Re: [RCU PATCH 00/14] Remove qdisc lock around ingress Qdisc

[John Fastabend <john.fastabend@gmail.com>]

On 03/13/2014 01:44 AM, Jamal Hadi Salim wrote:
> On 03/12/14 12:45, John Fastabend wrote:
>> On 03/11/2014 11:58 PM, Jamal Hadi Salim wrote:
>>> On 03/10/14 13:03, John Fastabend wrote:
>>
>> Sure I can provide this data as part of the patch description. My
>> expectation is now that the qdisc lock is not needed there should
>> be less impact to throughput/latency due to filter updates.
>
> Thats what i would think as well (which would be great).
>
>>
>> Yes upto one RCU grace period. Although I don't think this is an
>> issue because we get consistency eventually and even with the qdisc
>> lock there is no way to know if a set of skbs hit the filter list
>> before or after the update.
>>
>
> Probably not a big issue - but needs to be document clearly since the
> expectations have changed.
>
>> I'll get a v2 out tomorrow morning after making Eric's changes and
>> fixing the last compiler warning.
>>
>
> There is one more thing. Actions can be shared. I am wondering how this
> change is going to affect it - can you try this attached script? Sorry,
> I could have sent a much simpler script - but this is one sitting on
> my laptop (i have annotated it - notice the 2 policer instances
> are shared).
> Would be useful to do pings across two devices at the same time while
> concurently doing dumps and see if anything goes crazy (like an oops).

Ran the attached script and modified it to match UDP packets so I could
use it with pktgen to hammer it a bit more. And changed it to add/delete
filters with traffic while doing a watch on the filter and qdisc stats.
No oops, splats or anything else crazy.

I need to get some sleep now though so the performance numbers will have
to wait for tomorrow.

Thanks!
John

-- 
John Fastabend         Intel Corporation