_MAX
in the logs.

Then I patched the code again to check how many buffers are being used:
The max so far was 73:
LM_NOTICE("writev(rtpe_socks[node->idx], v , %d)\n", vcnt + 1);
got me:
NOTICE:rtpengine:send_rtpe_command: writev(rtpe_socks[node->idx], v , 73)

I will continue to monitor the system to see if there is a correlation
between the error and the number of buffers.

Thanks,
Ovidiu

On Wed, Jun 24, 2020 at 1:59 AM Răzvan Crainea <raz...@opensips.org> wrote:

Hi, Ovidiu!

I doubt this is a problem of OpenSIPS version, but rather of the OS you
are running on. I suspect that error comes from the fact that the bson
resulted has more than IOV_MAX elements, which if I recall correctly it
was 15 on some OSes.
We had a similar problem in rtpproy[1], where we had merged the buffers
in a single one just to pass over this limitation. Could you check if
you are facing a similar issue?

[1]
https://github.com/OpenSIPS/opensips/blob/master/modules/rtpproxy/rtpproxy.c#L2031

Răzvan Crainea
OpenSIPS Core Developer
http://www.opensips-solutions.com

On 6/24/20 7:32 AM, Ovidiu Sas wrote:
This is happening also on the latest 3.0.
The weird thing is that opensips doesn't send anything to rtpengine.
The first opensips/rtpengine exchange on the initial INVITE works ok,
but the opensips/rtpengine exchange on the 200ok fails (no command is
sent by opensips - confirmed by running ngrep on the loopback
interface).
On the next call, the initial offer works fine, but the answer fails
due to no command issued by opensips.

version: opensips 3.0.2 (x86_64/linux)
flags: STATS: On, DISABLE_NAGLE, USE_MCAST, SHM_MMAP, PKG_MALLOC,
Q_MALLOC, F_MALLOC, HP_MALLOC, DBG_MALLOC, FAST_LOCK-ADAPTIVE_WAIT
ADAPTIVE_WAIT_LOOPS=1024, MAX_RECV_BUFFER_SIZE 262144, MAX_LISTEN 16,
MAX_URI_SIZE 1024, BUF_SIZE 65535
poll method support: poll, epoll, sigio_rt, select.
git revision: 3a8f6f137
main.c compiled on 09:36:42 Jun 22 2020 with gcc 9

-ovidiu


On Wed, Jun 10, 2020 at 2:49 PM Ovidiu Sas <o...@voipembedded.com> wrote:

Hello all,

I'm running opensips 3.1.0-beta (latest version) and experiencing
connectivity issues to the rtpengine daemon running on the same host:
ERROR:rtpengine:send_rtpe_command: can't send command to a RTP proxy
(22:Invalid argument)
ERROR:rtpengine:send_rtpe_command: timeout waiting reply from a RTP proxy
ERROR:rtpengine:send_rtpe_command: proxy <udp:127.0.0.1:2223> does not
respond, disable it
ERROR:rtpengine:rtpe_function_call: no available proxies

After an opensips restart, everything comes to normal.
I was running previously opensips 3.1.0-dev and everything was working fine.

The issue starts showing up after a few days with very little traffic.
Is there anyone experiencing this issue?


Thanks,
Ovidiu

--
VoIP Embedded, Inc.
http://www.voipembedded.com




_______________________________________________
Users mailing list
Users@lists.opensips.org
http://lists.opensips.org/cgi-bin/mailman/listinfo/users



--
VoIP Embedded, Inc.
http://www.voipembedded.com



--
VoIP Embedded, Inc.
http://www.voipembedded.com



--
VoIP Embedded, Inc.
http://www.voipembedded.com




_______________________________________________
Users mailing list
Users@lists.opensips.org
http://lists.opensips.org/cgi-bin/mailman/listinfo/users

--
VoIP Embedded, Inc.
http://www.voipembedded.com




_______________________________________________
Users mailing list
Users@lists.opensips.org
http://lists.opensips.org/cgi-bin/mailman/listinfo/users




_______________________________________________
Users mailing list
Users@lists.opensips.org
http://lists.opensips.org/cgi-bin/mailman/listinfo/users

Reply via email to