Generating generic information for 'type: phandle-array' properties in
edtlib was difficult due to defining phandle-array as just a list of
phandles and numbers. To make sense of a phandle-array property like
'pwms', you have to know that #pwm-cells is expected to appear on
each referenced controller, and that the binding for the controller has
a #cells.
Because of this, handling of various 'type: phandle-array' properties
was previously hardcoded in edtlib and exposed through properties like
Node.pwms, instead of through the generic Node.props (though with a lot
of shared code).
In practice, it turns out that all 'type: phandle-array' properties in
Zephyr work exactly the same way: They all have names that end in -s,
the 's' is removed to derive the name of related properties, and they
all look up #cells in the binding for the controller, which gives names
to the data values.
Strengthen the definition of 'type: phandle-array' to mean a property
that works exactly like the existing phandle-array properties (which
also means requiring that the name ends in -s). This removes a ton of
hardcoding from edtlib and allows new 'type: phandle-array' properties
to be added without making any code changes.
If we ever need a property type that's a list of phandles and numbers
but that doesn't follow this scheme, then we could add a separate type
for it. We should check if the standard scheme is fine first though.
The only property type for which no information is generated is now
'compound'.
There's some inconsistency in how we generate identifiers for clocks
compared to other 'type: phandle-array' properties, so keep
special-casing them for now in gen_defines.py (see the comment in
write_clocks()).
This change also enabled a bunch of other simplifications, like reusing
the ControllerAndData class for interrupts.
Piggyback generalization of *-map properties so that they work for any
phandle-array properties. It's now possible to have things like
'io-channel-map', if you need to.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
API oversight. This was meant to be there all along together with
Device.parent, for navigating the devicetree, but since a need for it
never came up in gen_defines.py, it got overlooked.
Devices are just devicetree nodes augmented with binding information and
some interpretation of devicetree properties. I wonder if the name
should be changed to something like edtlib.Node to make that clearer.
Calling something like a flash partition a "device" is a bit weird, as
Galak pointed out.
I think I went with Device originally to avoid confusion with
dtlib.Node, but since edtlib users don't directly interact with dtlib,
it might not be that confusing in practice.
Piggyback some documentation clarifications.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Deprecate 'sub-node:' and add a more general 'child-binding:' mechanism
to bindings. Keep supporting 'sub-node:', but print a deprecation
warning when it's used.
Like 'sub-node:', 'child-binding:' gives a binding to child nodes, but
the binding is required to be a complete binding, and is treated (and
checked) like a normal binding.
'child-binding:' can in turn contain another 'child-binding:', up to any
number of levels. This is automatic from treating it like a normal
binding, and from the code initializing parent Devices before child
Devices.
This lets nodes give bindings to grandchildren.
For example, take this devicetree fragment:
parent {
compatible = "foo";
child-1 {
grandchild-1 {
...
};
grandchild-2 {
...
};
};
child-2 {
grandchild-3 {
...
};
};
};
The binding for 'foo' could provide bindings for grandchild-1/2/3 like
this:
compatible: "foo"
# Binding for children
child-binding:
title: ...
description: ...
...
# Binding for grandchildren
child-binding:
title: ...
description: ...
properties:
...
Due to implementation issues with the old devicetree scripts, only two
levels of 'child-binding:' is supported for now. This limitation will go
away in Zephyr 2.2.
Piggyback shortening 'description:' and 'title:' in some bindings that
provide child bindings. This makes the generated header a bit neater.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Instead of
child:
bus: foo
parent:
bus: bar
, have
child-bus: foo
parent-bus: bar
'bus' is the only key that ever appears under 'child' and 'parent'.
Support the old keys for backwards compatibility, with a deprecation
warning if they're used.
Also add 'child/parent-bus' tests to the edtlib test suite. It was
untested before.
I also considered putting more stuff under 'child' and 'parent', but
there's not much point when there's just a few keys I think. Top-level
stuff is cleaner and easier to read.
