CPU Overload with a TCP or MQTT connection/disconnection session.

[ralight]

migrated from Bugzilla #485143
status RESOLVED severity normal in component Mosquitto for 1.4
Reported in version 1.4 on platform PC
Assigned to: Roger Light

Original attachment names and IDs:

• TestMosquitto.zip (ID 259258)
• mosquitto.patch (ID 259275)

On 2016-01-04 09:11:22 -0500, Pierre-Yves BOISBUNON wrote:

On windows 7 / 10 /Server 2012 environment,

• A TCP client connection followed by a disconnection increases CPU loads for a 90s period. This behavior seems to be present in all versions.
• A MQTT client connection followed by a disconnection increases CPU loads for keepalive time duration x 1.5. This behavior seems to be present from version 1.4.x.

Any idea?

On 2016-01-04 11:09:00 -0500, Roger Light wrote:

What do you mean by a TCP connect/disconnect? A BSD socket connect() followed by close()? Or does e.g. "mosquitto_pub -t foo -m bar" produce this problem for you?

On 2016-01-05 04:35:18 -0500, Pierre-Yves BOISBUNON wrote:

Here is our test sequence:

mosquitto_pub.exe -t 0/rc/10 -m bar -p 5500

1451977925: New connection from ::1 on port 5500.

1451977925: New client connected from ::1 as mosqpub/3516-Pascal_Cha (c1, k60).

1451977925: Sending CONNACK to mosqpub/3516-Pascal_Cha (0, 0)

1451977925: Received PUBLISH from mosqpub/3516-Pascal_Cha (d0, q0, r0, m0, '0/rc/10', ... (3 bytes))

1451977925: Received DISCONNECT from mosqpub/3516-Pascal_Cha

1451977925: Client mosqpub/3516-Pascal_Cha disconnected.

Test OK

Connect Mqtt Client (keepalive : 10s)

1451979015: New connection from 127.0.0.1 on port 5500.

1451979015: New client connected from 127.0.0.1 as 1 (c0, k10, u'1').

1451979015: Sending CONNACK to 1 (0, 0)

1451979015: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

1451979020: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

1451979025: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

1451979030: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

Disconnect

1451979035: Received DISCONNECT from 1

1451979035: Client 1 disconnected.

Test OK

Connect Mqtt Client (keepalive : 10s)

1451978205: New connection from 127.0.0.1 on port 5500.

1451978205: New client connected from 127.0.0.1 as 1 (c0, k10, u'1').

1451978205: Sending CONNACK to 1 (0, 0)

1451978205: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

1451978210: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

1451978215: Received PUBLISH from 1 (d0, q0, r0, m0, '0/rc/1', ... (5 bytes))

Kill Mqtt Client

100% CPU 15s

1451978230: Client 1 has exceeded timeout, disconnecting.

1451978230: Socket error on client 1, disconnecting.

Connect Socket TCP

1451977972: New connection from 127.0.0.1 on port 5500.

Close Socket TCP

100% CPU 90s :

1451978061: Client has exceeded timeout, disconnecting.

1451978061: Socket error on client , disconnecting.

On 2016-01-06 11:38:48 -0500, Roger Light wrote:

Thanks, I'll take a look.

On 2016-01-14 08:49:23 -0500, Pierre-Yves BOISBUNON wrote:

Any update on this issue?

On 2016-01-14 16:21:27 -0500, Roger Light wrote:

Yes, sorry for the delay.

I believe this commit fixes the problem:

http:https://git.eclipse.org/c/mosquitto/org.eclipse.mosquitto.git/commit/?h=fixes&id=SHA: 00491da

Please reopen the bug if you disagree.

On 2016-01-15 05:49:27 -0500, Pierre-Yves BOISBUNON wrote:

Dear Roger

Thank you very much for you fix.

Could you please help to generate a native windows installer?
We do not found any trouble with cygwin based exe.

Thanks,

On 2016-01-15 07:33:09 -0500, Roger Light wrote:

This is set up to work easily with appveyor, so I've just pushed the changes to make it happen.

https://ci.appveyor.com/project/ralight/mosquitto/build/artifacts

The installer there is called 1.4.7, but it actually includes this fix.

On 2016-01-15 08:02:51 -0500, Pierre-Yves BOISBUNON wrote:

Thanks Roger for the installer.
I'm sorry but the issue is still there.
We do not see any change.

On 2016-01-15 10:14:41 -0500, Roger Light wrote:

In that case I can no longer duplicate the bug. What version of Windows are you using? I am on Windows 7.

It might be worth trying to compile yourself. You should use CMake to generate Visual Studio projects. The openssl files you need are at https://slproweb.com/products/Win32OpenSSL.html "Win32 OpenSSL v1.0.2e". You can disable pthreads support because it is only the broker we are worried about for this bug.

On 2016-01-19 08:17:00 -0500, Pierre-Yves BOISBUNON wrote:

Created attachment 259258
VS TCP Test Project

Dear Roger

The problem still exists on Windows 7.
I enclose you the TCP test project related to the TCP part problem of this thread.

Hope it helps.

Thanks.

On 2016-01-19 10:36:06 -0500, Roger Light wrote:

Thanks, I'll take a look on Thursday hopefully.

On 2016-01-20 10:08:40 -0500, Wilfried Chauveau wrote:

Created attachment 259275
add POLLHUP flag to disconnection handling code.

On 2016-01-20 10:10:07 -0500, Wilfried Chauveau wrote:

Hello,

I fixed the issue by adding the flag POLLHUP to (POLLERR | POLLNVAR).

Best regards,
Wilfried Chauveau.

On 2016-01-28 17:09:40 -0500, Roger Light wrote:

Thanks both. I forgot to update this bug after I found the POLLHUP problem myself and pushed an update. I believe it is definitely fixed this time - see the fixes branch again.