Message-ID: <20180610133611.51cf6651@tock>
Date:   Sun, 10 Jun 2018 13:36:11 +0200
From:   Alban <albeu@...e.fr>
To:     Srinivas Kandagatla <srinivas.kandagatla@...aro.org>
Cc:     Aban Bedel <albeu@...e.fr>, linux-kernel@...r.kernel.org,
        Rob Herring <robh+dt@...nel.org>,
        Mark Rutland <mark.rutland@....com>,
        David Woodhouse <dwmw2@...radead.org>,
        Brian Norris <computersforpeace@...il.com>,
        Boris Brezillon <boris.brezillon@...e-electrons.com>,
        Marek Vasut <marek.vasut@...il.com>,
        Richard Weinberger <richard@....at>,
        Cyrille Pitchen <cyrille.pitchen@...ev4u.fr>,
        devicetree@...r.kernel.org, linux-mtd@...ts.infradead.org
Subject: Re: [PATCH v3 1/3] nvmem: Update the OF binding to use a subnode
 for the cells list

On Sun, 10 Jun 2018 11:32:36 +0100
Srinivas Kandagatla <srinivas.kandagatla@...aro.org> wrote:

> On 08/06/18 18:07, Alban wrote:
> > On Fri, 8 Jun 2018 12:34:12 +0100
> > Srinivas Kandagatla <srinivas.kandagatla@...aro.org> wrote:
> >   
> ...
> > 
> > I looked into this. It would work fine for the cells but not so nicely
> > for the nvmem device API. The phandle for the nvmem device would have
> > to reference the node passed here and not the real device. We would end
> > up with a DT like this:
> > 
> > flash@0 {
> > 	compatible = "mtd";
> > 	...
> > 	nvmem_dev: nvmem-cells {
> > 		compatible = "nvmem-cells";
> > 		...
> >          };
> > };
> > 
> > other-device@10 {
> > 	...
> > 	nvmem = <&nvmem_dev>;
> > };
> > 
> > Now if there is no cell defined we have this empty child node that make
> > very little sense, it is just there to accommodate the nvmem API.
> >   
> NO. This just looks fine!
> nvmem-cells is the nvmem provider node without which you would not have 
> any provider instance.
> All this looks as expected!
> Am not sure what is the problem here!

The problem is that DT should represent the hardware, not the OS API.
What should be represented is that other drivers can access data stored
on this device. It is my understanding that this wouldn't be an
acceptable binding as the nvmem provider node would only exists because
of how the NVMEM API currently works, a correct binding would just
directly reference the storage device without this extra node.

> > What I would suggest now is to just change the wording. We don't
> > deprecate the current binding, but we extend it to allow grouping the
> > cells in a child node if required. The code to support this is trivial,
> > (4 lines including error handling) so even if we expect very few
> > bindings to make use of it it is not going to be maintenance drag.
> > That would look like this:  
> 
> > 
> > diff --git a/Documentation/devicetree/bindings/nvmem/nvmem.txt b/Documentation/devicetree/bindings/nvmem/nvmem.txt
> > index fd06c09..085d042 100644
> > --- a/Documentation/devicetree/bindings/nvmem/nvmem.txt
> > +++ b/Documentation/devicetree/bindings/nvmem/nvmem.txt
> > @@ -19,7 +19,10 @@ Optional properties:
> >   
> >   = Data cells =
> >   These are the child nodes of the provider which contain data cell
> > -information like offset and size in nvmem provider.
> > +information like offset and size in nvmem provider. Alternatively the data
> > +cells can be grouped in a node that has a compatible property set to
> > +"nvmem-cells".
> > +
> >   
> >   Required properties:
> >   reg:   specifies the offset in byte within the storage device.
> > diff --git a/drivers/nvmem/core.c b/drivers/nvmem/core.c
> > index 4e94a78..3e1369c 100644
> > --- a/drivers/nvmem/core.c
> > +++ b/drivers/nvmem/core.c
> > @@ -859,6 +859,14 @@ struct nvmem_cell *of_nvmem_cell_get(struct device_node *np,
> >          if (!nvmem_np)
> >                  return ERR_PTR(-EINVAL);
> >   
> > +       /* bindings that already have anonymous child nodes can instead put
> > +        * their cells in a child node with an nvmem-cells compatible. */
> > +       if (of_device_is_compatible(nvmem_np, "nvmem-cells")) {
> > +               nvmem_np = of_get_next_parent(nvmem_np);
> > +               if (!nvmem_np)
> > +                       return ERR_PTR(-EINVAL);
> > +       }
> > +
> >          nvmem = __nvmem_device_get(nvmem_np, NULL, NULL);
> >          of_node_put(nvmem_np);
> >          if (IS_ERR(nvmem))
> > 
> > What about it?  
> Let me repeat what I have said in my previous emails:
> 
> Having a subnode still sounds very fragile to me,
> and this could be much specific case of MTD provider. We might have
> instances where this could be sub-sub node of the the original provider
> for other providers. Also I do not want to bring in Provider specifics 
> layout into nvmem bindings.
> 
> I can not make myself any clearer than this, Its going to be a NAK from 
> my side for the above reasons!

I fully understand you concern but I think they are overblown. First I
highly doubt that more layouts will ever be needed, using a compatible
string pretty much guarantee that we won't clash with another binding.
Furthermore even if you consider this extension "MTD specific" the
amount of code is very small, non intrusive and only run once at
registration time. I would understand if we were talking about pages of
code nesting in various place, but not really when it is a single small
if block with an obvious condition. And finally I don't see that as MTD
specific as any other device could use this feature without any code
change.

> Also, patch I shared should give enough flexibility to various range of 
> providers which have different child node layouts without touching the 
> nvmem bindings. If it works please use it.

It works from a code POV but it break the basic guidelines of DT
bindings. As I want to have this done, I'm going to do a patch as you
want, but I see a high chance that the binding is going to be rejected
by the DT maintainers and we'll be back here again.

Alban

Content of type "application/pgp-signature" skipped

Powered by blists - more mailing lists