Re: [PATCH] net: mark skb NAPI IDs independently of busy polling

From: Ding Meng

Date: Sun Nov 15 2026 - 22:23:26 EST


On Fri, Nov 13, 2026 at 09:59:17PM +0100, Gerhard Engleder wrote:
> On 13.11.26 09:29, Paul Moses wrote:
> >
> > >
> > > Then, drivers that use deferred timestamp conversion (e.g. igc, engleder) must
> > > specify a busy poll dependency if they want to obtain the original net
> > > device upon consuming the skb. However, this dependency is far from
> > > obvious, and annotating it for future features of this kind will be very
> > > hard to discover.
> >
> > The "deferred timestamp conversion" was incomplete is what I'm saying.
> >
> > This started with 97dc7cd92ac6, 0abb62b68252 and 069b142f5819 were built upon
> > it, so maybe Gerhard should decide the path forward?
>
> The deferred timestamp conversion is a niche feature for TSN to solve
> the working clock / global time problem in TSN networks. IMO there is
> no need to make it working as soon as igc or tsnep drivers are enabled.
> It would be ok to fall back to normal vclock behavior like all other
> drivers if CONFIG_NET_RX_BUSY_POLL=n. For me it is ok that NETCONSOLE
> is not possible for PREEMPT_RT in combination with this feature.
>
> There are already so many things to consider to get TSN working. To
> not enable NETCONSOLE on PREEMPT_RT for deferred timestamp conversion
> would be just one of many.
>
> @Ding Meng: Do you actually need the NETCONSOLE in this combination?
> On productive systems likely not these days due to CRA.
>
> Gerhard
OK, understand. Thanks for your comments, Gerhard. Actuallly, we solved
this issue by disabling netconsole and enabling busy poll. This is
indeed a quite rare scenario, but when it happens, it is a bit difficult
to figure out that this was the root cause.

Ding Meng