[U-Boot] [PATCH] dm: allow setting driver_data before/during bind

[Stephen Warren]

On 05/10/2016 08:25 PM, Simon Glass wrote:
> Hi Stephen,
>
> On 4 May 2016 at 12:42, Stephen Warren <swarren at wwwdotorg.org> wrote:
>> On 05/01/2016 01:27 PM, Simon Glass wrote:
>>>
>>> Hi Stephen,
>>>
>>> On 28 April 2016 at 17:08, Stephen Warren <swarren at wwwdotorg.org> wrote:
>>>>
>>>> From: Stephen Warren <swarren at nvidia.com>
>>>>
>>>> This will allow a driver's bind function to use the driver data. One
>>>> example is the Tegra186 GPIO driver, which instantiates child devices
>>>> for each of its GPIO ports, yet supports two different HW instances each
>>>> with a different set of ports, and identified by the udevice_id .data
>>>> field.
>>>>
>>>> Signed-off-by: Stephen Warren <swarren at nvidia.com>
>>>> ---
>>>>    drivers/core/device.c            | 7 ++++---
>>>>    drivers/core/lists.c             | 6 +++---
>>>>    drivers/gpio/dwapb_gpio.c        | 2 +-
>>>>    drivers/gpio/s5p_gpio.c          | 2 +-
>>>>    drivers/gpio/sunxi_gpio.c        | 2 +-
>>>>    drivers/gpio/tegra_gpio.c        | 2 +-
>>>>    drivers/mtd/spi/sandbox.c        | 2 +-
>>>>    drivers/net/mvpp2.c              | 3 ++-
>>>>    drivers/pci/pci-uclass.c         | 5 ++---
>>>>    drivers/power/pmic/pmic-uclass.c | 2 +-
>>>>    drivers/usb/host/usb-uclass.c    | 5 ++---
>>>>    include/dm/device-internal.h     | 5 +++--
>>>>    12 files changed, 22 insertions(+), 21 deletions(-)
>>>
>>>
>>> I'm not sure this extra parameter carries its weight:
>>>
>>> - most callers just pass 0
>>
>> The same is true of the existing platdata field in many cases.
>
> Yes, but platdata is defined to be needed by bind(), whereas
> driver_data is supposed to be used in probe()  to find out which
> device tree compatible string matched.

This seems to conflict with the documentation in include/dm/device.h; it 
claims that platdata is created by calling ofdata_to_platdata() just 
before calling probe(), at least for the DT case (for the case where 
U_BOOT_DEVICE is used, the platdata is available right from the start).

I couldn't find anywhere in the documentation that states when 
driver_data is supposed to be used; could you point me at it so I can 
read it?

 > Remember that the device tree
> properties are not looked at during bind(), only later. So it makes
> sense to include platdata in the device_bind() call, but not
> driver_data.

Hmm, drivers/gpio/tegra_gpio.c:gpio_tegra_bind() uses DT (which you 
wrote or at least converted it to DM and chose where to put the DT 
accesses), and you very recently reviewed and applied "video: tegra: 
refuse to bind to disabled dcs" which modified 
drivers/video/tegra.c:tegra_lcd_bind() to use DT.

Admittedly I do now see the following in doc/driver-model/README.txt:

> The device's bind() method is permitted to perform simple actions, but
> should not scan the device tree node, not initialise hardware, nor set up
> structures or allocate memory. All of these tasks should be left for
> the probe() method.

... but then I wonder why that rule was enforced for the patch in this 
thread, but not in the other cases?

This inconsistency in review is extremely frustrating to me.

>>> - the field is supposed to be set up by device tree and probing tables,
>>> not code
>>
>> While the existence of this new parameter does allow arbitrary code to set
>> the parameter, this patch only actually sets the parameter in the case where
>> DT and probing tables have determined that value.
>
> I don't think so. That value is set in lists_bind_fdt().

Sure, but that function is only used from 3 places, and explicitly 
accepts a parameter to indicate which DT node to instantiate a device 
for. It won't work unless a valid DT node is passed to it, and therefore 
can only work for DT-based probing.

> I wonder if you could set it yourself after calling device_bind()?

The Tegra186 GPIO driver explicitly needs to use the data inside the 
driver's bind() function in order to know how many child devices to 
instantiate. Setting the value after calling device_bind() (which the 
core DM code already does) is too late.

For reference, you can see the exact code at:
http://lists.denx.de/pipermail/u-boot/2016-April/252238.html
"gpio: add Tegra186 GPIO driver"

Search for "tegra186_gpio_bind" and look at the assignment to, and use 
of, ctlr_data.

I had also quoted that part of the code in my previous email.

>>> - bind() methods should not care about the driver data (they are not
>>> allowed to touch hardware), so setting it later is fine
>>
>> Not touching HW is fine, but the driver data can still feed into purely SW
>> decisions that bind makes. More details below.
>>
>>> - you can already pass platform data to the driver which is the
>>> preferred communication method from a parent to its children
>>
>> I don't believe this is possible for devices instantiated from DT is it? In
>> that case, platform data is always NULL:
>
> That's right. For DT the paltform data is set up in the
> ofdata_to_platdata() method. Since you are using DT, you should follow
> that convention.

This is the opposite of what you said above, which was that platdata is 
for bind().

