-----Original Message-----
From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
Sent: Monday, March 6, 2017 8:07 PM
To: Dumitrescu, Cristian <redacted>
Cc: dev@dpdk.org; jerin.jacob@caviumnetworks.com;
balasubramanian.manoharan@cavium.com; hemant.agrawal@nxp.com;
shreyansh.jain@nxp.com; Wiles, Keith [off-list ref]; Richardson,
Bruce [off-list ref]
Subject: Re: [PATCH v3 2/2] ethdev: add hierarchical scheduler API

2017-03-06 16:59, Dumitrescu, Cristian:

quoted

From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]

quoted

2017-03-04 01:10, Cristian Dumitrescu:

quoted

This patch introduces the generic ethdev API for the traffic manager
capability, which includes: hierarchical scheduling, traffic shaping,
congestion management, packet marking.

We already have some API for QoS. Why integrating them in ethdev?
ethdev is an interface for networking drivers.
I think the QoS has nothing to do with drivers.
If there are some operations to offload in drivers, please identify them
and let's add the operations to ethdev.

The reason to add to ethdev is because QoS traffic

management/hierarchical scheduling is just another TX offload for Ethernet
devices. This TX offload is present in NICs, NPUs and SoCs from Broadcom,
Cavium, Intel, Mellanox, Netronome, NXP, others.

quoted

The API we currently have in DPDK (librte_sched) is great, but it refers to

an implementation for a fixed set of features for a BRAS-like hierarchy. The
current abstraction layer proposal is intended to support pretty much any
hierarchy and traffic management features such as hierarchical scheduling,
traffic shaping, congestion management, marking under the same API. It
targets pretty much any implementation, either HW, SW or hybrid; it does
support the existing librte_sched library feature set, but it is not limited to it.

OK I better understand now.
You should add this level of explanation in your patch.

However I am reluctant to add an API if there is no user.
I think we should wait to have at least one existing driver implementing
this API before integrating it.
It was the approach of eventdev which has a dedicated next- tree.

The next-tree solution could work, but IMO is not the best for this case, as this is purely driver development. This is just a TX offload feature that is well understood, as opposed to a new library with a huge design effort required like eventdev.

I think we are reasonably close to get agreement on the API from Cavium, Intel and NXP. When this is done, how about including it in DPDK with the experimental tag attached to it until several drivers implement it?

From Intel side, there are solid plans to implement it for ixgbe and i40e drivers in next DPDK releases, I am CC-ing Tim to confirm this. On Cavium and NXP side, Jerin and Hemant can comment on the plans to implement this API.