On Wed, 24 Aug 2022 10:09:55 +0200 Johannes Berg wrote:

On Tue, 2022-08-23 at 21:50 -0700, Jakub Kicinski wrote:

quoted

The @offset points to the
nest which should have contained the attribute  

I find this a bit tedious, tbh. You already kernel-side have patch 2 and
patch 3 that pass different things here.

Maybe it would be better to have this point to the _end_ of the {nest,
message} header, which - if there are any - would be equivalent to the
first sibling attribute?

Pointing at the start of a nest is easier because I can reuse the same
"attr walking" logic in user space as for finding invalid attributes 
to find the nest.

Though I guess one way or the other userspace has to have an if that
asks whether or not it's in a nest or the top-level namespace.

Hmm.

How about we just _remove_ the NLMSGERR_ATTR_MISS_NEST attribute if it's
not missing in a nested attribute? That would make sense from the naming
too:
 * NLMSGERR_ATTR_MISS_TYPE - which attribute type you missed
 * NLMSGERR_ATTR_MISS_NEST - which nesting you missed it in, _if any_


And that way the if simplifies down to something like

	if (tb[NLMSGERR_ATTR_MISS_NEST])

in the consumer too, and you don't need GENL_REQ_ATTR_CHECK() at all,
you just pass NULL to the second argument of NL_REQ_ATTR_CHECK().

Sounds good!