[PATCH net-next 2/6] software node: allow named software node to be created

[Andy Shevchenko]

On Mon, Jul 18, 2022 at 10:39:42PM +0200, Marek Behún wrote:
> On Mon, 18 Jul 2022 22:24:09 +0300
> Andy Shevchenko <andriy.shevchenko at linux.intel.com> wrote:
> 
> > On Mon, Jul 18, 2022 at 08:14:58PM +0100, Russell King (Oracle) wrote:
> > > On Mon, Jul 18, 2022 at 09:53:39PM +0300, Andy Shevchenko wrote:  
> > > > On Mon, Jul 18, 2022 at 09:43:42PM +0300, Andy Shevchenko wrote:  
> > > > > On Mon, Jul 18, 2022 at 02:27:02PM +0100, Russell King (Oracle) wrote:  
> > > > > > On Mon, Jul 18, 2022 at 03:29:52PM +0300, Andy Shevchenko wrote:  
> > > > > > > On Fri, Jul 15, 2022 at 11:48:41PM +0300, Vladimir Oltean wrote:  
> > > > > > > > So won't kobject_init_and_add() fail on namespace collision? Is it the
> > > > > > > > problem that it's going to fail, or that it's not trivial to statically
> > > > > > > > determine whether it'll fail?
> > > > > > > > 
> > > > > > > > Sorry, but I don't see something actionable about this.  
> > > > > > > 
> > > > > > > I'm talking about validation before a runtime. But if you think that is fine,
> > > > > > > let's fail it at runtime, okay, and consume more backtraces in the future.  
> > > > > > 
> > > > > > Is there any sane way to do validation of this namespace before
> > > > > > runtime?  
> > > > > 
> > > > > For statically compiled, I think we can do it (to some extent).
> > > > > Currently only three drivers, if I'm not mistaken, define software nodes with
> > > > > names. It's easy to check that their node names are unique.
> > > > > 
> > > > > When you allow such an API then we might have tracebacks (from sysfs) bout name
> > > > > collisions. Not that is something new to kernel (we have seen many of a kind),
> > > > > but I prefer, if possible, to validate this before sysfs issues a traceback.
> > > > >   
> > > > > > The problem in this instance is we need a node named "fixed-link" that
> > > > > > is attached to the parent node as that is defined in the binding doc,
> > > > > > and we're creating swnodes to provide software generated nodes for
> > > > > > this binding.  
> > > > > 
> > > > > And how you guarantee that it will be only a single one with unique pathname?
> > > > > 
> > > > > For example, you have two DSA cards (or whatever it's called) in the SMP system,
> > > > > it mean that there is non-zero probability of coexisting swnodes for them.
> > > > >   
> > > > > > There could be several such nodes scattered around, but in this
> > > > > > instance they are very short-lived before they are destroyed, they
> > > > > > don't even need to be published to userspace (and its probably a waste
> > > > > > of CPU cycles for them to be published there.)
> > > > > > 
> > > > > > So, for this specific case, is this the best approach, or is there
> > > > > > some better way to achieve what we need here?  
> > > > > 
> > > > > Honestly, I don't know.
> > > > > 
> > > > > The "workaround" (but it looks to me rather a hack) is to create unique swnode
> > > > > and make fixed-link as a child of it.
> > > > > 
> > > > > Or entire concept of the root swnodes (when name is provided) should be
> > > > > reconsidered, so somehow we will have a uniqueness so that the entire
> > > > > path(s) behind it will be caller-dependent. But this I also don't like.
> > > > > 
> > > > > Maybe Heikki, Sakari, Rafael can share their thoughts...
> > > > > 
> > > > > Just for my learning, why PHY uses "fixed-link" instead of relying on a
> > > > > (firmware) graph? It might be the actual solution to your problem.
> > > > > 
> > > > > How graphs are used with swnodes, you may look into IPU3 (Intel Camera)
> > > > > glue driver to support devices before MIPI standardisation of the
> > > > > respective properties.  
> > > > 
> > > > Forgot to say (yes, it maybe obvious) that this API will be exported,
> > > > anyone can use it and trap into the similar issue, because, for example,
> > > > of testing in environment with a single instance of the caller.  
> > > 
> > > I think we're coming to the conclusion that using swnodes is not the
> > > correct approach for this problem, correct?  
> > 
> > If I understand the possibilities of the usage in _this_ case, then it's
> > would be problematic (it does not mean it's incorrect). It might be due to
> > swnode design restrictions which shouldn't be made, I dunno. That' why
> > it's better to ask the others for their opinions.
> > 
> > By design swnode's name makes not much sense, because the payload there
> > is a property set, where _name_ is a must.
> > 
> > Now, telling you this, I'm questioning myself why the heck I added names
> > to swnodes in the intel_quark_i2c_gpio driver...
> 
> 1. the way we use this new named swnode (in patch 5/6 of this series) is
>    that it gets destroyed immediately after being parsed, so I don't
>    think there will be collisions in the namespace for forseeable future
> 
>    also, we first create an unnamed swnode for port and only then
>    fixed-link swnode as a child.
> 
>       new_port_fwnode = fwnode_create_software_node(port_props, NULL);
>       ...
>       fixed_link_fwnode =
>         fwnode_create_named_software_node(fixed_link_props,
>                                           new_port_fwnode, "fixed-link");
> 
>    so there shouldn't be a name collision, since the port node gets a
>    unique name, or am I misunderstanding this?

This is not problem, but what I was talking about is how to guarantee this
hierarchy? See what I answered to RNK.

> 2. even if there was a problem with name collision, I think the place
>    that needs to be fixed is swnode system. What use are swnodes if
>    they cannot be used like this?

Precisely, that's why I don't want to introduce an API that needs to be fixed.

-- 
With Best Regards,
Andy Shevchenko