On Thu, Jan 26, 2017 at 08:39:57PM +0000, Eads, Gage wrote:
>
>
> > -----Original Message-----
> > From: Jerin Jacob [mailto:jerin.ja...@caviumnetworks.com]
> > Sent: Thursday, January 26, 2017 3:39 AM
> > To: Eads, Gage <gage.e...@intel.com>
> > Cc: Richardson, Bruce <bruce.richard...@intel.com>; 'dev@dpdk.org'
> > <dev@dpdk.org>; 'thomas.monja...@6wind.com'
> > <thomas.monja...@6wind.com>; 'hemant.agra...@nxp.com'
> > <hemant.agra...@nxp.com>; Van Haaren, Harry
> > <harry.van.haa...@intel.com>; McDaniel, Timothy
> > <timothy.mcdan...@intel.com>
> > Subject: Re: [dpdk-dev] [PATCH v4 1/6] eventdev: introduce event driven
> > programming model
> >
> > On Wed, Jan 25, 2017 at 10:36:21PM +0000, Eads, Gage wrote:
> > > > > <jerin.ja...@caviumnetworks.com> > > Subject: [dpdk-dev] [PATCH
> > > > v4 > 1/6] eventdev: introduce event driven > > programming model
> > > > > > > <message truncated for brevity> > > +/** > > + * Enqueue
> > > > a burst > of events objects or an event object supplied > > in >
> > > > > > *rte_event* > > + * structure on an event device designated
> > > > by its > *dev_id* > > through the event + * port specified by
> > > > *port_id*. Each > event > > object specifies the event queue on +
> > > > * which it will be > enqueued.
> > > > > > > + *
> > > > > > > + * The *nb_events* parameter is the number of event
> > > > objects to > > > enqueue which are + * supplied in the *ev* array
> > > > of *rte_event* > > > structure.
> > > > > > > + *
> > > > > > > + * The rte_event_enqueue_burst() function returns the
> > > > number of > + > > * events objects it actually enqueued. A return
> > > > value equal to > > > *nb_events* + * means that all event objects
> > have
> > been enqueued.
> > > > > > > + *
> > > > > > > + * @param dev_id
> > > > > > > + * The identifier of the device.
> > > > > > > + * @param port_id
> > > > > > > + * The identifier of the event port.
> > > > > > > + * @param ev
> > > > > > > + * Points to an array of *nb_events* objects of type
> > *rte_event*
> > > > > > structure
> > > > > > > + * which contain the event object enqueue operations to be
> > > > > > processed.
> > > > > > > + * @param nb_events
> > > > > > > + * The number of event objects to enqueue, typically
> > number of
> > > > > > > + * rte_event_port_enqueue_depth() available for this port.
> > > > > > > + *
> > > > > > > + * @return
> > > > > > > + * The number of event objects actually enqueued on the
> > event
> > > > > > device. The
> > > > > > > + * return value can be less than the value of the
> > *nb_events*
> > > > > > parameter
> > > > > > > when
> > > > > > > + * the event devices queue is full or if invalid
> > parameters are
> > > > > > specified in a
> > > > > > > + * *rte_event*. If return value is less than *nb_events*,
> > the
> > > > > > remaining events
> > > > > > > + * at the end of ev[] are not consumed,and the caller has
> > to take
> > > > > > care of
> > > > > > > them
> > > > > > > + *
> > > > > > > + * @see rte_event_port_enqueue_depth() + */ +uint16_t >
> > > > > > +rte_event_enqueue_burst(uint8_t dev_id, uint8_t port_id,
> > > > > > > + const struct rte_event ev[], uint16_t
> > nb_events);
> > > > > >
> > > > > > There are a number of reasons this operation could fail to
> > > > enqueue > all > the events, including:
> > > > > > - Backpressure
> > > > > > - Invalid port ID
> > > > > > - Invalid queue ID
> > > > > > - Invalid sched type when a queue is configured for
> > > > ATOMIC_ONLY, > > ORDERED_ONLY, or PARALLEL_ONLY > - ...
