Re: [patch net-next v2] tools: ynl: introduce option to process unknown attributes or types

[Jakub Kicinski <kuba@kernel.org>]

On Tue, 17 Oct 2023 08:18:13 +0200 Jiri Pirko wrote:
> Tue, Oct 17, 2023 at 02:59:03AM CEST, kuba@kernel.org wrote:
> >On Mon, 16 Oct 2023 13:02:22 +0200 Jiri Pirko wrote:  
> >> +class FakeSpecAttr:
> >> +    def __init__(self, name):
> >> +        self.dict = {"name": name, "type": None}
> >> +        self.is_multi = False
> >> +
> >> +    def __getitem__(self, key):
> >> +        return self.dict[key]
> >> +
> >> +    def __contains__(self, key):
> >> +        return key in self.dict  
> >
> >Why the new class? Why not attach the NlAttr object directly?  
> 
> It's not NlAttr, it's SpecAttr. And that has a constructor with things I
> cannot provide for fake object, that's why I did this dummy object.

Just to be able to do spec["type"] on it?

There is an if "ladder", just replace the first

	if attr_spec["type"] == ...

with
	if attr_spec is None:
		# your code
	elif attr_spec["type"] == ...

hm?

> >I have an idea knocking about in my head to support "polymorphic"
> >nests (nests where decoding depends on value of another attr,
> >link rtnl link attrs or tc object attrs). The way I'm thinking 
> >about doing it is to return NlAttr / struct nla_attr back to the user.
> >And let the users call a sub-parser of choice by hand.  
> 
> Sounds parallel to this patch, isn't it?

I'm just giving you extra info to explain my thinking.
Given how we struggle to understand each other lately :S

> >So returning a raw NlAttr appeals to me more.  
> 
> Wait, you suggest not to print out attr.as_bin(), but something else?

Yea, it should not be needed. NlAttr has a __repr__ which *I think*
should basically do the same thing? Or you may need to call that
__repr__ from __str__, I don't know what PrettyPrinter uses internally

> >> +                if not self.process_unknown:
> >> +                    raise Exception(f'Unknown {attr_spec["type"]} with name {attr_spec["name"]}')
> >> +                if attr._type & Netlink.NLA_F_NESTED:
> >> +                    subdict = self._decode(NlAttrs(attr.raw), None)
> >> +                    decoded = subdict
> >> +                else:
> >> +                    decoded = attr.as_bin()  
> >
> >Again, I wouldn't descend at all.  
> 
> I don't care that much. I just thought it might be handy for the user to
> understand the topology. Actually, I found it quite convenient already.
> It's basically a direct dump. What is the reason not to do this exactly?

No strong reason but you need to rewrite it to at least not access
attr._type directly.

I have a weak preference for putting this code in NlAttr's __repr__,
could be more broadly useful?