Re: [PATCHv6 net-next 1/4] net: add a common function to compute features for upper devices

[Sabrina Dubroca <sd@queasysnail.net>]

2025-10-21, 10:46:22 +0200, Paolo Abeni wrote:
> On 10/20/25 11:10 AM, Sabrina Dubroca wrote:
> > 2025-10-17, 03:41:52 +0000, Hangbin Liu wrote:
> >> Some high level software drivers need to compute features from lower
> >> devices. But each has their own implementations and may lost some
> >> feature compute. Let's use one common function to compute features
> >> for kinds of these devices.
> >>
> >> The new helper uses the current bond implementation as the reference
> >> one, as the latter already handles all the relevant aspects: netdev
> >> features, TSO limits and dst retention.
> >>
> >> Suggested-by: Paolo Abeni <pabeni@redhat.com>
> >> Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
> > 
> > No objection to this patch/series, just a nit and some discussion below, so:
> > 
> > Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
> > 
> > 
> > [...]
> >> +/**
> >> + *	netdev_compute_master_upper_features - compute feature from lowers
> > 
> > nit: I'm slightly annoyed (that's not quite the right word, sorry)
> > that we're adding a new function to "compute features" that doesn't
> > touch netdev->features, but I can't come up with a better name
> > (the best I got was "compute extra features" and it doesn't help).
> 
> I'm not the right person to ask a good name, and I'm ok with the current
> one, but since the question is pending... what about:
> 
> netdev_{compute,update}_offloads_from_lower()
> 
> ?
> 
> As it actually updates (some of) the offloads available to the (upper)
> device?

(and the DST_RELEASE flags. at least the tso_max_* kind of fits into "offloads")

I think we can keep the current name. It's more "it kind of bothers
the pedantic part of me" than "annoyed", and we can't find a better
name, so let's ignore the pedantic part. Sorry for the noise.


> >> + *	@dev: the upper device
> >> + *	@update_header: whether to update upper device's header_len/headroom/tailroom
> >> + *
> >> + *	Recompute the upper device's feature based on all lower devices.
> >> + */
> >> +void netdev_compute_master_upper_features(struct net_device *dev, bool update_header)
> >> +{
> > [...]
> >> +	netif_set_tso_max_segs(dev, tso_max_segs);
> >> +	netif_set_tso_max_size(dev, tso_max_size);
> >> +
> >> +	netdev_change_features(dev);
> > 
> > Maybe a dumb idea: I'm wondering if we're doing this from the wrong
> > side.
> > 
> > Right now we have:
> > 
> > [some device op] -> [this new function] -> netdev_change_features -> __netdev_update_features -> ndo_fix_features
> > 
> > Would it make more sense to go instead:
> > 
> > [some device op] -> netdev_change_features -> __netdev_update_features -> ndo_fix_features -> [this new function]
> > 
> > ?
> 
> Uhmmm.... this function touches a few more things beyond dev->*features,
> calling it from ndo_fix_features() looks a bit out-of-scope.

True. And as Hangbin said, it's setting (so a bit more "update" than
"compute", as you wrote above) values whereas ndo_fix_features is just
returning a value.

So if we wanted to have this done by netdev_change_features, we'd
probably need a new ndo, or some kind of flag to tell
__netdev_update_features that this device needs the new function
called. Well, we have netif_is_bridge_master, netif_is_team_master,
netif_is_bond_master. But at this stage we don't know if update_header
should be true/false. So ndo would be cleaner, but a lot
heavier... it's probably not worth all this mess.

-- 
Sabrina