> > > > > >
> > > > > > The current API doesn't provide a straightforward way to
> > > > determine > the > cause of a failure. This is a particular issue
> > > > on event PMDs > that can > backpressure, where the app may want to
> > > > treat that case > differently > than the other failure cases.
> > > > > >
> > > > > > Could we change the return type to int16_t, and define a set
> > > > of > error > cases (e.g. -ENOSPC for backpressure, -EINVAL for an
> > > > invalid argument)?
> > > > > > (With corresponding changes needed in the PMD API) Similarly
> > > > we > could > change rte_event_dequeue_burst() to return an
> > > > int16_t, with > -EINVAL as > a possible error case.
> > > > >
> > > > > Use rte_errno instead, I suggest. That's what it's there for.
> > > > >
> > > > > /Bruce
> > > >
> > > > That makes sense. In that case, the API comment could be tweaked like
> > so:
> > > >
> > > > * If the return value is less than *nb_events*, the remaining
> > events at the
> > > > * end of ev[] are not consumed and the caller has to take care of
> > them,
> > and
> > > > * rte_errno is set accordingly. Possible errno values include:
> > > > * - EINVAL - The port ID is invalid, an event's queue ID is
> > invalid, or an
> > > > * event's sched type doesn't match the capabilities of
> > the
> > > > * destination queue.
> > > > * - ENOSPC - The event port was backpressured and unable to enqueue
> > one or
> > > > * more events.
> > >
> > > However it seems better to use a signed integer for the dequeue burst
> > return
> > value, if it is to use rte_errno. Application code could be simplified:
> > >
> > > (signed return value)
> > > ret = rte_event_dequeue_burst(...);
> > > if (ret < 0)
> > > rte_panic("Dequeued returned errno %d\n", rte_errno);
> > >
> > > vs.
> > >
> > > (unsigned return value)
> > > ret = rte_event_dequeue_burst(...);
> > > if (ret == 0 && rte_errno != 0)
> > > rte_panic("Dequeued returned errno %d\n", rte_errno);
> > >
> > > And with an unsigned return value, all dequeue implementations would have
> > to clear rte_errno when no events are dequeued.
>
> After some internal discussion, I don't think the signed return value is
> necessary for burst dequeue. Burst enqueue is the more interesting case...
>
> >
> > Gage,
> >
> > Just to understand, what is the expected application behavior if the
> > implementation returns -ENOSPC
>
> It's application-dependent -- depending on the importance of the event, the
> application could decide to retry the enqueue some number of times or decide
> to drop the event.
>
> >
> > Apart for the above SW driver behavior, I think, HW implementation has two
> > more different behavior
> > a) Implementation make sure that it never returns -ENOSPC by allocating
> > more
> > space on the fly or any other scheme
> > b) Tail drop
> >
>
> By "tail drop," do you mean the hardware drops the event (and presumably
> frees any memory it points to)? Or the enqueue is unsuccessful and the
> application drops the event?
>
> > Considering different implementation has different behaviors, How about
> > enumerating the overflow policy at the port configuration time? and let
> > implementation act accordingly to avoid fast-patch change in
> > application(effects in all implementation irrespective of the capability)
> >
> > possible enumerating value at the port configuration time,
> > - PANIC or similar scheme to denote it cannot proceed
> > - TAIL DROP
> > or any expected application behavior you want to add
>
> I wonder if that's necessary? Hardware behavior a) means the function will
> always return nb_events. If hardware drops the event(s), I assume
> enqueue_burst would still return nb_events and the app behaves as if all
> events were sent. If the enqueue fails (ret < nb_events), app software could
> check rte_errno and take the action it deems necessary. So all fast-path
> enqueue code could look like:
>
> ret = rte_event_enqueue_burst(..., nb_events);
> if (ret < nb_events) {
> ....
> }
I would agree with that.
I think both enqueue and dequeue should have unsigned return values.
Both should set rte_errno on unsuccessful or partially successful
operation i.e.:
enqueue: sets errno where ret < nb_events
dequeue: sets errno where ret == 0 (errno may be set to no-error
if queue is just empty)
/Bruce
