[dpdk-dev] [PATCH] doc: abstract the behaviour of rte_ctrl_thread_create

[Olivier Matz]

Hi Honnappa,

Back from holidays, sorry for the late answer.

On Mon, Aug 09, 2021 at 01:18:42PM +0000, Honnappa Nagarahalli wrote:
> <snip>
> > 
> > 30/07/2021 23:44, Honnappa Nagarahalli:
> > > The current expected behaviour of the function rte_ctrl_thread_create
> > > is rigid which makes the implementation of the function complex.
> > > Make the expected behaviour abstract to allow for simplified
> > > implementation.
> > >
> > > With this change, the calls to pthread_setaffinity_np can be moved to
> > > the control thread. This will avoid the use of pthread_barrier_wait
> > > and simplify the synchronization mechanism between
> > > rte_ctrl_thread_create and the calling thread.
> > >
> > > Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli at arm.com>
> > > ---
> > > +* eal: The expected behaviour of the function
> > > +``rte_ctrl_thread_create``
> > > +  abstracted to allow for simplified implementation. The new
> > > +behaviour is
> > > +  as follows:
> > > +  Creates a control thread with the given name. The affinity of the
> > > +new
> > > +  thread is based on the CPU affinity retrieved at the time
> > > +rte_eal_init()
> > > +  was called, the dataplane and service lcores are then excluded.
> > 
> > I don't understand what is different of the current API:
> >  * Wrapper to pthread_create(), pthread_setname_np() and
> >  * pthread_setaffinity_np(). The affinity of the new thread is based
> >  * on the CPU affinity retrieved at the time rte_eal_init() was called,
> >  * the dataplane and service lcores are then excluded.
> My concern is for the word "Wrapper". I am not sure how much we are bound by that to keep the code as a "wrapper".
> The new patch does not change the high level behavior.

I am ok to remove the word "wrapper" from the description, and I agree
it can be better described without quoting the pthread_* functions.

> Are you saying you are ok with the patch without the deprecation notice?

I don't think it requires a deprecation notice if the API and ABI is
left unchanged. To be honnest, I find a bit hard to understand what is
really changed by reading the deprecation notice:

> +* eal: The expected behaviour of the function ``rte_ctrl_thread_create``
> +  abstracted to allow for simplified implementation. The new behaviour is
> +  as follows:
> +  Creates a control thread with the given name. The affinity of the new
> +  thread is based on the CPU affinity retrieved at the time rte_eal_init()
> +  was called, the dataplane and service lcores are then excluded.

I'll send my comments to your patch:
http://patches.dpdk.org/project/dpdk/patch/20210802051652.3611-1-honnappa.nagarahalli@arm.com/


Thanks,
Olivier