On Mon, Aug 18, 2014 at 11:54:29AM +0100, Mark Rutland wrote:
> Hi Mika,
>
> While I am very much in favour of having a structured way of describing
> device specific data in ACPI I am very concerned by the idea of assuming
> (a false) equivalence with DT. More on that below.
Hi Mark,
The equivalence is intended in the sense that we should be able to represent any
of the existing schemas with the same properties in ACPI as in DT.
...
> > +An example device where we might need properties is a GPIO keys device.
> > +In addition to the GpioIo/GpioInt resources the driver needs to know how
> > +to map each GpioIo resource to the corresponding Linux input event.
> > +
> > +To solve this we add the following ACPI device properties from the
> > +gpio-keys schema:
> > +
> > + Device (KEYS) {
> > + Name (_DSD, Package () {
> > + ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
> > + Package () {
> > + Package () {"poll-interval", 100},
> > + Package () {"autorepeat", 1}
> > + }
> > + })
> > +
> > + Device (BTN0) {
> > + Name (_DSD, Package () {
> > + ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
> > + Package () {
> > + Package () {"linux,code", 105},
> > + Package () {"linux,input-type", 1},
>
> The leakage of Linux implementation details into DTBs is bad enough. I
> would really not like to see this kind of thing in ACPI tables.
>
> Leaking this into HW description of any sort completely rids said
> description of any OS independence, and it's usually short-sighted,
> nonsensical, or ill-defined (what does 'autorepeat' actually mean?).
>
> I am very worried by the prospect of _DSD making it harder rather than
> easier for an arbitrary OS to use ACPI data due to assumptions being
> made about (a particular revision of) a particular OS.
This is perhaps not the best example to include in the documentation. A guiding
principle has been to only represent in the DSD that which is an actual property
of the hardware.
It is also a goal of the mechanism to enable ACPI systems to reuse the 200+
existing OF-aware drivers in the kernel today. If those schemas are
ill-conceived, we should work to improve the schemas.
On the point of OS-independence, while I agree that using something like the USB
keycodes here would be better, the namespacing employed in these properties
makes it possible to use the same DSD for muliple OSes as another OS could add
it's own keycode, for example, without conflict. Again, that is really an issue
with the schemas, not the _DSD mechanism itself.
Perhaps we should use a different example in the documentation as I would agree
this is not a model schema.
> > +Of course the device driver then needs to iterate over these devices but
> > +it can be done easily via the ->children field of the companion ACPI
> > +device. This will be demonstrated later on in the document.
> > +
> > +If there is an existing Device Tree binding for a device, it is expected
> > +that the same bindings are used with ACPI properties, so that the driver
> > +dealing with the device needs only minor modifications if any.
>
> For simple devices like a UART which just has a "clock-frequency"
> property in DT, importing both the concept of the property and the key
> name itself is a sensible approach.
>
> However, I do not think this makes sense as a general design rule. ACPI
> and DT already have different idioms for describing certain things (e.g.
> GPIOs, interrupts) and care needs to be taken to distinguish simple
> device-specific properties from class-specific properties or implied
> linkages between devices.
I believe we have GPIO handled well in this series. In DT terms, ACPI is the
controller, providing a list of GpioIO|GpioInt resources in the CRS of the
referenced object. Arguments are Resource Index, Pin Index, Polarity. The rest
of the properties are handled via the GpioIO|GpioInt resources in the _CRS.
>
> The "#clock-cells" and "#pwm-cells" _DSD examples in ACPI 5.1 are
> complete nonsense. They import an idiom from DT into ACPI without
Yeah... coming up with good examples for a generic mechanism which are both
descriptive and universally unobjectionable. Good point. We might be able to
improve this via an errata update. What would you recommend?
> considering the surrounding context (cells, lack of typing, phandles,
How are the phandles not covered via the ACPI namespace references?
The point on cells is a good one, and I believe some of this can be mitigated by
providing subsystem specific accessors as we have for GPIO, for example.
> etc). As an example, I worry that why it may be trivial to map pinctrl
> bindings into ACPI _DSD, there is no clear way that they would interact
> with ACPI's HW management. This is going to significantly expand our
> problem surface area.
Pinctrl is an area where I suspect we will need more specific support, but
it is still on my list for a detailed review.
> There are known deficiencies with many DT bindings that we can't fix for
> legacy reasons on the DT side (e.g. the clock detection for ARM
> primecell devices). Some of these issues can be trivially fixed for a
> legacy-free environment like ACPI _DSD.
