Re: [PATCH net-next 0/5] net: dsa: Multi-CPU ground work

[Florian Fainelli <f.fainelli@gmail.com>]

On 06/06/2017 11:25 AM, Vivien Didelot wrote:
> Hi Florian,
> 
> Florian Fainelli <f.fainelli@gmail.com> writes:
> 
>> This patch series prepares the ground for adding mutliple CPU port support to
> 
>                                                    multiple
>                                                    
>> DSA, and starts by removing redundant pieces of information such as
>> master_netdev which is cpu_dp->ethernet. Finally drivers are moved away from
> 
>                          cpu_dp->netdev
>                          
>> directly accessing ds->dst->cpu_dp and use appropriate helper functions.
>>
>> Note that if you have Device Tree blobs/platform configurations that are
>> currently listing multiple CPU ports, the proposed behavior in
>> dsa_ds_get_cpu_dp() will be to return the last bit set in ds->cpu_port_mask.
>>
>> Future plans include:
>> - making dst->cpu_dp a flexible data structure (array, list, you name it)
>> - having the ability for drivers to return a default/preferred CPU port (if
>>   necessary)
> 
> The overall patchset looks good. I have questions for future work
> though.
> 
> I am still not sure that we need CPU port references in
> dsa_switch_tree. When device tree or pdata is parsed, we have allocated
> dsa_switch and dsa_port structures. We should be able validate and
> assign all ds->ports[x].cpu_dp, before setting up the switches and
> creating the slave devices. What do you think?

True, we should be able to do that, thanks for the suggestion.

> 
> Also I see dsa_ptr becoming a pointer to the assosicated dsa_port, and
> dsa_port should contain the tagging ops for quick access. That is more
> rigourous with the physical representation and much easier for
> transparent multi-CPU port support.

Ultimately, I agree we should probably have dev->dsa_ptr be the actual
CPU port within the switch, and from the switch be able to go to the
collection of switches (dst). This should indeed be a bit more optimized
as there should be less traversal of structures in such a case.

Thanks!
-- 
Florian