> -----Original Message----- > From: Jerin Jacob [mailto:jerin.ja...@caviumnetworks.com] > Sent: Monday, June 12, 2017 6:46 AM > To: dev@dpdk.org > Cc: Richardson, Bruce <bruce.richard...@intel.com>; Van Haaren, Harry > <harry.van.haa...@intel.com>; hemant.agra...@nxp.com; Eads, Gage > <gage.e...@intel.com>; nipun.gu...@nxp.com; Vangati, Narender > <narender.vang...@intel.com>; Rao, Nikhil <nikhil....@intel.com>; Jerin Jacob > <jerin.ja...@caviumnetworks.com> > Subject: [dpdk-dev] [PATCH] eventdev: add producer enqueue hint > > Some PMD like OCTEONTX can have optimized handling of events if the PMD > knows it is a producer pattern in advance. > For instance, OCTEONTX PMD can support burst mode if op == > RTE_EVENT_OP_NEW. > > Since the event producer initialize(set all_op_new == 1) the event object > before > the main producer loop, This scheme does not call for any performance > regression on other PMDs. > > Signed-off-by: Jerin Jacob <jerin.ja...@caviumnetworks.com> > --- > Another option is to add a flag in enqueue API or have parallel enqueue API. > --- > drivers/event/octeontx/ssovf_worker.c | 12 ++++++++++-- > lib/librte_eventdev/rte_eventdev.h | 10 +++++++++- > 2 files changed, 19 insertions(+), 3 deletions(-) > > diff --git a/drivers/event/octeontx/ssovf_worker.c > b/drivers/event/octeontx/ssovf_worker.c > index ad3fe684d..209c595cf 100644 > --- a/drivers/event/octeontx/ssovf_worker.c > +++ b/drivers/event/octeontx/ssovf_worker.c > @@ -196,8 +196,16 @@ ssows_enq(void *port, const struct rte_event *ev) > uint16_t __hot ssows_enq_burst(void *port, const struct rte_event ev[], > uint16_t nb_events) { > - RTE_SET_USED(nb_events); > - return ssows_enq(port, ev); > + uint16_t i; > + struct ssows *ws = port; > + > + if (ev[0].all_op_new) { > + rte_smp_wmb(); > + for (i = 0; i < nb_events; i++) > + ssows_new_event(ws, &ev[i]); > + return i; > + } else > + return ssows_enq(port, ev); > } > > void > diff --git a/lib/librte_eventdev/rte_eventdev.h > b/lib/librte_eventdev/rte_eventdev.h > index a248fe90e..1c1a46593 100644 > --- a/lib/librte_eventdev/rte_eventdev.h > +++ b/lib/librte_eventdev/rte_eventdev.h > @@ -933,7 +933,15 @@ struct rte_event { > * and is undefined on dequeue. > * @see RTE_EVENT_OP_NEW, (RTE_EVENT_OP_*) > */ > - uint8_t rsvd:4; > + uint8_t all_op_new:1; > + /**< Valid only with event enqueue operation - This hint > + * indicates that the enqueue request has only the > + * events with op == RTE_EVENT_OP_NEW. > + * The event producer, typically use this pattern to > + * inject the events to eventdev. > + * @see RTE_EVENT_OP_NEW > rte_event_enqueue_burst() > + */ > + uint8_t rsvd:3; > /**< Reserved for future use */ > uint8_t sched_type:2; > /**< Scheduler synchronization type > (RTE_SCHED_TYPE_*) > -- > 2.13.1
I slightly prefer the parallel enqueue API -- I can see folks making the mistake of setting all_op_new without setting the op to RTE_EVENT_OP_NEW, and later adding a "forward-only" enqueue API could be interesting for the sw PMD -- but this looks fine to me. Curious if others have any thoughts.
