[dpdk-dev] [PATCH 1/8] ethdev: introduce sample action for rte flow

[Andrew Rybchenko]

On 6/25/20 7:26 PM, Jiawei Wang wrote:
> When using full offload, all traffic will be handled by the HW, and
> directed to the requested vf or wire, the control application loses
> visibility on the traffic.
> So there's a need for an action that will enable the control application
> some visibility.
> 
> The solution is introduced a new action that will sample the incoming
> traffic and send a duplicated traffic in some predefined ratio to the
> application, while the original packet will continue to the target
> destination.
> 
> The packets sampled equals is '1/ratio', if the ratio value be set to 1
> , means that the packets would be completely mirrored. The sample packet
> can be assigned with different set of actions from the original packet.
> 
> In order to support the sample packet in rte_flow, new rte_flow action
> definition RTE_FLOW_ACTION_TYPE_SAMPLE and structure rte_flow_action_sample
> will be introduced.
> 
> Signed-off-by: Jiawei Wang <jiaweiw at mellanox.com>

[snip]

> @@ -2709,6 +2716,28 @@ struct rte_flow_action {
>  struct rte_flow;
>  
>  /**
> + * @warning
> + * @b EXPERIMENTAL: this structure may change without prior notice
> + *
> + * RTE_FLOW_ACTION_TYPE_SAMPLE
> + *
> + * Adds a sample action to a matched flow.
> + *
> + * The matching packets will be duplicated to a special queue or vport
> + * in the predefined probabiilty, All the packets continues processing
> + * on the default flow path.
> + *
> + * When the sample ratio is set to 1 then the packets will be 100% mirrored.
> + * Additional action list be supported to add for sampled or mirrored packets.
> + */
> +struct rte_flow_action_sample {
> +	/* packets sampled equals to '1/ratio' */
> +	const uint32_t ratio;
> +	/* sub-action list specific for the sampling hit cases */
> +	const struct rte_flow_action *actions;

This design idea does not look good to me from the very
beginning. IMHO it does not fit flow API overall design.
I mean sub-action list.

As I understand Linux iptables solves it on match level
(i.e. in pattern). E.g. "limit" extension which is basically
sampling. Sampling using meta pattern item in combination
with PASSTHRU action (to make sampling actions non-terminating
if required) is a better solution from design point of view.