Thanks Jerin, agree with you on graceful shutdown in rc2.
> -----Original Message----- > From: Jerin Jacob <jerinjac...@gmail.com> > Sent: Sunday, May 3, 2020 3:28 PM > To: Varghese, Vipin <vipin.vargh...@intel.com> > Cc: Pavan Nikhilesh Bhagavatula <pbhagavat...@marvell.com>; Jerin Jacob > Kollanukkaran <jer...@marvell.com>; Burakov, Anatoly > <anatoly.bura...@intel.com>; dpdk-dev <dev@dpdk.org> > Subject: Re: [dpdk-dev] [PATCH] eventdev: fix device probe and remove for > secondary process > > On Sun, May 3, 2020 at 6:45 AM Varghese, Vipin <vipin.vargh...@intel.com> > wrote: > > > > Hi Pavan, > > > > Snipped > > > > > >> > > > > > >> > > When probing event device in secondary process skip > > > >> > > reinitializing the device data structure as it is already > > > >> > > done in primary > > > process. > > > >> > > > > > >> > > When removing event device in secondary process skip closing > > > >> > > the event device as it should be done by primary process. > > > >If primary has crashed or closed before secondary abnormally. > > > >Should not close of secondary trigger removal of Eventdev services? > > > > > > Closing event device on exit of one secondary doesn’t make sense as > > > there might be systems where there might be one primary and multiple > > > secondaries and secondaries are spawned/destroyed on demand. > > > > > > Behavior of secondaries on primary process crash is undefined. > > Absolutely true, there are work scenarios where primary configures ports > and Eventdev queues-ports pair. It would be multiple secondaries which > process packets via event dev. In such cases, when primary segfaults or > crashes > it will lead to undefined states. > > > > In such scenarios, would not it be preferer able for all secondaries to > subscribe to service function like health check. If the primary is not alive > anymore, then gracefully handle inflight events and events to dequeue. If this > is right understanding, should not there be option in secondary to gracefully > shut down it's worker queue and ports (rather than event device instance)? > > The health check function may not be limited to eventdev, it must apply for > all > the subsystems in multiprocess mode if primary dies. > Such features can be designed/agreed based on every subsystem in mind. > For rc2, Let's have this bug fix. New features can be implemented after an > agreement with all stakeholders wrt multi-process semantics which applicable > for all subsystems. > > > > > > snipped
