Hi,

Thank you for your e-mail.

I considered using devlink-rate, and it seems like a good fit. However 
we would also

need support for rate-limiting for individual queues on the VF. 
Currently we have

two types of rate objects in devlink-rate: leaf and node. Would adding a 
third one - queue be accepted ?

Also we might want to add some other object rate parameters to currently 
existing ones, for

example 'priority'.


If this sounds acceptable I will work on the patch and submit it as 
soon, as it's ready.


Thanks,

Michał


On 7/6/2022 12:15 AM, Jakub Kicinski wrote:

On Mon,  4 Jul 2022 13:45:13 +0200 Michal Wilczynski wrote:

quoted

If we were to follow normal flow, we would now use tc-filter family of
commands to direct types of interesting traffic to the correct nodes.
That is NOT the case in this implementation. In this POC, meaningful
classid number identifies scheduling node. Number of qdisc handle is a
queue number in a PF space. Reason for this - we want to support ALL
queues on the card including SR-IOV ones that are assigned to VF
netdevs.

Have you looked at the devlink rate API? It should do what you need.
Dima has been working on extending that API recently you may want to
compare notes with him as well:

https://lore.kernel.org/all/20220620152647.2498927-1-dchumak@nvidia.com/ (local)