>> int lists_bind_fdt(struct udevice *parent, const void *blob, int offset,
>>                     struct udevice **devp)
>> ...
>>                  ret = device_bind(parent, entry, name, NULL, id->data,
>>                                    offset, &dev);
>>
>> (That quoted code is with this patch applied, and the NULL value is the
>> platform data parameter.)
>>
>>> Also it's not clear from your Tegra 186 GPIO patch where you are using
>>> this.
>>
>> Here's the relevant part from the Tegra186 GPIO driver patch I posted:
>>
>>> +static int tegra186_gpio_bind(struct udevice *parent)
>>> +{
>>> +       struct tegra186_gpio_platdata *parent_plat = parent->platdata;
>>> +       struct tegra186_gpio_ctlr_data *ctlr_data =
>>> +               (struct tegra186_gpio_ctlr_data *)parent->driver_data;
>>
>> ...
>>>
>>> +       /* If this is a child device, there is nothing to do here */
>>> +       if (parent_plat)
>>> +               return 0;
>>
>> ...
>>>
>>> +       for (port = 0; port < ctlr_data->port_count; port++) {
>>
>> ...
>>>
>>> +               plat->name = ctlr_data->ports[port].name;
>>> +               plat->regs = &(regs[ctlr_data->ports[port].offset / 4]);
>>
>>
>> The data is used to determine how many child devices (one per port) to
>> create, and the name and register offset of each one. This is modelled after
>> the logic in the previous Tegra GPIO driver that you wrote, with the
>> unfortunate modification that the register layout is more "interesting" on
>> Tegra186, and so we can't determine the number of and parameters for the
>> child devices purely algorithmically, since the register layout is decidedly
>> non-linear.
>
> OK I see. This feels like something that your device tree should
> describe.

DT generally describes the presence of HW blocks, not the internal 
details of those HW blocks, since that is a fixed facet of the hardware 
design and can be statically derived from the value of the compatible 
property.

Equally, I'm not sure how describing the details of the HW differences 
in DT would help. The details need to be known by the driver's bind() 
function, which you and the U-Boot documentation both state isn't 
allowed to access DT.

> Failing that, how about a hard-coded table of information in
> the source code? You can look through the table and create the
> appropriate child devices.

That is EXACTLY what the code is doing. The only issue is that the 
driver supports two different compatible properties and needs to know 
which one was found in DT in order to use the right table. That's what 
driver_data is; the .data value from the udevice_id/of_match table.

>> I suppose an alternative would be to create separate U_BOOT_DRIVER()s for
>> each compatible value with different register layout, and then have the
>> bind() for each of those call into some common implementation with a
>> hard-coded parameter. Still, it seems like the usage in the current code is
>> exactly what udevice_id.data is for; to avoid having to implement separate
>> functions that do that.
>
> Yes, but you should use different compatible strings for the nodes.

That's EXACTLY what the code does. However, there is currently no way 
for bind() to find out which entry in the udevice_id table matched, 
since the DM core currently sets driver_data after calling bind() rather 
than before.

> As
> I understand it, you only have a single node, so re-purposing this
> does not seem right to me.

I must not understand what you're conveying by "single node" here. To 
quote the example from the Tegra186 GPIO DT binding documentation, here 
is what the DT looks like:

	gpio at 2200000 {
		compatible = "nvidia,tegra186-gpio";
		reg-names = "security", "gpio";
		reg =
			<0x0 0x2200000 0x0 0x10000>,
			<0x0 0x2210000 0x0 0x10000>;
		interrupts =
			<0 47 IRQ_TYPE_LEVEL_HIGH>,
			<0 50 IRQ_TYPE_LEVEL_HIGH>,
			<0 53 IRQ_TYPE_LEVEL_HIGH>,
			<0 56 IRQ_TYPE_LEVEL_HIGH>,
			<0 59 IRQ_TYPE_LEVEL_HIGH>,
			<0 180 IRQ_TYPE_LEVEL_HIGH>;
		gpio-controller;
		#gpio-cells = <2>;
		interrupt-controller;
		#interrupt-cells = <2>;
	};

	gpio at c2f0000 {
		compatible = "nvidia,tegra186-gpio-aon";
		reg-names = "security", "gpio";
		reg =
			<0x0 0xc2f0000 0x0 0x1000>,
			<0x0 0xc2f1000 0x0 0x1000>;
		interrupts =
			<0 60 IRQ_TYPE_LEVEL_HIGH>;
		gpio-controller;
		#gpio-cells = <2>;
		interrupt-controller;
		#interrupt-cells = <2>;
	};

Those two HW blocks are roughly but not exactly the same HW. Hence, 
there are two compatible values. The two HW blocks are similar enough to 
be handled by a single driver though, and hence the udevice_id table has 
two entries:

static const struct udevice_id tegra186_gpio_ids[] = {
	{
		.compatible = "nvidia,tegra186-gpio",
		.data = (ulong)&tegra186_gpio_main_data,
	},
	{
		.compatible = "nvidia,tegra186-gpio-aon",
		.data = (ulong)&tegra186_gpio_aon_data,
	},
	{ }
};

The .data value there provides all necessary information for the driver 
to handle both HW block designs.

All I need is to be able to access the ".data" from that table in 
bind(), whereas the DM core currently doesn't allow that.

>> Perhaps the creation of the child devices could happen in probe() rather
>> than bind()? I imagine there's some reason this wouldn't work (such as this
>> causing the devices to be created too late to be referenced by other
>> drivers?) or you would have done this in the existing Tegra GPIO driver.
>
> Best not - it is good to have the devices known on start-up. Let me
> know if the above solution doesn't work.

I must admit, I didn't see any solution offered in your email.

If the child devices must be created in bind(), which I do agree makes 
sense, then bind() must be able to determine which udevice_id table 
entry is related to the device being bound. Is there another way to do 
that besides using the driver_data from the udevice_id table? I suppose 
I could make bind() go read the compatible property from the DT node for 
the device, and manually search through the udevice_id table and find 
the matching entry. However, that would be duplicating work the DM core 
has already done, and exposes by setting the device's driver_data right 
after calling the driver's bind(), so doesn't make sense to me.