I'm planning to add a 'child-binding' key a bit later (like 'sub-node',
but more flexible), and child-* is consistent with that.
Also add an unrelated test-bindings/grandchild-3.yaml that was
accidentally left out earlier.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
When foo.yaml set some property 'required: true' and bar.yaml set the
same property 'required: false', the check for changing
'required: false' to 'required: true' would raise an error for
include: [bar.yaml, foo.yaml]
(with that particular order due to implementation details).
The order files are included in shouldn't matter. To fix it, change the
logic so that 'required' values are ORed together between included files
(so that 'required: true' is always respected), and remove the
'required' true-to-false check when merging included files.
Keep the true-to-false check when merging the (merged) included files
into the main binding (the binding with the 'include:' in it). This
might give a good organization, and the old scripts do it too.
Piggyback two fixes/cleanups:
- 'compatible' should be allowed to appear in included files
- No need to allow an 'inherits' key in _check_binding(), because
it has been removed before then, when merging bindings
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Except for a few special properties like 'interrupts' and '#size-cells',
require all devicetree properties on nodes with bindings to be declared
in the binding.
This will help catch misspellings, and makes refactoring and cleanups
safer.
Suggested by Peter A. Bigot.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
For missing optional properties, it can be handy to generate a default
value instead of no value, to cut down on #ifdefs.
Allow a default value to be specified in the binding, via a new
'default: <default value>' setting for properties in bindings.
Defaults are supported for both scalar and array types. YAML arrays are
used to specify the value for array types.
'default:' also appears in json-schema, with the same meaning.
Include misc. sanity checks, like the 'default' value matching 'type'.
The documentation changes in binding-template.yaml explain the syntax.
Suggested by Peter A. Bigot in
https://github.com/zephyrproject-rtos/zephyr/issues/17829.
Fixes: #17829
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Having backwards compatibility for !include and 'constraint:' is silly
without also having backwards compatibility for 'category:', because
that forces a binding change anyway.
Add backwards compatibility for 'category:', and just print a
deprecation warning when it's used.
Also move tests for deprecated features into a dedicated
test-bindings/deprecated.yaml binding, instead of piggybacking on other
tests.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Have
include: foo.dts
include: [foo.dts, bar.dts]
instead of
inherits:
!include foo.dts
inherits:
!include [foo.dts, bar.dts]
This is a nicer and shorter and less cryptic syntax, and will make it
possible to get rid of the custom PyYAML constructor for '!include'
later.
'inherits: !include ...' is still supported for backwards compatibility
for now. Later on, I'm planning to mass-replace it, add a deprecation
warning if it's used, and document 'include:'. Then the '!include'
implementation can be removed a bit later.
'!include' has caused issues in the past (see the comment above the
add_constructor() call), gets iffy with multiple EDT instances, and
makes the code harder to follow.
I'm guessing '!include' might've been intended to be useful outside of
'inherits:' originally, but that's the only place where it's used. It's
undocumented that it's possible to put it elsewhere.
To implement the backwards compatibility, the code just transforms
inherits:
!include foo.dts
into
inherits:
- foo.dts
and treats 'inherits:' similarly to 'include:'. Previously, !include
inserted the contents of the included file instead.
Some more sanity checks for 'include:'/'inherits:' are included as well.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Add two new type-checked property types 'phandles' and 'phandle-array'
to edtlib.
'phandles' is for pure lists of phandles, with no other data, like
foo = < &bar &baz ... >
'phandle-array' is for lists of phandles and (possibly) numbers, like
foo = < &bar 1 2 &baz 3 4 ... >
dt-schema also has the 'phandle-array' type.
Property.val (in edtlib) is set to an array of Device objects for the
'phandles' type.
For the 'phandle-array' type, no Property object is created. This type
is only used for type checking.
Also refactor how types that do not create a Property object
('phandle-array' and 'compound') are handled. Have _prop_val() return
None for them.
