Re: [PATCH v2 1/6] device property: Helper to match multiple connections

[Andy Shevchenko <andriy.shevchenko@linux.intel.com>]

On Sun, Feb 20, 2022 at 08:55:10PM -0800, Bjorn Andersson wrote:
> On Sun 20 Feb 03:16 PST 2022, Andy Shevchenko wrote:
> > On Fri, Feb 18, 2022 at 11:00:45AM -0800, Bjorn Andersson wrote:
> > > On Wed 09 Feb 04:30 PST 2022, Andy Shevchenko wrote:
> > > > On Mon, Feb 07, 2022 at 07:19:39PM -0800, Bjorn Andersson wrote:

...

> > > > > +int fwnode_connection_find_matches(struct fwnode_handle *fwnode,
> > > > > +				   const char *con_id, void *data,
> > > > > +				   devcon_match_fn_t match,
> > > > > +				   void **matches, unsigned int matches_len)
> > > > > +{
> > > > > +	unsigned int count;
> > > > > +
> > > > > +	if (!fwnode || !match || !matches)
> > > > 
> > > > !matches case may be still useful to get the count and allocate memory by
> > > > caller. Please, consider this case.
> > > 
> > > As discussed in previous version, and described in the commit message,
> > > the returned value of "match" is a opaque pointer to something which
> > > has to be passed back to the caller in order to be cleaned up.
> > > 
> > > E.g. the typec mux code returns a pointer to a typec_mux/switch object
> > > with a refcounted struct device within, or an ERR_PTR().
> > > 
> > > So unfortunately we can must gather the results into matches and pass it
> > > back to the caller to take consume or clean up.
> > 
> > It's fine. You have **matches, means pointer of an opaque pointer.
> > What I'm talking about is memory allocation for and array of _pointers_.
> > That's what caller very much aware of and can allocate on heap. So, please
> > consider this case.
> 
> I'm sorry, but I'm not sure what you're looking for.
> 
> 
> I still interpret your comment as that it would be nice to be able to do
> something like:
> 
> count = fwnode_connection_find_matches(fwnode, "orientation-switch",
> 				       NULL, typec_switch_match, NULL, 0);
> 
> based on the returned value the caller could allocate an array of
> "count" pointers and then call the function again to actually fill out
> the count elements.

Yes, that's what I want from the generic fwnode APIs.
(Keyword: generic)

> The problem with this is that, typec_switch_match() does:

As you stated, the problem is in the typec_switch_match(). So, it's not related
to the fwnode, but how you are using it.

> void *typec_switch_match(fwnode, id, data) {
> 	struct device *dev = find_struct_device(fwnode, id);
> 	if (!dev)
> 		return NULL;
> 	get_device(dev);
> 	return container_of(dev, struct typec_switch, dev);
> }
> 
> So if we call the match function and if that finds a "dev" it will
> return a struct typec_switch with a refcounted struct device within.

fwnode (as being an abstraction on top of the others) has no knowledge
about this. And more important should not know that.

> We can see if that's NULL or not and will be able to return a "count",
> but we have no way of releasing the reference acquired - we must return
> the void pointer back to the client, so that it can release it.

The caller (if it wants to!) may create different callbacks for count and real
matching, no?

> My claim is that this is not a problem, because this works fine with any
> reasonable size of fwnode graphs we might run into - and the client will
> in general have a sense of the worst case number of matches (in this
> series its 3, as there's 3 types of lanes that can be switched/muxed
> coming out of a USB connector).

-- 
With Best Regards,
Andy Shevchenko