Message-ID: <7b0c2e0132b71b131fc9a5407abd27bc0be700ee.camel@redhat.com>
Date: Fri, 15 Dec 2023 12:06:52 +0100
From: Paolo Abeni <pabeni@...hat.com>
To: Jakub Kicinski <kuba@...nel.org>, Jiri Pirko <jiri@...nulli.us>
Cc: netdev@...r.kernel.org, anthony.l.nguyen@...el.com, 
 intel-wired-lan@...ts.osuosl.org, qi.z.zhang@...el.com, Wenjun Wu
 <wenjun1.wu@...el.com>, maxtram95@...il.com, "Chittim, Madhu"
 <madhu.chittim@...el.com>, "Samudrala, Sridhar"
 <sridhar.samudrala@...el.com>,  Simon Horman <simon.horman@...hat.com>
Subject: Re: [Intel-wired-lan] [PATCH iwl-next v4 0/5] iavf: Add devlink and
 devlink rate support'

On Thu, 2023-12-14 at 17:46 -0800, Jakub Kicinski wrote:
> On Thu, 14 Dec 2023 21:29:51 +0100 Paolo Abeni wrote:
> > Together with Simon, I spent some time on the above. We think the
> > ndo_setup_tc(TC_SETUP_QDISC_TBF) hook could be used as common basis for
> > this offloads, with some small extensions (adding a 'max_rate' param,
> > too).
> 
> uAPI aside, why would we use ndo_setup_tc(TC_SETUP_QDISC_TBF)
> to implement common basis?
> 
> Is it not cleaner to have a separate driver API, with its ops
> and capabilities?

We understand one of the end goal is consolidating the existing rate-
related in kernel interfaces.  Adding a new one does not feel a good
starting to reach that goal, see [1] & [2] ;). ndo_setup_tc() feels
like the natural choice for H/W offload and TBF is the existing
interface IMHO nearest to the requirements here.

The devlink rate API could be a possible alternative...

> > The idea would be:
> > - 'fixing' sch_btf so that the s/w path became a no-op when h/w offload
> > is enabled
> > - extend sch_btf to support max rate
> > - do the relevant ice implementation
> > - ndo_set_tx_maxrate could be replaced with the mentioned ndo call (the
> > latter interface is a strict super-set of former)
> > - ndo_set_vf_rate could also be replaced with the mentioned ndo call
> > (with another small extension to the offload data)
> > 
> > I think mqprio deserves it's own separate offload interface, as it
> > covers multiple tasks other than shaping (grouping queues and mapping
> > priority to classes)
> > 
> > In the long run we could have a generic implementation of the
> > ndo_setup_tc(TC_SETUP_QDISC_TBF) in term of devlink rate adding a
> > generic way to fetch the devlink_port instance corresponding to the
> > given netdev and mapping the TBF features to the devlink_rate API.
> > 
> > Not starting this due to what Jiri mentioned [1].
> 
> Jiri, AFAIU, is against using devlink rate *uAPI* to configure network
> rate limiting. That's separate from the internal representation.

... with a couples of caveats:

1) AFAICS devlink (and/or devlink_port) does not have fine grained, per
queue representation and intel want to be able to configure shaping on
per queue basis. I think/hope we don't want to bring the discussion to
extending the devlink interface with queue support, I fear that will
block us for a long time. Perhaps I’m missing or misunderstanding
something here. Otherwise in retrospect this looks like a reasonable
point to completely avoid devlink here.

2) My understanding of Jiri statement was more restrictive. @Jiri it
would great if could share your genuine interpretation: are you ok with
using the devlink_port rate API as a basis to replace
ndo_set_tx_maxrate() (via dev->devlink_port->devlink->) and possibly
ndo_set_vf_rate(). Note the given the previous point, this option would
still feel problematic.

Cheers,

Paolo

[1] https://xkcd.com/927/
[2] https://www.youtube.com/watch?v=f8kO_L-pDwo

Powered by blists - more mailing lists