>
> We have a lot of stupid (and broken) DT bindings that were either
> ill-designed or which newer hardware has exposed weaknesses in. I don't
> want us to turn these into even more broken ACPI _DSD bindings.
That's surely the case. Could you provide a concrete example or two I can use to
build context?
> For the reasons above I am against having a single API which accesses
> both interfaces. I believe at present it is better to separate the two.
> For those devices which truly need a small amount of data there will not
> be much duplication. For those which need large amounts of data we
> clearly need to consider whether those bindings are sensible to pull
> across into ACPI at all.
The issue here is there are over 200 drivers in the kernel today that are not
usable by ACPI systems, and many more external to the kernel which are only
usable via board files. The alternative would be to modify every one of these
drivers with an ACPI-specific pdata creation function and to create a new schema
library for ACPI which would differ from DT. Ultimiately it would be preferable
for the hardware vendor to define the properties required to describe the
hardware in a way independent from the firmware implementation.
> > +2. Formal definition of properties
> > +----------------------------------
> > +The following chapters define the currently supported properties. For
> > +these there exists a helper function that can be used to extract the
> > +property value.
> > +
> > +2.1 Integer types
> > +-----------------
> > +ACPI integers are always 64-bit. However, for drivers the full range is
> > +typically not needed so we provide a set of functions which convert the
> > +64-bit integer to a smaller Linux integer type.
>
> Does that always make sense, or should that be left to the driver?
>
I don't follow - the driver chooses which size integer to request. Josh has
suggested these functions should check for overflow, which I thought made sense.
> In some cases properties form DT having a max value of 0xffffffff is
> simply a result of using the default DT number representation rather
> than a conscious decision that 32 bits are enough.
>
> > +An integer property looks like this:
> > +
> > + Package () {"poll-interval", 100},
> > + Package () {"i2c-sda-hold-time-ns", 300},
> > + Package () {"clock-frequency", 400000},
> > +
> > +To read a property value, use a unified property accessor as shown
> > +below:
> > +
> > + u32 val;
> > + int ret;
> > +
> > + ret = device_property_read(dev, "poll-interval", DEV_PROP_U32, &val);
> > + if (ret)
> > + /* Handle error */
> > +
> > +The function returns 0 if the property is copied to 'val' or negative
> > +errno if something went wrong (or the property does not exist).
> > +
> > +2.2 Integer arrays
> > +------------------
> > +An integer array is a package holding only integers. Arrays can be used to
> > +represent different things like Linux input key codes to GPIO mappings, pin
> > +control settings, dma request lines, etc.
>
> As I mentioned above, I am extremely concerned that it is not sensible
> to blindly import the class bindings you mention here.
>
> > +An integer array looks like this:
> > +
> > + Package () {
> > + "max8952,dvs-mode-microvolt",
> > + Package () {
> > + 1250000,
> > + 1200000,
> > + 1050000,
> > + 950000,
> > + }
> > + }
> > +
> > +The above array property can be accessed like:
> > +
> > + u32 voltages[4];
> > + int ret;
> > +
> > + ret = device_property_read_array(dev, "max8952,dvs-mode-microvolt",
> > + DEV_PROP_U32, voltages,
> > + ARRAY_SIZE(voltages));
> > + if (ret)
> > + /* Handle error */
> > +
> > +
> > +All functions copy the resulting values cast to a requested type to the
> > +caller supplied array. If you pass NULL in the value pointer ('voltages' in
> > +this case), the function returns number of items in the array. This can be
> > +useful if caller does not know size of the array beforehand.
>
> On the DT side we have of_property_count_u{8,16,32,64}_elems. Having
> separate functions makes the code easier to read.
That makes sense to me, especially as we would be replacing such usage in
existing drivers, it would be nice for it to be as close to a direct mapping as
possible.
...
> > +An object reference looks like this:
> > +
> > + Package () {"clock", \_SB.CLK0},
>
> For examples I would strongly recommend using something slightly more
> abstract so that (for example) it is not assumed that this is the
> canonical way to describe a clock using _DSD.
>
Agreed.
> > +At the time of writing this, there is no unified device_property_* accessor
> > +for references so one needs to use the following ACPI helper function:
> > +
> > + int acpi_dev_get_property_reference(struct acpi_device *adev,
> > + const char *name,
> > + const char *size_prop, int index,
> > + struct acpi_reference_args *args);
> > +
> > +The referenced ACPI device is returned in args->adev if found.
