On Thu, Aug 08, 2013 at 07:34:28PM +0100, Stephen Warren wrote:

On 08/08/2013 03:11 AM, Mark Rutland wrote:

quoted

On Wed, Aug 07, 2013 at 05:12:12PM +0100, Stephen Warren wrote:

quoted

On 08/07/2013 08:07 AM, Mark Rutland wrote:

quoted

On Tue, Aug 06, 2013 at 12:00:50PM +0100, Pawel Moll wrote:

quoted

On Wed, 2013-07-31 at 16:56 +0100, Mark Rutland wrote:

quoted

quoted

Ah, I guess the question more: Do we want generic bindings that describe
the low-level details of the HW in a completely generic fashion so that
new HW can be supported with zero kernel changes, or do we want a simple
driver with a lookup table that maps from compatible value to the HW
configuration? One of the potential benefits of DT is to be able to
support new HW without code changes, although perhaps that's more
typically considered in the context of new boards rather than new IP
blocks or SoCs.

... or FPGAs that can be synthesized with random collection of standard
IP blocks. With Xilinx's Zynq and Altera's SOCFPGA this is getting
simpler and simpler...

quoted

I think that going forward it would be better to have have a compatible
string per different device. As Olof pointed out, we're leaking the way
we currently handle things in Linux into the binding, rather than
precisely describing the hardware (with a unique compatible string). I'm
not sure if this is much better than embedding a bytecode describing how
to poke devices.

This really isn't leaking information about how Linux handles the
device. It's simply saying that there is a GPIO controller whose HW is
able to be described by a simple/generic binding, and that binding
provides full details of the register layout. Other OSs can handle this
differently; see below ...

I worry that it doesn't provide a full description, but rather a
description of the subset of the hardware used by the driver.

I don't see that as a problem.

A particular DT file provides a description of an interface to HW. To my
mind, if that particular DT doesn't describe everything about a
particular HW module (e.g. some advanced feature can't be enabled),
that's basically equivalent to not describing aspect of the board/system
(so e.g. some I2C device isn't represented at all, and hence some
temperature probe can't be monitored).

My concern is simply that if people can get basic functionality with the
generic driver they won't bother describing the details of the more
specific binding, and you'll never be able to use the board with a
better driver without having to modify the dt later.

I guess the question boils down to how much we care about that
situation.

I think both the simple(!) simple-gpio interface to HW, and the more
complex tegra20-gpio interface are both equally valid interfaces; they
simply allow a different set of features to be accessed.

They're certianly both valid. I fear they may become mutually exclusive
(people only define one or the other, and the bindings drift and become
incompatible).

quoted

quoted

...

quoted

quoted

Frankly speaking I don't know where to draw the line, but I feel that in
this particular case - a "generic" GPIO binding - is worth the effort.
SOCs are literally littered with control registers driving random bits.
My favourite example - Versatile Express ;-) - have random registers
representing things like LEDs or MMC status lines. Depending on the
motherboard/FPGA version they can live in different places. And yes, I
can have a Versatile Express "platform" driver registering different set
of them depending on the particular variant of the FPGA bitfile. Or try
to represent them in the tree...

I worry that going down that route encourages bindings to describe a
single way to use a given device, rather than describing the device
itself and allowing the OS to decide how to use it. This limits what we
can do in future, and I worry about how we can handle quirks sanely if
we describe devices in this way.

Well, each DT node that uses this binding must still have a compatible
property that fully defines the exact instance of the HW. In other
words, assuming this binding worked fine for Tegra, the DT must contain:

compatible = "nvidia,tegra20-gpio", "simple-gpio";

and not just:

compatible = "simple-gpio";

In that case, an OS could choose to match on "nvidia,tegra20-gpio" and
ignore most of the other properties to instantiate a more "custom"
driver, or to enable HW-specific quirks.

In that case, does the "nvidia,tegra20-gpio" require any extra reg
fields for registers not used by the "simple-gpio" binding? If peopel
are given a shortcut, I don't see why they'd bother to describe the
hardware they're not using.

Do you mean entries in the reg field itself, or the various
dat/set/clr/... properties that describe the register layout?

I mean entries. If people describe the "dat", "set," "clr" registers
from the simple-gpio binding, they may also need some
"nvidia,tegra20-gpio"-specific reg entry describing the whole register
set (which will likely overlap). Describing things twice seems odd to
me.

I would expect the reg property for a DT node to be entirely complete in
all cases, even if writing a simple-gpio node when the HW could instead
be later described as a tegra20-gpio node. In other words, if the GPIO
HW block has 2 sets of disjoint registers, then the DT should include
those both in reg even if features accessible through the simple-gpio
view don't need the second bank.

If that's the case, we need to ensure that the bindings for any more
specific binding compatible with simple-gpio is a strict superset of the
simple-gpio binding (and this needs to be clear from the start in the
binding document for simple-gpio). Presumably, any more specific binding
*must* have one or more reg-names entries for its more useful data.

Re: dat/set/clr/... - yes, I imagine the whole point of later switching
to a more specific binding would be to enable more features, which would
likely in turn require the use of more registers. However, I would
expect the tegra20-gpio binding to embody the specific set of registers
that are present in HW. In other words, the simple-gpio binding would
require dat/set/clr/... properties and drivers would solely use those to
know how to access the various registers. However, use of the more
specific tegra20-gpio interface would entail the user having pre-defined
explicit knowledge of the register layout, and hence it would entirely
ignore dat/set/clr/... properties, and use its own built-in layout
knowledge.

Ok, that sounds like what I was getting at above. I'd want a push-back
on simple-gpio usage without a more-specific binding also being
documented.

quoted

What about the case where some mfd IP block can act as a gpio
controller, compatible with simple-gpio, and also provides some other
functionality requiring a separate driver? I suspect people will
describe this as two devices, mirroring the Linux driver model, rather
than describing the hardware itself.

Are the GPIO registers truly an entirely separate HW block? If so,
simple-gpio v.s. something else shouldn't matter.

If the GPIO registers are co-mingled with other features, then I think
that it'd be legal for the DT to contain either:

a)

Just a simple-gpio block. The other features would not be available.

b)

A more precise compatible value, thus allowing SW to expose all the
features.

The one issue here is that perhaps we could never "re-purpose" a plain
simple-gpio node simply by binding a different driver to it; all the
other properties could be missing. As such, perhaps when using
simple-gpio we shouldn't actually include the more precise entry in the
compatible property.

The possible solutions would be:

a) Very carefully craft every GPIO-related binding so that all
properties beyond what simple-gpio provides were optional, and hence a
pure simple-gpio node would work fine with the more complex driver. This
could be difficult.

b) Make sure that the DT already contained the union of the properties
required by simple-gpio and the more complex binding from the start. In
this case, there might not be much point doing simple-gpio; just use the
more advanced binding from the start.

That would be my preference. That also caters for the case of a device
that happens to be usable with more than one generic binding
simultaneously.

I guess this boils down to: When an old DT is used with a new kernel, is
the minimum expectation that all previous functionality will still
operate correctly, *or* that no DT changes are required to enable any
new functionality that the new kernel could provide on that hardware?

I'd tend to lean towards new kernels maintaining at least old
functionality, rather than requiring that no DT changes are required to
enable new functionality.

That certainly makes sense. I'd also like to aim for new dts maintaining
the old functionality on old kernels, or we're still stuck with a
dtb<->kernel version requirement.

Thanks,
Mark.