Re: [PATCH v1 1/2] platform/x86: serial-multi-instantiate: Set fwnode for i2c

[Hans de Goede]

Hi,

On 11/24/22 12:47, Hans de Goede wrote:
> Hi Stefan,
> 
> On 11/24/22 12:07, Stefan Binding wrote:
>> This allows the i2c driver to obtain the ACPI_COMPANION.
>>
>> Signed-off-by: Stefan Binding <sbinding@xxxxxxxxxxxxxxxxxxxxx>
>> ---
>>  drivers/platform/x86/serial-multi-instantiate.c | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/platform/x86/serial-multi-instantiate.c b/drivers/platform/x86/serial-multi-instantiate.c
>> index 5362f1a7b77c..15ef2f3c442e 100644
>> --- a/drivers/platform/x86/serial-multi-instantiate.c
>> +++ b/drivers/platform/x86/serial-multi-instantiate.c
>> @@ -194,6 +194,7 @@ static int smi_i2c_probe(struct platform_device *pdev, struct smi *smi,
>>  		strscpy(board_info.type, inst_array[i].type, I2C_NAME_SIZE);
>>  		snprintf(name, sizeof(name), "%s-%s.%d", dev_name(dev), inst_array[i].type, i);
>>  		board_info.dev_name = name;
>> +		board_info.fwnode = acpi_fwnode_handle(adev);
>>  
>>  		ret = smi_get_irq(pdev, adev, &inst_array[i]);
>>  		if (ret < 0)
> 
> I'm afraid that making this change is not as straight forward as it looks.
> 
> I know that I have tried to do this in the past and it failed.
> 
> IIRC there were 3 problems:
> 
> 1. I was expecting this to also allow the driver for the instantiated
> i2c-client to be able to bind using an acpi_match_table but that
> unfortunately does not work. acpi_match_table matches only work for
> the first physical_node linked under
> /sys/bus/acpi/devices/xxxx:xx/physical_node and that is the platform
> device to which serial-multi-instantiate.c binds. The i2c_client becomes
> the second physical node.  Note this is not really an issue,
> just something to be aware of.
> 
> 
> 2. This causes the i2c-core to use the first IRQ resource in the ACPI
> fwnode as client->irq for any clients for which we do not set an
> IRQ when instantiating. Which may very well be wrong. Sometimes that
> IRQ is only valid for the first i2c-client which we instantiate; and
> not for the others! And sometimes it is a problem because it may
> point to an irqchip for which we never wrote a driver leading to
> all probes of the i2c-client failing with -EPROBE_DEFER, see:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=d1d84bb95364ed604015c2b788caaf3dbca0262f
> 
> Note that patch has been reverted since that specific -EPROBE_DEFER
> issue has been solved by making the ACPI core instantiate a
> platform_device instead of an i2c_client (in this case we
> did not need the actual i2c_client at all).
> 
> The current i2c-core code has a (!client-irq) test guarding its
> code of trying to use the first ACPI fwnode IRQ resource.
> 
> So we could disable this by setting client->irq = -ENOENT in
> serial-multi-instantiate.c when (inst->flags & IRQ_RESOURCE_TYPE) ==
> IRQ_RESOURCE_NONE). But that will introduce a new problem. Many
> i2c-drivers check if there is an IRQ for them to use by doing:
> "if (client->irq) request_irq(client->irq, ...)" but then with
> error checking/so setting client->irq to -ENOENT will cause
> the request_irq to fail, leading the probe to fail.
> 
> So before you can write a patch setting client->irq = -ENOENT
> when (inst->flags & IRQ_RESOURCE_TYPE) == IRQ_RESOURCE_NONE),
> you would first need to patch all i2c-drivers for clients
> instantiated through serial-multi-instantiate.c changing:
> 
> 	if (client->irq) {
> 		...
> 	}
> 
> to:
> 
> 	if (client->irq > 0) {
> 		...
> 	}
> 
> Note this is not as bad as it sounds, since there are only
> a few drivers for clients instantiated by serial-multi-instantiate.c .

Possibly a  nicer way to fix this would be to make the i2c-core change
client->irq to 0 if it is -ENOENT before calling the i2c_driver's
probe method, thus fixing things centrally for all i2c-drivers without
needing to audit/patch them all. Specifically in:

drivers/i2c/i2c-core-base.c: i2c_device_probe() change:

	if (!client->irq) {
		...
	}

to:

	if (!client->irq) {
		...
	} else if (client->irq == -ENOENT) {
		client->irq = 0; /* Drivers expect 0 for "no-IRQ" */
	}

And maybe as Andy suggested, handle at least the IRQ in
i2c_acpi_new_device_by_fwnode() by adding something like that there:

	/* Disable the i2c-core attempting to get an IRQ from ACPI itself */
	if (!board_info->irq)
		board_info->irq= -ENOENT;

I also agree with Andy that setting board_info->fw_node would be done
there ideally too. But then you would need to extend the audit of
impacted drivers mentioned below to also include drivers for
i2c-clients instantiated through other code-paths calling
i2c_acpi_new_device_by_fwnode()  (of which there are not many,
but there are a few others).

> 3. Some drivers may check for an ACPI companion device and then
> change their behavior. So all drivers for clients instantiated
> through serial-multi-instantiate.c will need to be audited for
> this (and a summary of this audit needs to be added to the commit
> msg).

Regards,

Hans