Hi Andy - thanks for comments!

On 22/02/2021 14:58, Andy Shevchenko wrote:

On Mon, Feb 22, 2021 at 3:12 PM Daniel Scally [off-list ref] wrote:

quoted

ACPI devices with _HID INT3472 are currently matched to the tps68470
driver, however this does not cover all situations in which that _HID
occurs. We've encountered three possibilities:

1. On Chrome OS devices, an ACPI device with _HID INT3472 (representing
a physical TPS68470 device) that requires a GPIO and OpRegion driver
2. On devices designed for Windows, an ACPI device with _HID INT3472
(again representing a physical TPS68470 device) which requires GPIO,
Clock and Regulator drivers.
3. On other devices designed for Windows, an ACPI device with _HID
INT3472 which does **not** represent a physical TPS68470, and is instead
used as a dummy device to group some system GPIO lines which are meant
to be consumed by the sensor that is dependent on this entry.

This commit adds a new module, registering a platform driver to deal
with the 3rd scenario plus an i2c driver to deal with #1 and #2, by
querying the CLDB buffer found against INT3472 entries to determine
which is most appropriate.

Can you split CLK parts (and maybe regulators as well) to something
like intel_skl_int3472_clk.c?

Sure, no problem

quoted

+#include <linux/acpi.h>
+#include <linux/i2c.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+               dev_err(&adev->dev, "%s object is not an ACPI buffer\n", id);

Perhaps acpi_handle_err() et al. instead of dev_*(&adev->dev, ...)
where it's applicable?

Ah - yes, ok, thanks. TIL those exist

quoted

+       if (obj->buffer.length > sizeof(*cldb)) {
+               dev_err(&adev->dev, "The CLDB buffer is too large\n");
+               ret = -EINVAL;

ENOSPC? ENOMEM?

I still think EINVAL actually, as in this case the problem isn't that
space couldn't be allocated but that the buffer in the SSDB is larger
than I expect it to be, which means the definition of it has changed /
this device isn't actually supported.

quoted

+       ret = platform_driver_register(&int3472_discrete);
+       if (ret)
+               return ret;
+
+       ret = i2c_register_driver(THIS_MODULE, &int3472_tps68470);
+       if (ret)
+               platform_driver_unregister(&int3472_discrete);

Not a fan of the above, but let's see what others will say...

Yeah; happy to discuss this more if needed.

quoted

+#include <linux/clk-provider.h>

This is definitely not for *.h. (Not all C files needed this)

quoted

+#include <linux/gpio/machine.h>

Ditto.

quoted

+#include <linux/regulator/driver.h>
+#include <linux/regulator/machine.h>

Ditto.

Yep; I'll move them to *_clk.c and *_regulator.c files.

quoted

+static int skl_int3472_clk_prepare(struct clk_hw *hw)
+{
+       struct int3472_gpio_clock *clk = to_int3472_clk(hw);
+
+       gpiod_set_value(clk->ena_gpio, 1);
+       if (clk->led_gpio)

Make it optional and drop this check. Same for other places of use of this GPIO.

Oops, of course, thanks

quoted

+static int skl_int3472_clk_enable(struct clk_hw *hw)
+{
+       /*
+        * We're just turning a GPIO on to enable, which operation has the
+        * potential to sleep. Given enable cannot sleep, but prepare can,
+        * we toggle the GPIO in prepare instead. Thus, nothing to do here.
+        */

Missed . and / or  () in some words? (Describing callbacks, personally
I use the form "->callback()" in such cases)

OK, I'll fix the comment to match that style.

quoted

+static unsigned int skl_int3472_get_clk_frequency(struct int3472_discrete_device *int3472)
+{
+       union acpi_object *obj;

+       unsigned int ret = 0;

unsigned for ret is unusual. Looking into the code, first of all it
doesn't need this assignment; second, it probably can gain a better
name: "frequency"?

Yep ok, I'll rename to freq/frequency

quoted

+       if (!IS_ERR_OR_NULL(sensor_config) && sensor_config->function_maps) {

Hmm...

Would

if (IS_ERR_OR_NULL(sensor_config))
  return 0;

if (!_maps)
  return 0;

with respective comments working here?

