Re: [PATCH v2 1/2] eal: expose lcore pthread id

[Tyler Retzlaff <roretzla@linux.microsoft.com>]

On Thu, Oct 20, 2022 at 10:03:01PM +0200, Michael Pfeiffer wrote:
> On Thu, 2022-10-20 at 08:36 -0700, Stephen Hemminger wrote:
> > On Thu, 20 Oct 2022 13:20:40 +0200
> > David Marchand <david.marchand@redhat.com> wrote:
> > 
> > > On Fri, Oct 14, 2022 at 9:54 AM Markus Theil <markus.theil@tu-ilmenau.de>
> > > wrote:
> > > > 
> > > > From: Michael Pfeiffer <michael.pfeiffer@tu-ilmenau.de>
> > > > 
> > > > Also expose the pthread id of each lcore, in
> > > > order to allow modification of pthread attributes,
> > > > for example use rte_thread_setname without executing
> > > > pthread_self() on the maybe already running lcore.
> > > > 
> > > > The rte_lcore_to_thread_id function is added to API.
> > > > 
> > > > Signed-off-by: Michael Pfeiffer <michael.pfeiffer@tu-ilmenau.de>  
> > > 
> > > We are trying to abstract the use of pthread in DPDK API.
> > > So I don't think this patch is going in the right direction.
> > 
> > Agree, exposing this will make Windows support harder
> > and who knows what next OS to come will need.
> 
> Hi,
> thanks for your feedback. I understand your concerns regarding abstraction and
> portability.
> 
> Markus and I ultimately use the function in the patch to call
> rte_thread_setname() (which takes the pthread id as an argument) to rename our
> lcore workers from "lcore-worker-X" to something more meaningful in the scope
> of our application. Having descriptive thread names makes debugging
> significantly easier. For example, verifying CPU pinning worked as intended
> with ps -T ..., or identifying threads in the Intel VTune profiler.
> 
> Would you consider something like
> - int rte_lcore_setname(unsigned int lcore_id, const char *name)
> - int rte_lcore_getname(unsigned int lcore_id, char *name, size_t len)
> a more appropriate API? That would still allow us to set names from the main
> lcore, but would not expose any pthread internals.

if only setname, getname are needed for rte_thread_t i imagine it
shouldn't be too objectionable to add them to rte_thread.h

you can either wait until i have time to do it (could be a while) or you
can put a new patch + unit test together.

if testing the windows implementation part of the addition is a barrier
i can apply and run the supplied unit test, assist with review.

> 
> Thanks & regards
> Michael