[ovs-dev] [PATCH 0/9] Translation fixes for revalidation

[Jarno Rajahalme]

IMO this series should be backported to 2.4 and 2.3 as well, where applicable.

  Jarno


> On Dec 10, 2015, at 18:17, Daniele Di Proietto <diproiettod at vmware.com> wrote:
> 
> 
> 
>> On 10/12/2015 16:27, "Jesse Gross" <jesse at kernel.org> wrote:
>> 
>>> On Thu, Dec 10, 2015 at 4:11 PM, Jarno Rajahalme <jarno at ovn.org> wrote:
>>> 
>>>> On Dec 10, 2015, at 3:20 PM, Jesse Gross <jesse at kernel.org> wrote:
>>>> 
>>>> On Wed, Dec 9, 2015 at 6:27 PM, Daniele Di Proietto
>>>> <diproiettod at vmware.com> wrote:
>>>>> Sometimes the ofproto layer creates a flow which is not liked by the
>>>>> revalidation for various reasons.  This behavior, while not critical
>>>>> might impact the performance.  This series aims to fix a lot of these
>>>>> bugs.
>>>>> 
>>>>> The detection has been done by modifying OVS to revalidate a flow as
>>>>> soon as it is installed (this is not included in the series, I'd be
>>>>> happy to discuss strategies to merge something like that upstream).
>>>>> If the revalidation complains there's a bug. This series fixes all the
>>>>> bugs found in the testsuite.
>>>>> 
>>>>> The first commits are trivial fixes to various components in OVS. The
>>>>> last three commits address more complicated problems and I'd be happy
>>>>> to discuss alternative (maybe simpler) solutions.
>>>> 
>>>> I just wanted to say that this is really great work. Thanks a lot for
>>>> tracking all of these corner cases down!
>>> 
>>> +2
> 
> Thanks!
> 
> I forgot to add my signoffs and your acks, Jarno, sorry about that.
> 
>>> 
>>>> I think it would definitely be worthwhile to upstream your detection
>>>> code when you have a chance.
>>> 
>>> Maybe add ovs-appctl commands to turn on/off the immediate revalidation
>>> and related error reporting. I have a customer bug case were this would
>>> be immediately useful.
>> 
>> I think it would be useful to both have a command to turn on immediate
>> revalidation and also automatically do it for a low percentage of
>> flows that are installed (which is what we used to have and is nice
>> since it catches errors that nobody knows about yet). We can also turn
>> this on for unit tests and fail them if it detects an inconsistency.
> 
> An appctl is definitely a good idea.  Maybe we can make the percentage
> configurable and set it to 100% in the testcases.  I'll work on that.
>