Re: Boolean properties

Hi Mitch,

On Wed, Dec 7, 2011 at 12:04 AM, Mitch Bradley [off-list ref] wrote:

On 12/7/2011 7:12 AM, David Gibson wrote:

quoted

On Mon, Dec 05, 2011 at 04:29:22PM -0800, Simon Glass wrote:

quoted

Hi,

As I understand it, boolean properties in the device tree are defined
as follows:

- if the property is present, then it is true
- if the property is absent, then it is false

That's the common convention.  Although an individual binding could
define something else, and I think some already do define a
boolean~=int approach instead (e.g. the ibm,dfp cpu property).  In
some cases that's anticipating future expansion to values other than 0
and 1.

quoted

The fdtget and fdtput utilities deal with reading and changing these.
At present fdtget returns the value of the property and is not
suitable for boolean properties. fdtput allows setting of a property
value but not removing it.

To deal with boolean properties these utilities need to change. One
option is to use something like:

$ fdtget -t f /path/to/node property
(prints 'true' or 'false')
$ fdtput -t f /path/to/node property false/true
$ fdtput -t f /path/to/node property 0/1
(deletes the property for false and 0; adds it with no value for true and
1)

(Unfortunately the 'b' type is taken for byte at present, so it is
unclear to be so far what letter we should use for 'f' in the -t flag)

Hrm.  Rather than treating this as a boolean "type", I think it would
be less ambiguous to instead add options to explicitly delete a
property (either e.g. fdtput -d or a new fdtdel), and to create an
empty valued property.

quoted

But first I would like to suggest that we support boolean properties
with a value, i.e. change to:

- if the property is absent, then it is false
- if the property is present, then it is true
    - unless the property has a 0 value, iwc it is false

So the first question here is who/what exactly constitutes "we" here -
kernel, u-boot, dtc?  Using the above logic sounds like a good idea
for fdt readers on the "be liberal in what you accept" principle.  But
it's something I'd like to see Mitch Bradley and/or Segher comment on
- they'd have a better idea of whether there's existing practice that
might break with that.

Since the producer and consumer of a named property must agree on the
interpretation of its value, you can do whatever you want with a newly-named
property without breaking anything else - although you might not be able to
use existing encoding/decoding methods if you make up something new.

The use case for absence/presence boolean properties was as follows:

You develop a device and describe it via a set of properties.  Some time
later, someone develops a newer version of that device that has a new
feature that was not contemplated in the original design.  You need to
declare that the feature exists, but the device node for the older version
has no property relating to that hitherto-unknown feature.  So you invent a
new property, with the understanding that the absence of that property
indicates that the feature is not present.

That scheme does not preclude a "name"=0/1 formulation, with the
understanding that absence implies 0, but it's unclear to me that the 0/1
value adds much.  Either way, for your system to work, you have to get the
right information in the device node.

Yes that's right. Can you please look at my three reasons below and
comment on those?

The =0/1 formulation could in principle be used for 3 states, with absence
meaning either "don't know" or "error".  To my way of thinking, that just
makes things more complicated, putting the burden on the programmer to
somehow handle a vague condition.  I prefer "get it right and be done with
it".

Absence means 0 and I don't propose to change that, nor introduce a
vague condition. My suggest is purely that we allow boolean properties
to have value which is not ignored. Please see below for my reasons,
which may or may not be persuasive:

quoted

quoted

There are three reasons:
1. it is very confusing for the poor user to set a property value to 0
and have it come back as 'true' from a function that checks these
sorts of things. This is particularly true since fdt properties do not
have types, or even type hints.  The function I have in mind is
something like this (modified to fit the discussion):

/**
 * Look up a boolean property in a node and return it.
 *
 * A boolean properly is true if present in the device tree and false if
not
 * present (regardless of any property data/value).
 *
 * @param blob FDT blob
 * @param node node to examine
 * @param prop_name    name of property to find
 * @return 1 if the properly is present; 0 if it isn't present
 */
int fdtdec_get_bool(const void *blob, int node, const char *prop_name);

An alternative might be to print an error (and refuse to return a
value) when a boolean property has a value:

 * @return 1 if the properly is present; 0 if it isn't present, -1 if
it has any data (and is therefore illegal as a boolean property)

Kind of a pain for C.

quoted

2. it allows us to adjust the boolean value without requiring a change
in the size of the device tree blob (faster, easier)

3. Related to 2 it allows us to put template values in the fdt which
can then be *optionally* updated in the build system or in U-Boot to
the required value before passing to the kernel.

I presume this discussion has happened before - can anyone please
point me to the thread?

Regards,
Simon