> > +
> > +In addition to simple object references it is also possible to have object
> > +references with arguments. These are represented in ASL as follows:
> > +
> > + Device (\_SB.PCI0.PWM) {
> > + Name (_DSD, Package () {
> > + ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
> > + Package () {
> > + Package () {"#pwm-cells", 2}
> > + }
> > + })
> > + }
> > +
> > + Device (\_SB.PCI0.BL) {
> > + Name (_DSD, Package () {
> > + ToUUID("daffd814-6eba-4d8c-8a91-bc9bbf4aa301"),
> > + Package () {
> > + Package () {
> > + "pwms",
> > + Package () {
> > + \_SB.PCI0.PWM, 0, 5000000,
> > + \_SB.PCI0.PWM, 1, 4500000,
> > + }
>
> I'm not all that familiar with ASL so perhaps this is not possible, but
> IMO it would make far more sense to have the two elements described as
> individual packages:
>
> Package {
> "pwms",
> Package () { \ref1, 0, 5235 }
> Package () { \ref2, 1, 42423 }
> }
>
> That way there's no need for #x-cells properties, it's possible to do
> better type checking, variadic arguments can be supported, etc.
While the above is valid ASL (minus a comma), the prp_uuid doesn't allow for
multiple "values" in the key value pairs. If wanted to encapsulate each pwm
entry, we would need to give each a unique name, "pwm0", "pwm1" etc. The above
is intended to parallel the DT to make it easiest to reuse drivers.
>
> As it stands this live in the intersection of the limitations of both
> ACPI and DT, and I suspect this will only cause pain on both sides.
>
> > + }
> > + }
> > + })
> > + }
> > +
> > +In the above example, the referenced device declares a property that
> > +returns the number of expected arguments (here it is "#pwm-cells"). If
> > +no such property is given we assume that all the integers following the
> > +reference are arguments.
> > +
> > +In the above example PWM device expects 2 additional arguments. This
> > +will be validated by the ACPI property core.
> > +
> > +The additional arguments must be integers. Nothing else is supported.
>
> Is that an ASL limitation or a Linux limitation?
This is a Linux implementation (or our choice of) limitation. Are there other
argument types we should be considering?
>
> > +It is possible, as in the above example, to have multiple references
> > +with varying number of integer arguments. It is up to the referenced
> > +device to declare how many arguments it expects. The 'index' parameter
> > +selects which reference is returned.
> > +
> > +One can use acpi_dev_get_property_reference() as well to extract the
> > +information in additional parameters:
> > +
> > + struct acpi_reference_args args;
> > + struct acpi_device *adev = /* this will point to the BL device */
> > + int ret;
> > +
> > + /* extract the first reference */
> > + acpi_dev_get_property_reference(adev, "pwms", "#pwm-cells", 0, &args);
> > +
> > + BUG_ON(args.nargs != 2);
> > + BUG_ON(args.args[0] != 0);
> > + BUG_ON(args.args[1] != 5000000);
> > +
> > + /* extract the second reference */
> > + acpi_dev_get_property_reference(adev, "pwms", "#pwm-cells", 1, &args);
> > +
> > + BUG_ON(args.nargs != 2);
> > + BUG_ON(args.args[0] != 1);
> > + BUG_ON(args.args[1] != 4500000);
> > +
> > +In addition to arguments, args.adev now points to the ACPI device that
> > +corresponds to \_SB.PCI0.PWM.
> > +
> > +It is intended that this function is not used directly but instead
> > +subsystems like pwm implement their ACPI support on top of this function
> > +in such way that it is hidden from the client drivers, such as via
> > +pwm_get().
>
> If we're expecting device class bindings to use high-level interfaces
> (like pwm_get, clk_get, etc) then as far as I can see there is no reason
> that the low-level representation in ACPI and DT need to be identical.
There are properties that will not be subsystem specific, or rather that are
simple key:value pairs that are convenient to access with the common mechanism.
Even if we get a number of these abstractions in place, there is still value in
a simple common property API in that it enables drivers to be ignorant of the
underlying firmware-mechanism.
We have an example in this series for gpiolib - is this a good example in your
opinion? Would adding examples/mechanism for pwm_get etc in the series help?
> Importing DT bindings wholesale into ACPI only optimises for getting
> things supported quickly, not getting things supported sanely.
>
> Thanks,
> Mark.
We're also trying to lower the barriers to enabling and writing new drivers that
work on both DT and ACPI based systems without domain specific knowledge of
each.
Thanks for taking the time to review and comment Mark, very much appreciated.
--
Darren Hart
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
