Re: ynl - mutiple policies for one nested attr used in multiple cmds

[Jakub Kicinski <kuba@kernel.org>]

On Sat, 5 Aug 2023 08:33:28 +0200 Jiri Pirko wrote:
> >I'm not sure if you'll like it but my first choice would be to skip
> >the selector attribute. Put the attributes directly into the message.
> >There is no functional purpose the wrapping serves, right?  
> 
> Well, the only reason is backward compatibility.
> Currently, there is no attr parsing during dump, which is ensured by
> GENL_DONT_VALIDATE_DUMP flag. That means if user passes any attr, it is
> ignored.
> 
> Now if we allow attrs to select, previously ignored attributes would be
> processed now. User that passed crap with old kernel can gen different
> results with new kernel.
> 
> That is why I decided to add selector attr and put attrs inside, doing
> strict parsing, so if selector attr is not supported by kernel, user
> gets message back.
> 
> So what do you suggest? Do per-dump strict parsing policy of root
> attributes serving to do selection?

Even the selector attr comes with a risk, right? Not only have we
ignored all attributes, previously, we ignored the payload of the
message. So the payload of a devlink dump request could be entirely
uninitialized / random and it would work.

IOW we are operating on a scale of potential breakage here, unless
we do something very heavy handed.

How does the situation look with the known user apps? Is anyone
that we know of putting attributes into dump requests?