On Fri, Feb 09, 2024 at 02:44:04PM +0530, Jerin Jacob wrote:
> On Fri, Feb 2, 2024 at 6:11 PM Bruce Richardson
> <bruce.richard...@intel.com> wrote:
> >
> > Clarify the meaning of the NEW, FORWARD and RELEASE event types.
> > For the fields in "rte_event" struct, enhance the comments on each to
> > clarify the field's use, and whether it is preserved between enqueue and
> > dequeue, and it's role, if any, in scheduling.
> >
> > Signed-off-by: Bruce Richardson <bruce.richard...@intel.com>
> > ---
> > V3: updates following review
> > ---
> > lib/eventdev/rte_eventdev.h | 161 +++++++++++++++++++++++++-----------
> > 1 file changed, 111 insertions(+), 50 deletions(-)
> >
> > diff --git a/lib/eventdev/rte_eventdev.h b/lib/eventdev/rte_eventdev.h
> > index 8d72765ae7..58219e027e 100644
> > --- a/lib/eventdev/rte_eventdev.h
> > +++ b/lib/eventdev/rte_eventdev.h
> > @@ -1463,47 +1463,54 @@ struct rte_event_vector {
> >
> > /* Event enqueue operations */
> > #define RTE_EVENT_OP_NEW 0
> > -/**< The event producers use this operation to inject a new event to the
> > - * event device.
> > +/**< The @ref rte_event.op field must be set to this operation type to
> > inject a new event,
> > + * i.e. one not previously dequeued, into the event device, to be scheduled
> > + * for processing.
> > */
> > #define RTE_EVENT_OP_FORWARD 1
> > -/**< The CPU use this operation to forward the event to different event
> > queue or
> > - * change to new application specific flow or schedule type to enable
> > - * pipelining.
> > +/**< The application must set the @ref rte_event.op field to this
> > operation type to return a
> > + * previously dequeued event to the event device to be scheduled for
> > further processing.
> > *
> > - * This operation must only be enqueued to the same port that the
> > + * This event *must* be enqueued to the same port that the
> > * event to be forwarded was dequeued from.
> > + *
> > + * The event's fields, including (but not limited to) flow_id, scheduling
> > type,
> > + * destination queue, and event payload e.g. mbuf pointer, may all be
> > updated as
> > + * desired by the application, but the @ref rte_event.impl_opaque field
> > must
> > + * be kept to the same value as was present when the event was dequeued.
> > */
> > #define RTE_EVENT_OP_RELEASE 2
> > /**< Release the flow context associated with the schedule type.
> > *
> > - * If current flow's scheduler type method is *RTE_SCHED_TYPE_ATOMIC*
> > - * then this function hints the scheduler that the user has completed
> > critical
> > - * section processing in the current atomic context.
> > - * The scheduler is now allowed to schedule events from the same flow from
> > - * an event queue to another port. However, the context may be still held
> > - * until the next rte_event_dequeue_burst() call, this call allows but
> > does not
> > - * force the scheduler to release the context early.
> > - *
> > - * Early atomic context release may increase parallelism and thus system
> > + * If current flow's scheduler type method is @ref RTE_SCHED_TYPE_ATOMIC
> > + * then this operation type hints the scheduler that the user has
> > completed critical
> > + * section processing for this event in the current atomic context, and
> > that the
> > + * scheduler may unlock any atomic locks held for this event.
> > + * If this is the last event from an atomic flow, i.e. all flow locks are
> > released,
>
>
> Similar comment as other email
> [Jerin] When there are multiple atomic events dequeue from @ref
> rte_event_dequeue_burst()
> for the same event queue, and it has same flow id then the lock is ....
>
> [Mattias]
> Yes, or maybe describing the whole lock/unlock state.
>
> "The conceptual per-queue-per-flow lock is in a locked state as long
> (and only as long) as one or more events pertaining to that flow were
> scheduled to the port in question, but are not yet released."
>
> Maybe it needs to be more meaty, describing what released means. I don't
> have the full context of the documentation in my head when I'm writing this.
