[ovs-dev] [PATCH ovn v2 6/6] ovn-northd-ddlog: New implementation of ovn-northd based on ddlog.

[Dumitru Ceara]

On 11/6/20 5:59 PM, Ben Pfaff wrote:
> On Fri, Nov 06, 2020 at 05:25:36PM +0100, Dumitru Ceara wrote:
>> On 11/6/20 4:18 AM, Ben Pfaff wrote:
>>> Some of these are from ovn-northd, not ovn-northd-ddlog, and so I don't
>>> think it's likely that my patch series causes them, since it doesn't
>>> really touch ovn-northd.  The OVN testsuite has a regrettable number of
>>> race conditions in it.
>>
>> I agree, there are probably races in the testsuite but I think there's
>> also a bug with "ovn-nbctl --wait=hv sync".
>>
>> For example, adding a "sleep 1" here:
>> https://github.com/ovn-org/ovn/blob/c108f23e1c10910031f9409b79001d001aae0c8f/tests/ovn.at#L21478
>>
>> makes this test pass on my machine with both ovn-northd and
>> ovn-northd-ddlog.
> 
> This hints toward a bug in ovn-controller.  It suggests that
> ovn-controller reports that it is caught up before it has pushed all the
> flows to ovs-vswitchd.
> 

Right, I tracked it down to ovn-controller setting nb_cfg in the SB
chassis record while there's still an unreplied monitor_cond_change from SB.

If some logical_flows were added to the SB at nb_cfg == X.
And if at nb_cfg == X+1 some changes happen to the SB that would also
make ovn-controller request a monitor condition change that includes
flows added at X then ovn-controller "acks" nb_cfg == X too early.

I think it might be enough to just delay reporting that ovn-controller
caught up if there are still in flight monitor_cond_change requests.
I'll see if I can come up with a fix and if there are other corner cases.

Regards,
Dumitru