[dpdk-dev] rte_flow ageing

[David Bouyeure]

Thanks a lot Asaf, for your answer, so fast.

depending on the feature we want, the table you mentioned in the doc may 
give different combinations. Mine, DPDK-20.08/OFED 5.1-2, is part of the 
list.

Anyway, my question is more about the API design. Please, find my 
comments below.

On 3/29/21 8:02 PM, Asaf Penso wrote:
> Hello David,
>
> Thanks for reaching out, I'll try to answer as best as I know and I added Matan who will be able to provide further info during next week.
> First, according to our pmd documentation (http://doc.dpdk.org/guides/nics/mlx5.html#supported-hardware-offloads) we recommend using DPDK20.11 and OFED5.2, and not the combo you are referring to.
> Second, we can always improve our documentation and I appreciate your queries.
>
> Please see my comments inline.
>
> Regards,
> Asaf Penso
>
>> -----Original Message-----
>> From: dev <dev-bounces at dpdk.org> On Behalf Of David Bouyeure
>> Sent: Monday, March 29, 2021 11:35 AM
>> To: dev at dpdk.org
>> Subject: [dpdk-dev] rte_flow ageing
>>
>> Hi,
>>
>>
>> I've found out the pretty useful experimental brand new flow ageing API
>> implemented in the mlx5 PMD.
> It is useful and I hope you'll fully understand at the end why 😊
>   
>
>> I'm trying it (rte_eth_dev_callback_register(RTE_ETH_EVENT_FLOW_AGED),
>> RTE_FLOW_ACTION_TYPE_AGE) to recover any flow that I previously
>> offloaded.
>>
>> The DPDK version is 20.08 and Mellanox(Connect-X6) OFED drivers are 5.1-
>> 2.5.8.0.
>>
> See above the suggested versions for this feature
>
>> I eventually don't see the usefulness of the callback since it's actually triggered
>> indirectly by us(the DPDK application) when calling
>> rte_flow_get_aged_flows().
> The main intention is to offload the aging logic from the application level to the pmd level.
> There is so saving of cpu cycles, and the gain here is with simplicity.
> The application doesn't need to have complex logic of comparison between counters or other HW info that can be retrieve.
> Now, the pmd hides all of that and leaves the application only to decide what to do with the flows that are aged out.
> Please note, the pmd does not delete any flow, just provide the list of all the flows that are aged.
I fully understand that and this is a very very useful feature to us.
>> If we don't call it, the callback is called only once.
>>
>> And, calling rte_flow_get_aged_flows() from the callback won't trigger it next
>> time(MLX5_AGE_TRIGGER is reset after the callback call)
> Once you call the function the pmd will not trigger more events. Now it's up to the application to decide what to do.
> Doing it differently, will cause an interrupt storm and the pmd avoids that.If new flows are aged then the pmd will trigger a new event.

Sorry, I wasn't realizing that the callback isn't called for each flow 
but rather for each port, though it's clear in the PMD code. But, the 
fact that we can register several RTE_ETH_EVENT_FLOW_AGED event handlers 
is surprising.

So, you suggest to use the callback as an indicator to later retrieve 
the aged-out flows, that's it?

Wouldn't calling rte_flow_get_aged_flows with NULL param just to get the 
number of aged_flows do the same, without the need to un/register a 
callback, and DPDK to call it?

Another thing, the explanation here 
http://doc.dpdk.org/api/rte__flow_8h.html#a43763e0794d2696b18b6272619aafc2a 
*"...to get the aged flows usynchronously from the event callback..."* 
seems wrong to me because age_info->flags is set to 0 just after the 
callback, thus ML5_AGE_TRIGGER is canceled and no event will be 
triggered before we'll call rte_flow_get_aged_flows() outside of the 
callback.
>> Furthermore, I don't see the point of computing ageing flows in
>> mlx5_fow.c::mlx5_flow_aging_check() if the client callback isn't called.
>>
> Can you elaborate? I'm not sure I understand your intention.
Please forgot :-)
>
>> So far, I can handle the flow ageing from the same thread as the one which is
>> handling the flow direction(rte_flow), it even avoid threads synchronization.
>> But, in the future, I may need to be noticed as soon as possible of a single flow
>> ageing, and thus handle this flow logic from the ageing callback.
>>
>>
>> I may misunderstand the whole ageing API... Thanks a lot for any clarification.