Re: [net-next, v2, 3/7] net: dsa: free skb->cb usage in core driver

[Vladimir Oltean]

On Fri, May 07, 2021 at 11:26:32AM +0000, Y.b. Lu wrote:
> > From: Tobias Waldekranz <tobias@xxxxxxxxxxxxxx>
> > On Mon, Apr 26, 2021 at 17:37, Yangbo Lu <yangbo.lu@xxxxxxx> wrote:
> > > Free skb->cb usage in core driver and let device drivers decide to use
> > > or not. The reason having a DSA_SKB_CB(skb)->clone was because
> > > dsa_skb_tx_timestamp() which may set the clone pointer was called
> > > before p->xmit() which would use the clone if any, and the device
> > > driver has no way to initialize the clone pointer.
> > >
> > > Although for now putting memset(skb->cb, 0, 48) at beginning of
> > > dsa_slave_xmit() by this patch is not very good, there is still way to
> > > improve this. Otherwise, some other new features, like one-step
> >
> > Could you please expand on this improvement?
> >
> > This memset makes it impossible to carry control buffer information from
> > driver callbacks that run before .ndo_start_xmit, for
> > example .ndo_select_queue, to a tagger's .xmit.
> >
> > It seems to me that if the drivers are to handle the CB internally from now on,
> > that should go for both setting and clearing of the required fields.
>
> For the timestamping case, dsa_skb_tx_timestamp may not touch
> .port_txtstamp callback, so we had to put skb->cb initialization at
> beginning of dsa_slave_xmit.
> To avoid breaking future skb->cb usage you mentioned, do you think we
> can back to Vladimir's idea initializing only field required, or even
> just add a callback for cb initialization for timestamping?

I would like to avoid overengineering, which a callback for skb->cb
initialization would introduce, given the needs we have now.

FWIW, we may not even be able to touch skb->cb in .ndo_select_queue for
Tobias's use case, that discussion is here:
https://patchwork.kernel.org/project/netdevbpf/patch/20210426170411.1789186-7-tobias@xxxxxxxxxxxxxx/