Re: [RFC] MFD's relationship with Device Tree (OF)

[Lee Jones <lee.jones@linaro.org>]

On Tue, 09 Jun 2020, Rob Herring wrote:

Thanks for replying Rob.

> On Tue, Jun 9, 2020 at 5:01 AM Lee Jones <lee.jones@linaro.org> wrote:
> >
> > Good morning,
> >
> > After a number of reports/queries surrounding a known long-term issue
> > in the MFD core, including the submission of a couple of attempted
> > solutions, I've decided to finally tackle this one myself.
> >
> > Currently, when a child platform device (sometimes referred to as a
> > sub-device) is registered via the Multi-Functional Device (MFD) API,
> > the framework attempts to match the newly registered platform device
> > with its associated Device Tree (OF) node.  Until now, the device has
> > been allocated the first node found with an identical OF compatible
> > string.  Unfortunately, if there are, say for example '3' devices
> > which are to be handled by the same driver and therefore have the same
> > compatible string, each of them will be allocated a pointer to the
> > *first* node.
> >
> > Let me give you an example.
> >
> > I have knocked up an example 'parent' and 'child' device driver.  The
> > parent utilises the MFD API to register 3 identical children, each
> > controlled by the same driver.  This happens a lot.  Fortunately, in
> > the majority of cases, the OF nodes are also totally identical, but
> > what if you wish to configure one of the child devices with different
> > attributes or resources supplied via Device Tree, like a clock?  This
> > is currently impossible.
> >
> > Here is the Device Tree representation for the 1 parent and the 3
> > child (sub) devices described above:
> >
> >         parent {
> >                 compatible = "mfd,of-test-parent";
> >
> >                 child@0 {
> 
> Just a note, unit-address implies there is a 'reg' property. Why
> that's important below.

Right.  This is just an example to express the problem more easily.

> >                         compatible = "mfd,of-test-child";
> >                         clocks = <&clock 0>;
> >                 };
> >
> >                 child@1 {
> >                         compatible = "mfd,of-test-child";
> >                         clocks = <&clock 1>;
> >                 };
> >
> >                 child@2 {
> >                         compatible = "mfd,of-test-child";
> >                         clocks = <&clock 2>;
> >                 };
> >         };
> >
> > This is how we register those devices from MFD:
> >
> > static const struct mfd_cell mfd_of_test_cell[] = {
> >         OF_MFD_CELL("mfd_of_test_child", NULL, NULL, 0, 0, "mfd,of-test-child"),
> >         OF_MFD_CELL("mfd_of_test_child", NULL, NULL, 0, 1, "mfd,of-test-child"),
> >         OF_MFD_CELL("mfd_of_test_child", NULL, NULL, 0, 2, "mfd,of-test-child")
> > };
> >
> > ... which we pass into mfd_add_devices() for processing.
> >
> > In an ideal world.  The devices with the platform_id; 0, 1 and 2 would
> > be matched up to Device Tree nodes; child@0, child@1 and child@2
> > respectively.  Instead all 3 devices will be allocated a pointer to
> > child@0's OF node, which is obviously not correct.
> >
> > This is how it looks when each of the child devices are probed:
> >
> >  [0.708287] mfd-of-test-parent mfd_of_test: Registering 3 devices
> >  [...]
> >  [0.712511] mfd-of-test-child mfd_of_test_child.0: Probing platform device: 0
> >  [0.712710] mfd-of-test-child mfd_of_test_child.0: Using OF node: child@0
> >  [0.713033] mfd-of-test-child mfd_of_test_child.1: Probing platform device: 1
> >  [0.713381] mfd-of-test-child mfd_of_test_child.1: Using OF node: child@0
> >  [0.713691] mfd-of-test-child mfd_of_test_child.2: Probing platform device: 2
> >  [0.713889] mfd-of-test-child mfd_of_test_child.2: Using OF node: child@0
> >
> > "Why is it when I change child 2's clock rate, it also changes 0's?"
> >
> > Whoops!
> >
> > So in order to fix this, we need to make MFD more-cleverer!
> >
> > However, this is not so simple.  There are some rules we should abide
> > by (I use "should" intentionally here, as something might just have to
> > give):
> >
> >  a) Since Device Tree is designed to describe hardware, inserting
> >     arbitrary properties into DT is forbidden.  This precludes things
> >     we would ordinarily be able to match on, like 'id' or 'name'.
> >  b) As an extension to a) DTs should also be OS agnostic, so
> >     properties like 'mfd-device', 'mfd-order' etc are also not
> >     not suitable for inclusion.
> >  c) The final solution should ideally be capable of supporting both
> >     newly defined and current trees (without retroactive edits)
> >     alike.
> 
> Presumably anything current already works. If you had the above
> example already, requiring updating the DT to make it work seems fine.

