SWX table question

SWX table question

[Morten Brørup]

Thank you for an interesting presentation today, Cristian.

It made me aware of the existence of the SWX Table, which could be used for connection tracking.

I hadn't noticed the library before, because it is documented (and named) as part of the SWX/Pipeline Framework, and we don't use the SWX/Pipeline Framework, so I have ignored anything related to that framework.

My question is: Is the SWX Table library tied to the SWX/Pipeline Framework, or can a DPDK application use it independently of that framework, like any other DPDK library? If so, are there any limitations or restrictions - e.g. is it lockless, and can it be accessed by multiple cores simultaneously?


Med venlig hilsen / Kind regards,
-Morten Brørup

RE: SWX table question

[Dumitrescu, Cristian]

Hi Morten,

> -----Original Message-----
> From: Morten Brørup <mb@smartsharesystems.com>
> Sent: Thursday, September 8, 2022 2:49 PM
> To: Dumitrescu, Cristian <cristian.dumitrescu@intel.com>
> Cc: dev@dpdk.org
> Subject: SWX table question
> 
> Thank you for an interesting presentation today, Cristian.
> 
> It made me aware of the existence of the SWX Table, which could be used
> for connection tracking.
> 
> I hadn't noticed the library before, because it is documented (and named) as
> part of the SWX/Pipeline Framework, and we don't use the SWX/Pipeline
> Framework, so I have ignored anything related to that framework.
> 
> My question is: Is the SWX Table library tied to the SWX/Pipeline Framework,
> or can a DPDK application use it independently of that framework, like any
> other DPDK library? If so, are there any limitations or restrictions - e.g. is it
> lockless, and can it be accessed by multiple cores simultaneously?
> 

Yes, the table library can absolutely be used directly by the applications.

No, the table library is not lockless. For a given table object, the create, add, delete, free operations need to be serialized, while for the lookup operation it depends on the table type: for some table types, the lookup operation is read only, hence multiple lookup operations launched from different threads can overlap, while for some other table types, notably the learner table (used to implement the P4 PNA add-on-miss tables), the lookup operation is also modifying the data structures (e.g. the key timestamp), hence this operation also needs serial access to the table.

> 
> Med venlig hilsen / Kind regards,
> -Morten Brørup

Regards,
Cristian