Re: [dpdk-dev] [PATCH 1/3] evendev: fix inconsistency in event queue config

[Pavan Nikhilesh Bhagavatula <pbhagavatula@caviumnetworks.com>]

On Fri, Oct 20, 2017 at 09:54:36AM +0000, Van Haaren, Harry wrote:
> > From: Pavan Nikhilesh [mailto:pbhagavatula@caviumnetworks.com]
> > Sent: Thursday, October 12, 2017 2:16 PM
> > To: jerin.jacob@caviumnetworks.com; hemant.agrawal@nxp.com; Van Haaren,
> > Harry <harry.van.haaren@intel.com>
> > Cc: dev@dpdk.org; Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>
> > Subject: [dpdk-dev] [PATCH 1/3] evendev: fix inconsistency in event queue
> > config
> >
> > With the current scheme of event queue configuration the cfg schedule
> > type macros (RTE_EVENT_QUEUE_CFG_*_ONLY) are inconsistent with the
> > event schedule type (RTE_SCHED_TYPE_*) this requires unnecessary
> > conversion between the fastpath and slowpath API's while scheduling
> > events or configuring event queues.
> >
> > This patch aims to fix such inconsistency by using event schedule
> > types (RTE_SCHED_TYPE_*) for event queue configuration.
>
> True - worth cleaning up.
>
>
> > This patch also fixes example/eventdev_pipeline_sw_pmd as it doesn't
> > convert RTE_EVENT_QUEUE_CFG_*_ONLY to RTE_SCHED_TYPE_* which leads to
> > improper events being enqueued to the eventdev.
> >
> > Fixes: adb5d5486c39 ("examples/eventdev_pipeline_sw_pmd: add sample app")
> >
> > Signed-off-by: Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>
> >
> >
> >  app/test-eventdev/evt_common.h           | 21 -------------
> >  app/test-eventdev/test_order_queue.c     |  4 +--
> >  app/test-eventdev/test_perf_queue.c      |  4 +--
> >  drivers/event/dpaa2/dpaa2_eventdev.c     |  4 +--
> >  drivers/event/sw/sw_evdev.c              | 28 +++++------------
> >  examples/eventdev_pipeline_sw_pmd/main.c | 18 +++++------
> >  lib/librte_eventdev/rte_eventdev.c       | 20 +++++-------
> >  lib/librte_eventdev/rte_eventdev.h       | 54 ++++++++++-------------------
> > ---
> >  test/test/test_eventdev.c                | 12 +++----
> >  test/test/test_eventdev_sw.c             | 16 +++++-----
> >  10 files changed, 60 insertions(+), 121 deletions(-)
> >
> > diff --git a/app/test-eventdev/evt_common.h b/app/test-eventdev/evt_common.h
> > index 4102076..ee896a2 100644
> > --- a/app/test-eventdev/evt_common.h
> > +++ b/app/test-eventdev/evt_common.h
> > @@ -92,25 +92,4 @@ evt_has_all_types_queue(uint8_t dev_id)
> >  			true : false;
> >  }
> >
> > -static inline uint32_t
> > -evt_sched_type2queue_cfg(uint8_t sched_type)
> > -{
> > -	uint32_t ret;
> > -
> > -	switch (sched_type) {
> > -	case RTE_SCHED_TYPE_ATOMIC:
> > -		ret = RTE_EVENT_QUEUE_CFG_ATOMIC_ONLY;
> > -		break;
> > -	case RTE_SCHED_TYPE_ORDERED:
> > -		ret = RTE_EVENT_QUEUE_CFG_ORDERED_ONLY;
> > -		break;
> > -	case RTE_SCHED_TYPE_PARALLEL:
> > -		ret = RTE_EVENT_QUEUE_CFG_PARALLEL_ONLY;
> > -		break;
> > -	default:
> > -		rte_panic("Invalid sched_type %d\n", sched_type);
> > -	}
> > -	return ret;
> > -}
>
>
> We should note here, that SCHED_TYPE are integer values:
> #define RTE_SCHED_TYPE_ORDERED          0
> #define RTE_SCHED_TYPE_ATOMIC           1
> #define RTE_SCHED_TYPE_PARALLEL         2
>
> While the EVENT_QUEUE_CFG_ types were bitmasks (before being removed in this patch)
> #define RTE_EVENT_QUEUE_CFG_ATOMIC_ONLY        (1ULL << 0)
> #define RTE_EVENT_QUEUE_CFG_ORDERED_ONLY       (2ULL << 0)
> #define RTE_EVENT_QUEUE_CFG_PARALLEL_ONLY      (3ULL << 0)
> #define RTE_EVENT_QUEUE_CFG_SINGLE_LINK        (1ULL << 2)
>
>
> I'm not against this change - but we must be careful that if there was any bit-masking being used previously,
> that that will subtly have broken if we change to sched types. I'm reviewing with that in mind..
>
> The RTE_EVENT_QUEUE_CFG_ALL_TYPES  config flag now means that all SCHED_TYPEs
> are valid. Previously this was contained in the bitmask.. this may lead to issues.
>
> See patch 2/3, where *only* the schedule_type is read, and returned. What if it the "ALL_TYPES" flag is
> set on the queue? It will be reported wrongly. Currently there is no integer value for "ALL_TYPES",
> so we cannot represent "ALL TYPES" in the return value from get_attr().
>
> Am I explaining my reasoning clearly enough?
>

Hey Harry,

I do understand what you mean, my initial thought was to include "ALL_TYPES" as
a schedule_type in queue config but this would just complicate things.

As these values are only used in config phase we could have a check to see if
event_queue_cfg is not "ALL_TYPES" and only then return the value of sched_type
else return a error value in case of get_attr().

I think most of the places this specific check is handled, one such missed
place would be get_attr(). If we could come to a conclusion to fix it in a
correct way I will send out a v2.

Thanks,
Pavan.

>
> The patch correctly leaves QUEUE_CFG_SINGLE_LINK as a bitmask, so that bit is ok.
>
>
>
> >  #endif /*  _EVT_COMMON_*/
> > diff --git a/app/test-eventdev/test_order_queue.c b/app/test-
> > eventdev/test_order_queue.c
> > index beadd9c..1fa4082 100644
>
>
> <snip> Remaining diff was updating the QUEUE_CFG to SCHED_TYPE </snip>