[AAA-28] Race condition between AAA and netconf connector

Thu, 21 Mar 2019 11:56:40 +0000

From Devin Avery

Hi Folks,

Following up on this. So the problem appears to be an internal race condition inside of MD-SAL / AAA / config system combined with logic flaws about retrying connections.

Basically, the way I get this to reproduce is Iist all of the features that I want to install in the featuresBoot list and then start a clean controller (I.e. delete data directory, delete /etc/opendaylight). When the controller starts up there is a race condition from the config subsystem injecting the self-referencing netconf connection and the AAA database being locked. When the problem occurs the AAA database file is locked and thus when netconf tries to connect and specifically authenticate the AAA fails because of an exception. Then because the config subsystem, or more specifically the netconf mounter is a try once and done (no retries) it fails to connect the self-referencing controller.

Potential fixes from what I can tell:

enhance AAA logic to handle issues when the database might be locked, and consider a wait retry mechanism (we would need to answer the question of “when would we expect this error” to implement something like this).
and/or enhance the netconf mounting logic to retry even if it got a failed authentication (though that seems a little strange)

I think ultimately #1 is the longer term fix. If you are trying to reproduce this note that I see this every once in a while - on 10 machines, I see it consistently on 1 of them every time I run a suite of tests.

For now I have worked around this by installing the features more slowly / independently, but this is clearly not an ideal solution. I’ll see if I can’t come up with a more consistent reproduction scenario but figured I would share what I know in the mean time.

OpenDaylight JIRA

[AAA-28] Race condition between AAA and netconf connector