On Mon, 05 Oct 2020 21:25:57 +0200 Johannes Berg wrote:
> On Mon, 2020-10-05 at 12:22 -0700, Jakub Kicinski wrote:
>
> > > > + if (value & ~(u64)pt->mask) {
> > > > + NL_SET_ERR_MSG_ATTR(extack, nla, "reserved bit set");
> > > > + return -EINVAL;
> > >
> > > You had an export of the valid bits there in ethtool, using the cookie.
> > > Just pointing out you lost it now. I'm not sure I like using the cookie,
> > > that seems a bit strange, but we could easily define a different attr?
> > >
> > > OTOH, one can always query the policy export too (which hopefully got
> > > wired up) so it wouldn't really matter much.
> >
> > My thinking is that there are no known uses of the cookie, it'd only
> > have practical use to test for new flags - and we're adding first new
> > flag in 5.10.
>
> Hm, wait, not sure I understand?
>
> You _had_ this in ethtool, but you removed it now. And you're not
> keeping it here, afaict.
>
> I can't disagree on the "no known uses of the cookie" part, but it feels
> odd to me anyway - since that is just another netlink message (*), you
> could as well add a "NLMSGERR_ATTR_VALID_FLAGS" instead of sticking the
> data into the cookie?
>
> But then are you saying the new flags are only in 5.10 so the policy
> export will be sufficient, since that's also wired up now?
Right, I was commenting on the need to keep the cookie for backward
compat.
My preference is to do a policy dump to check the capabilities of the
kernel rather than shoot messages at it and then try to work backward
based on the info returned in extack.
> johannes
>
> (*) in a way - the ack message has the "legacy" fixed part before the
> attrs, of course
>