>
Will take a look to reword a bit
>
> > + * the scheduler is now allowed to schedule events from that flow from to
> > another port.
> > + * However, the atomic locks may be still held until the next
> > rte_event_dequeue_burst()
> > + * call; enqueuing an event with opt type @ref RTE_EVENT_OP_RELEASE allows,
>
> Is ";" intended?
>
> > + * but does not force, the scheduler to release the atomic locks early.
>
> instead of "not force", can use the term _hint_ the driver and reword.
Ok.
>
> > + *
> > + * Early atomic lock release may increase parallelism and thus system
> > * performance, but the user needs to design carefully the split into
> > critical
> > * vs non-critical sections.
> > *
> > - * If current flow's scheduler type method is *RTE_SCHED_TYPE_ORDERED*
> > - * then this function hints the scheduler that the user has done all that
> > need
> > - * to maintain event order in the current ordered context.
> > - * The scheduler is allowed to release the ordered context of this port and
> > - * avoid reordering any following enqueues.
> > - *
> > - * Early ordered context release may increase parallelism and thus system
> > - * performance.
> > + * If current flow's scheduler type method is @ref RTE_SCHED_TYPE_ORDERED
> > + * then this operation type informs the scheduler that the current event
> > has
> > + * completed processing and will not be returned to the scheduler, i.e.
> > + * it has been dropped, and so the reordering context for that event
> > + * should be considered filled.
> > *
> > - * If current flow's scheduler type method is *RTE_SCHED_TYPE_PARALLEL*
> > - * or no scheduling context is held then this function may be an NOOP,
> > - * depending on the implementation.
> > + * Events with this operation type must only be enqueued to the same port
> > that the
> > + * event to be released was dequeued from. The @ref rte_event.impl_opaque
> > + * field in the release event must have the same value as that in the
> > original dequeued event.
> > *
> > - * This operation must only be enqueued to the same port that the
> > - * event to be released was dequeued from.
> > + * If a dequeued event is re-enqueued with operation type of @ref
> > RTE_EVENT_OP_RELEASE,
> > + * then any subsequent enqueue of that event - or a copy of it - must be
> > done as event of type
> > + * @ref RTE_EVENT_OP_NEW, not @ref RTE_EVENT_OP_FORWARD. This is because
> > any context for
> > + * the originally dequeued event, i.e. atomic locks, or reorder buffer
> > entries, will have
> > + * been removed or invalidated by the release operation.
> > */
> >
> > /**
> > @@ -1517,56 +1524,110 @@ struct rte_event {
> > /** Event attributes for dequeue or enqueue operation */
> > struct {
> > uint32_t flow_id:20;
> > - /**< Targeted flow identifier for the enqueue and
> > - * dequeue operation.
> > - * The value must be in the range of
> > - * [0, nb_event_queue_flows - 1] which
> > - * previously supplied to rte_event_dev_configure().
> > + /**< Target flow identifier for the enqueue and
> > dequeue operation.
> > + *
> > + * For @ref RTE_SCHED_TYPE_ATOMIC, this field is
> > used to identify a
> > + * flow for atomicity within a queue & priority
> > level, such that events
> > + * from each individual flow will only be scheduled
> > to one port at a time.
> > + *
> > + * This field is preserved between enqueue and
> > dequeue when
> > + * a device reports the @ref
> > RTE_EVENT_DEV_CAP_CARRY_FLOW_ID
> > + * capability. Otherwise the value is
> > implementation dependent
> > + * on dequeue.
> > */
> > uint32_t sub_event_type:8;
> > /**< Sub-event types based on the event source.
> > + *
> > + * This field is preserved between enqueue and
> > dequeue.
> > + * This field is for application or event adapter
> > use,
> > + * and is not considered in scheduling decisions.
>
>
> cnxk driver is considering this for scheduling decision to
> differentiate the producer i.e event adapters.
> If other drivers are not then we can change the language around it is
> implementation defined.