The new types are implemented with two new TYPE_PHANDLES and
TYPE_PHANDLES_AND_NUMS types at the dtlib level. There is also a new
Property.to_nodes() functions for fetching the Nodes for an array of
phandles, with type checking.
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Property type-checking has been pretty rudimentary until now, only
checking things like the length being divisible by 4 for 'type: array',
and strings being null-terminated. In particular, no checking was done
for 'type: uint8-array', letting
jedec-id = < 0xc8 0x28 0x17 >;
slip through when
jedec-id = [ 0xc8 0x28 0x17 ];
was intended.
Fix it by adding a syntax-based type checker:
1. Add Property.type, which gives a high-level type for the property,
derived from the markers added in the previous commit.
This includes types like TYPE_EMPTY ('foo;'),
TYPE_NUM ('foo = < 3 >;'), TYPE_BYTES ('foo = [ 01 02 ];'),
TYPE_STRINGS ('foo = "bar", "baz"'),
TYPE_PHANDLE ('foo = < &bar >;'), and TYPE_COMPOUND (everything not
recognized).
See the Property.type docstring in dtlib for more info.
2. Use the high-level type in
Property.to_num()/to_string()/to_node()/etc. to verify that the
property was assigned in an expected way for the type.
If the assignment looks bad, give a helpful error:
expected property 'nums' on /foo/bar in some.dts to be assigned
with 'nums = < (number) (number) ... >', not 'nums = "oops";'
Some other related changes are included as well:
- There's a new Property.to_bytes() function that works like accessing
Property.bytes, except with an added check for the value being
assigned like 'foo = [ ... ]'.
This function solves problems like the jedec-id one.
- There's a new Property.to_path() function for fetching the
referenced node for assignments like 'foo = &node;', with type
checking. (Strings are accepted too, as long as they give the path
to an existing node.)
This function is used for /chosen and /aliases.
- A new 'type: phandle' type can now be given in bindings, for
properties that are assigned like 'foo = < &node >;'.
- Property.__str__() now displays phandles and path references as they
were written (e.g. '< &foo >' instead of '< 0x1 >', if the
allocated phandle happened to be 1).
- Property.to_num() and Property.to_nums() no longer take a 'length'
parameter, because it makes no sense with the type checking.
- The global dtlib.to_string() and dtlib.to_strings() functions were
removed, because they're not that useful.
- More tests were added, along with misc. minor cleanup in various
places.
- Probably other stuff I forgot.
The more strict type checking in dtlib indirectly makes some parts of
edtlib more strict as well (wherever Property.to_*() is used).
Fixes: #18131
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
Add a new DTS/binding parser to scripts/dts/ for generating
generated_dts_board.conf and generated_dts_board_unfixed.h.
The old code is kept to generate some deprecated defines, using the
--deprecated-only flag. It will be removed later.
The new parser is implemented in three files in scripts/dts/:
dtlib.py:
A low-level .dts parsing library. This is similar to devicetree.py in
the old code, but is a general robust DTS parser that doesn't rely on
preprocessing.
edtlib.py (e for extended):
A library built on top of dtlib.py that brings together data from DTS
files and bindings and creates Device instances with all the data for
a device.
gen_defines.py:
A script that uses edtlib.py to generate generated_dts_board.conf and
generated_dts_board_unfixed.h. Corresponds to extract_dts_includes.py
and the files in extract/ in the old code.
testdtlib.py:
Test suite for dtlib.py. Can be run directly as a script.
testedtlib.py (uses test.dts and test-bindings/):
Test suite for edtlib.py. Can be run directly as a script.
The test suites will be run automatically in CI.
The new code turns some things that were warnings (or not checked) in
the old code into errors, like missing properties that are specified
with 'category: required' in the binding for the node.
The code includes lots of documentation and tries to give helpful error
messages instead of Python errors.
Co-authored-by: Kumar Gala <kumar.gala@linaro.org>
Signed-off-by: Ulf Magnusson <Ulf.Magnusson@nordicsemi.no>
