On 03.02.2016 03:32, Joe Stringer wrote:
> On 29 January 2016 at 04:03, Ilya Maximets <i.maxim...@samsung.com> wrote:
>> Hi.
>>
>> On 29.01.2016 04:02, Daniele Di Proietto wrote:
>>> Hi guys,
>>>
>>> Sorry for resurrecting this thread, but I think we should reconsider
>>> this patch at this point.
>>>
>>> Let me explain: I'm still perfectly happy with the fix we applied,
>>> e4e74c3a2b9a("dpif-netdev: Purge all ukeys when reconfigure pmd."),
>>> it is necessary to properly collect stats and it works perfectly
>>> when reconfiguring pmd threads.
>>>
>>> I'm revisiting this, because I'm not that comfortable anymore assuming
>>> that packets with the exact same flow have to come in from the
>>> same queue. Here's a couple of scenarios:
>>>
>>> - With vhost-user multi queue, I'm not sure that anything prevents a
>>> guest from sending the same flow to different queues.
>>
>> Nothing prevents. That is true. But this situation isn't really clear.
>> Can you imagine situation where normal application in normal mode
>> sends same flow to different queues and expects no issues?
>>
>>> - We've encountered the same bug in dpif-netdev when adding/removing
>>> ports. We fixed it by never moving the queues from the original
>>> pmd thread, which makes sense, but that's just an example
>>> - At some point it would be nice to reassign queues among pmd threads
>>> without deleting the megaflow cache.
>>
>> I agree that this two scenarios can be useful. I want to implement
>> some kind of scheduler for pmd threads to manage port/rxqs assignment.
>> In this case ability to reassign queues without destroying of threads
>> will be very nice.
>
> I was hoping Alex might chime in on-list as he spent a bit more time
> thinking about this than me, but the patch seems reasonable to me.
> Longer term I think there are various things we can do to improve
> revalidation in the userspace datapath, but this change doesn't really
> make a difference for the direction of that work.
>
> For the actual patch, I think typically we would mix in values using
> something like hash_2words(ufid->u32[0], pmd_id) rather than addition.
>
> Ilya, do you mind resending? I can review/merge it unless anyone else
> has thoughts/objections.
OK. I'll resend this patch with above fix.
Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