>
How does the event type influence the scheduling decision? I can drop the
last line here, but it seems strange to me that the type of event could affect
things. I would have thought that with the eventdev API only the queue,
flow id, and priority would be factors in scheduling?
>
> > + *
> > * @see RTE_EVENT_TYPE_CPU
> > */
> > uint32_t event_type:4;
> > - /**< Event type to classify the event source.
> > - * @see RTE_EVENT_TYPE_ETHDEV, (RTE_EVENT_TYPE_*)
> > + /**< Event type to classify the event source.
> > (RTE_EVENT_TYPE_*)
> > + *
> > + * This field is preserved between enqueue and
> > dequeue
> > + * This field is for application or event adapter
> > use,
> > + * and is not considered in scheduling decisions.
>
>
> cnxk driver is considering this for scheduling decision to
> differentiate the producer i.e event adapters.
> If other drivers are not then we can change the language around it is
> implementation defined.
>
> > */
> > uint8_t op:2;
> > - /**< The type of event enqueue operation -
> > new/forward/
> > - * etc.This field is not preserved across an
> > instance
> > - * and is undefined on dequeue.
> > - * @see RTE_EVENT_OP_NEW, (RTE_EVENT_OP_*)
> > + /**< The type of event enqueue operation -
> > new/forward/ etc.
> > + *
> > + * This field is *not* preserved across an instance
> > + * and is implementation dependent on dequeue.
> > + *
> > + * @see RTE_EVENT_OP_NEW
> > + * @see RTE_EVENT_OP_FORWARD
> > + * @see RTE_EVENT_OP_RELEASE
> > */
> > uint8_t rsvd:4;
> > - /**< Reserved for future use */
> > + /**< Reserved for future use.
> > + *
> > + * Should be set to zero on enqueue.
>
> I am worried about some application explicitly start setting this to
> zero on every enqueue.
> Instead, can we say application should not touch the field, Since every
> eventdev
> operations starts with dequeue() driver can fill to the correct value.
>
I'll set this to zero on "NEW", or untouched on FORWARD/RELEASE.
If we don't state that it should be zeroed on NEW or untouched
otherwise we cannot use the space in future without ABI break.
> > + */
> > uint8_t sched_type:2;
> > /**< Scheduler synchronization type
> > (RTE_SCHED_TYPE_*)
> > * associated with flow id on a given event queue
> > * for the enqueue and dequeue operation.
> > + *
> > + * This field is used to determine the scheduling
> > type
> > + * for events sent to queues where @ref
> > RTE_EVENT_QUEUE_CFG_ALL_TYPES
> > + * is configured.
> > + * For queues where only a single scheduling type
> > is available,
> > + * this field must be set to match the configured
> > scheduling type.
> > + *
> > + * This field is preserved between enqueue and
> > dequeue.
> > + *
> > + * @see RTE_SCHED_TYPE_ORDERED
> > + * @see RTE_SCHED_TYPE_ATOMIC
> > + * @see RTE_SCHED_TYPE_PARALLEL
> > */
> > uint8_t queue_id;
> > /**< Targeted event queue identifier for the
> > enqueue or
> > * dequeue operation.
> > - * The value must be in the range of
> > - * [0, nb_event_queues - 1] which previously
> > supplied to
> > - * rte_event_dev_configure().
> > + * The value must be less than @ref
> > rte_event_dev_config.nb_event_queues
> > + * which was previously supplied to
> > rte_event_dev_configure().
>
> Some reason, similar text got removed for flow_id. Please add the same.
>
That was deliberate based on discussion on V2. See:
http://inbox.dpdk.org/dev/zby3nb4ngs8t5...@bricha3-mobl.ger.corp.intel.com/
and wider thread discussion starting here:
http://inbox.dpdk.org/dev/zbvotaepzja0g...@bricha3-mobl.ger.corp.intel.com/
Basically, the comment is wrong based on what the code does now. No event
adapters or apps are limiting the flow-id, and nothing seems broken, so we
can remove the comment.
