Re: [PATCH v2 04/22] of/platform: add of_platform_device_find()

[Tomeu Vizoso]

On 28 July 2015 at 17:31, Rob Herring <robherring2@xxxxxxxxx> wrote:
> On Tue, Jul 28, 2015 at 8:54 AM, Tomeu Vizoso
> <tomeu.vizoso@xxxxxxxxxxxxx> wrote:
>> On 28 July 2015 at 15:39, Rob Herring <robherring2@xxxxxxxxx> wrote:
>>> On Tue, Jul 28, 2015 at 8:19 AM, Tomeu Vizoso
>>> <tomeu.vizoso@xxxxxxxxxxxxx> wrote:
>>>> From an arbitrary node in the tree, find the enclosing node that
>>>> corresponds to a platform device, as registered by
>>>> of_platform_populate().
>>>>
>>>> This can be used to find out what device needs to be probed so the
>>>> dependency described by a given node is made available.
>>>>
>>>> Signed-off-by: Tomeu Vizoso <tomeu.vizoso@xxxxxxxxxxxxx>
>>>> ---
>>>>
>>>> Changes in v2:
>>>> - Move the logic for finding a platform device from its firmware node to
>>>>   of/platform.c as it's not needed for ACPI nodes.
>>>>
>>>>  drivers/of/platform.c       | 60 +++++++++++++++++++++++++++++++++++++++++++++
>>>>  include/linux/of_platform.h |  1 +
>>>>  2 files changed, 61 insertions(+)
>>>>
>>>> diff --git a/drivers/of/platform.c b/drivers/of/platform.c
>>>> index ff27494cda8c..89c5cd513027 100644
>>>> --- a/drivers/of/platform.c
>>>> +++ b/drivers/of/platform.c
>>>> @@ -501,6 +501,66 @@ void of_platform_depopulate(struct device *parent)
>>>>  }
>>>>  EXPORT_SYMBOL_GPL(of_platform_depopulate);
>>>>
>>>> +static bool of_is_platform(struct device_node *np)
>>>> +{
>>>> +       int count;
>>>> +
>>>> +       count = of_property_count_strings(np, "compatible");
>>>> +
>>>> +       /* The node has to have a compatible string */
>>>> +       if (!count)
>>>> +               return false;
>>>> +
>>>> +       /* But it cannot be just a simple memory-mapped bus */
>>>> +       if (count == 1 && of_match_node(of_default_bus_match_table, np))
>>>> +               return false;
>>>> +
>>>> +       /* But AMBA devices aren't platform devices */
>>>> +       if (of_device_is_compatible(np, "arm,primecell"))
>>>> +               return false;
>>>> +
>>>> +       /* Node is immediately below root */
>>>> +       if (!np->parent || !np->parent->parent)
>>>> +               return true;
>>>> +
>>>> +       /* If it's a node in a simple memory-mapped bus */
>>>> +       if (of_match_node(of_default_bus_match_table, np->parent))
>>>> +               return true;
>>>
>>> This seems really fragile.
>>
>> I think this finding logic matches the logic for registering platform
>> devices in of_platform_populate and also what is documented in
>> Documentation/devicetree/usage-model.txt.
>
> Right. So now we have that logic in 2 places. One is descending from
> the root and one is walking up from the child. That's an opportunity
> for some mismatch.

Definitely.

>>> What about platform devices which are
>>> children of MFDs but are not "simple-mfd"?
>>
>> This code should deal fine with those (the boards I tested with do
>> have them). It probes the mfd master, and that in turn will call
>> mfd_add_devices causing the target device to be probed.
>
> I don't see how this function would ever return true in this case
> unless you put the MFD at the root level. The only other way to return
> true is matching on of_default_bus_match_table for the parent (i.e.
> the MFD).

Oops, you are right.

>>> Does of_find_device_by_node not work for you?
>>
>> Well, the dependencies aren't always platform devices, that's why I
>> need to go up the tree until I find a node that corresponds to a
>> platform device that I can query and probe.
>>
>> If I had a way to get, say, a i2c device from its fwnode then I would
>> just need to make sure that a device's parent is probed before probing
>> it and everything would be cleaner in the OF case.
>
> If you have the struct device from the device_node, then you should be
> able to do this, right?

Yes, if I could go back from the device_node to the struct device that
was registered from it, for all buses, then all this would be much
simpler and more robust. It would basically work like in the ACPI
case.

I will play with this idea.

>>> That is probably not the
>>> most efficient search, but we could fix that. We could add struct
>>> device ptr to struct device_node and check without searching for
>>> example.
>>
>> That would be great, but I thought there was an issue with a OF node
>> being able to be related to more than one struct device (but I haven't
>> found this myself yet).
>
> I think it pretty much should be one to one. I'm not aware of any
> examples where that is not the case. This function would already be
> broken if you could have more than one struct device.

Well, for platform devices we currently know that there can only be
one struct device for a given device_node, but that's not so clear for
other devices.

> There is an issue where you could have 2 drivers match the same node,
> but on different compatible strings. But that's a different issue.

Yup.

Thanks,

Tomeu

> Rob
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/