"works" it a matter of opinion.  Some instances "work" out of luck.
Some "work" because they have been worked-around or an alternative
implementation sought.

For instance, 'ab8500-pwm' only has 1 DT node present, yet 3 devices
are registered via MFD.  Since MFD matches devices with DT nodes
containing identical compatible strings using first-found, all PWM
instances are assigned a pointer to the 1 existing DT node.
Fortunately in this case they all share the same clock, so it "works",
but that's clearly not the intended implementation.

> >  d) Existing properties could be used, but not abused.  For example,
> >     one of my suggestions (see below) is to use the 'reg' property.
> >     This is fine in principle but loading 'reg' with arbitrary values
> >     (such as; 0, 1, 2 ... x) which 1) clearly do not have anything to
> >     do with registers and 2) would be meaningless in other OSes/
> >     implementations, just to serve our purpose, is to be interpreted
> >     as an abuse.
> 
> Multiple instances of something implies you have some way to address
> them and 'reg' is what defines the address of something. 0,1,2,etc.
> looks suspiciously like just some kernel defined indexes, but if
> that's how things are defined in the datasheet I'm okay with them.
> 
> The one wrinkle is there's only one address space at one level, so
> gpio@0, gpio@1, pwm@0, pwm@1, etc. doesn't really work (well, it
> works, but having overlapping addresses is not good practice). Either
> we relax that in this case or we can add another level to group nodes.

All agreed.  Sounds promising.

> > Proposal 1:
> >
> > As mentioned above, my initial thoughts were to use the 'reg' property
> > to match an MFD cell entry with the correct DT node.  However, not
> > all Device Tree nodes have 'reg' properties.  Particularly true in the
> > case of MFD, where memory resources are usually shared with the parent
> > via Regmap, or (as in the case of the ab8500) the MFD handles all
> > register transactions via its own API.
> 
> Just to pick on ab8500, it should have had 'reg' property IMO. The
> 'bank' is clearly a h/w property and how you address each sub-device.
> 
> >
> > Proposal 2:
> >
> > If we can't guarantee that all DT nodes will have at least one
> > property in common to be used for matching and we're prevented from
> > supplying additional, potentially bespoke properties, then we must
> > seek an alternative procedure.
> >
> > It should be possible to match based on order.  However, the developer
> > would have to guarantee that the order in which the child devices are
> > presented to the MFD API are in exactly the same order as they are
> > represented in the Device Tree.  The obvious draw-back to this
> > strategy is that it's potentially very fragile.
> 
> I don't think we should use order.

If it's always possible to have a 'reg' property then we won't need
to.

> > Current Proposal:
> >
> > How about a collection of Proposal 1 and Proposal 2?  First we could
> > attempt a match on the 'reg' property.  Then, if that fails, we would
> > use the fragile-but-its-all-we-have Proposal 2 as the fall-back.
> 
> Yes, we should use 'reg' whenever possible. If we don't have 'reg',
> then you shouldn't have a unit-address either and you can simply match
> on the node name (standard DT driver matching is with compatible,
> device_type, and node name (w/o unit-address)). We've generally been
> doing 'classname-N' when there's no 'reg' to do 'classname@N'.
> Matching on 'classname-N' would work with node name matching as only
> unit-addresses are stripped.

Let me try and knock something up.

I'll get back to you when it's done.

-- 
Lee Jones [李琼斯]
Senior Technical Lead - Developer Services
Linaro.org │ Open source software for Arm SoCs
Follow Linaro: Facebook | Twitter | Blog

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel