Re: [dpdk-dev] [PATCH 01/27] eventdev: dlb upstream prerequisites

["McDaniel, Timothy" <timothy.mcdaniel@intel.com>]

-----Original Message-----
From: Jerin Jacob <jerinjacobk@gmail.com> 
Sent: Saturday, June 27, 2020 2:45 AM
To: McDaniel, Timothy <timothy.mcdaniel@intel.com>; Ray Kinsella <mdr@ashroe.eu>; Neil Horman <nhorman@tuxdriver.com>
Cc: Jerin Jacob <jerinj@marvell.com>; Mattias Rönnblom <mattias.ronnblom@ericsson.com>; dpdk-dev <dev@dpdk.org>; Eads, Gage <gage.eads@intel.com>; Van Haaren, Harry <harry.van.haaren@intel.com>
Subject: Re: [dpdk-dev] [PATCH 01/27] eventdev: dlb upstream prerequisites

> +
> +/** Event port configuration structure */
> +struct rte_event_port_conf_v20 {
> +       int32_t new_event_threshold;
> +       /**< A backpressure threshold for new event enqueues on this port.
> +        * Use for *closed system* event dev where event capacity is limited,
> +        * and cannot exceed the capacity of the event dev.
> +        * Configuring ports with different thresholds can make higher priority
> +        * traffic less likely to  be backpressured.
> +        * For example, a port used to inject NIC Rx packets into the event dev
> +        * can have a lower threshold so as not to overwhelm the device,
> +        * while ports used for worker pools can have a higher threshold.
> +        * This value cannot exceed the *nb_events_limit*
> +        * which was previously supplied to rte_event_dev_configure().
> +        * This should be set to '-1' for *open system*.
> +        */
> +       uint16_t dequeue_depth;
> +       /**< Configure number of bulk dequeues for this event port.
> +        * This value cannot exceed the *nb_event_port_dequeue_depth*
> +        * which previously supplied to rte_event_dev_configure().
> +        * Ignored when device is not RTE_EVENT_DEV_CAP_BURST_MODE capable.
> +        */
> +       uint16_t enqueue_depth;
> +       /**< Configure number of bulk enqueues for this event port.
> +        * This value cannot exceed the *nb_event_port_enqueue_depth*
> +        * which previously supplied to rte_event_dev_configure().
> +        * Ignored when device is not RTE_EVENT_DEV_CAP_BURST_MODE capable.
> +        */
>         uint8_t disable_implicit_release;
>         /**< Configure the port not to release outstanding events in
>          * rte_event_dev_dequeue_burst(). If true, all events received through
> @@ -733,6 +911,14 @@ struct rte_event_port_conf {
>  rte_event_port_default_conf_get(uint8_t dev_id, uint8_t port_id,
>                                 struct rte_event_port_conf *port_conf);
>
> +int
> +rte_event_port_default_conf_get_v20(uint8_t dev_id, uint8_t port_id,
> +                               struct rte_event_port_conf_v20 *port_conf);
> +
> +int
> +rte_event_port_default_conf_get_v21(uint8_t dev_id, uint8_t port_id,
> +                                     struct rte_event_port_conf *port_conf);

Hi Timothy,

+ ABI Maintainers (Ray, Neil)

# As per my understanding, the structures can not be versioned, only
function can be versioned.
i.e we can not make any change to " struct rte_event_port_conf"

# We have a similar case with ethdev and it deferred to next release v20.11
http://patches.dpdk.org/patch/69113/

Regarding the API changes:
# The slow path changes general looks good to me. I will review the
next level in the coming days
# The following fast path changes bothers to me. Could you share more
details on below change?

diff --git a/app/test-eventdev/test_order_atq.c
b/app/test-eventdev/test_order_atq.c
index 3366cfc..8246b96 100644
--- a/app/test-eventdev/test_order_atq.c
+++ b/app/test-eventdev/test_order_atq.c
@@ -34,6 +34,8 @@
                        continue;
                }

+               ev.flow_id = ev.mbuf->udata64;
+
# Since RC1 is near, I am not sure how to accommodate the API changes
now and sort out ABI stuffs.
# Other concern is eventdev spec get bloated with versioning files
just for ONE release as 20.11 will be OK to change the ABI.
# While we discuss the API change, Please send deprecation notice for
ABI change for 20.11,
so that there is no ambiguity of this patch for the 20.11 release.

Hello Jerin,

Thank you for the review comments.

With regard to your comments regarding the fast path flow_id change, the Intel DLB hardware
is not capable of transferring the flow_id as part of the event itself. We therefore require a mechanism
to accomplish this. What we have done to work around this is to require the application to embed the flow_id
within the data payload. The new flag, #define RTE_EVENT_DEV_CAP_CARRY_FLOW_ID (1ULL << 9), can be used
by applications to determine if they need to embed the flow_id, or if its automatically propagated and present in the
received event.

What we should have done is to wrap the assignment with a conditional.  

if (!(device_capability_flags & RTE_EVENT_DEV_CAP_CARRY_FLOW_ID))
	ev.flow_id = ev.mbuf->udata64;

This would minimize/eliminate any performance impact due to the processor's branch prediction logic.
The assignment then becomes in essence a NOOP for all event devices that are capable of carrying the flow_id as part of the event payload itself.

Thanks,
Tim



Thanks,
Tim