No, because the absence of either sensor_config or
sensor_config->function_maps is not a failure mode. We only need to
provide sensor_configs for some platforms, and function_maps for even
fewer. So if that check is false, the rest of the function should still
execute.

quoted

+static int skl_int3472_register_clock(struct int3472_discrete_device *int3472)
+{
+       struct clk_init_data init = {
+               .ops = &skl_int3472_clock_ops,
+               .flags = CLK_GET_RATE_NOCACHE,
+       };
+       int ret = 0;
+
+       init.name = kasprintf(GFP_KERNEL, "%s-clk",
+                             acpi_dev_name(int3472->adev));

devm_*() ? Or is the lifetime different?

No it's not; I'll use devm_*(), thanks

quoted

+       sensor_config = int3472->sensor_config;
+       if (IS_ERR_OR_NULL(sensor_config)) {
+               dev_err(int3472->dev, "No sensor module config\n");
+               return PTR_ERR(sensor_config);

NULL -> 0. Is it okay?

Ah, no it's not - good catch thank you.

quoted

+       if (ares->type != ACPI_RESOURCE_TYPE_GPIO ||
+           ares->data.gpio.connection_type != ACPI_RESOURCE_GPIO_TYPE_IO)
+               return 1; /* Deliberately positive so parsing continues */

I don't like to lose control over ACPI_RESOURCE_TYPE_GPIO, i.e.
spreading it over kernel code (yes, I know about one existing TS
case).
Consider to provide a helper in analogue to acpi_gpio_get_irq_resource().

Sure, but I probably name it acpi_gpio_is_io_resource() - a function
named "get" which returns a bool seems a bit funny to me.

quoted

+       if (ret < 0 && ret != -EPROBE_DEFER)
+               dev_err(int3472->dev, err_msg);

dev_err_probe() will make the above conditional go away. And you may even do...

Ah-ha - thought that must exist but couldn't find it - thank you.

quoted

+       if (int3472->clock.ena_gpio) {

Not sure you need this here.

We haven't seen a device that lacks a clock enable GPIO it's true, but
since all the other kinds seem optional it didn't seem impossible that
that one is optional too. I can remove if you prefer and we can just
deal with it when we encounter one like that though?

quoted

+       /* Max num GPIOs we've seen plus a terminator */
+       int3472 = kzalloc(struct_size(int3472, gpios.table,
+                         INT3472_MAX_SENSOR_GPIOS + 1), GFP_KERNEL);

Wonder of you can use devm_*() APIs in this function.

Yeah I can, I'll switch to that.

quoted

+int skl_int3472_discrete_remove(struct platform_device *pdev)
+{
+       struct int3472_discrete_device *int3472 = platform_get_drvdata(pdev);
+       if (int3472->gpios.dev_id)
+               gpiod_remove_lookup_table(&int3472->gpios);

gpiod_remove_lookup_table() is now NULL-aware.
But in any case I guess you don't need the above check.

Sorry; forgot to call out that I didn't follow that suggestion;
int3472->gpios is a _struct_ rather than a pointer, so &int3472->gpios
won't be NULL, even if I haven't filled anything in to there yet because
it failed before it got to that point. So, not sure that it quite works
there.

quoted

+       if (!IS_ERR(int3472->regulator.rdev))
+               regulator_unregister(int3472->regulator.rdev);

Shouldn't it be the pointer to the regulator itself?

int3472->regulator is type struct int3472_gpio_regulator, the .rdev is
the normal regulator_dev

quoted

+       if (!IS_ERR(int3472->clock.clk))

If you get it optional, you won't need this additional check.

Yes - here it will definitely work; thanks, I'll add that patch

quoted

+       ret = skl_int3472_fill_cldb(adev, &cldb);
+       if (!ret && cldb.control_logic_type != 2) {
+               dev_err(&client->dev, "Unsupported control logic type %u\n",
+                       cldb.control_logic_type);
+               return -EINVAL;
+       }
+
+       if (ret)
+               cldb_present = false;

if (ret)
  ...
else if (...)  {
  ...
  return ...;
}

Oh yeah...now you point that out I have no idea what I was thinking there...