On Fri, Sep 4, 2015 at 8:14 PM, Frank Rowand <frowand.list@xxxxxxxxx> wrote:
> On 9/4/2015 12:12 PM, David Daney wrote:
>> From: David Daney <david.daney@xxxxxxxxxx>
>>
>> It is perfectly legitimate for a PCI device to have an
>> PCI_INTERRUPT_PIN value of zero. This happens if the device doesn't
>> use interrupts, or on PCIe devices, where only MSI/MSI-X are
>> supported.
>>
>> Silence the annoying "of_irq_parse_pci() failed with rc=-19" error
>> messages by making them conditional on !-ENODEV (which can only be
>> produced in the PCI_INTERRUPT_PIN == 0 case).
>>
>> Signed-off-by: David Daney <david.daney@xxxxxxxxxx>
>> ---
>> drivers/of/of_pci_irq.c | 4 +++-
>> 1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/of/of_pci_irq.c b/drivers/of/of_pci_irq.c
>> index 1710d9d..33d242a 100644
>> --- a/drivers/of/of_pci_irq.c
>> +++ b/drivers/of/of_pci_irq.c
>> @@ -106,7 +106,9 @@ int of_irq_parse_and_map_pci(const struct pci_dev *dev, u8 slot, u8 pin)
>>
>> ret = of_irq_parse_pci(dev, &oirq);
>> if (ret) {
>> - dev_err(&dev->dev, "of_irq_parse_pci() failed with rc=%d\n", ret);
>> + if (ret != -ENODEV)
>> + dev_err(&dev->dev,
>> + "of_irq_parse_pci() failed with rc=%d\n", ret);
>> return 0; /* Proper return code 0 == NO_IRQ */
>> }
>>
>>
>
> It is not safe to assume that the functions that of_irq_parse_pci() calls
> will never be modified to return -ENODEV, thus resulting in of_irq_parse_pci()
> returning -ENODEV for a reason other than PCI_INTERRUPT_PIN == 0.
Yes, but we're talking about a print statement.
>
> A more robust solution would be something like:
>
>
> (1) Change of_irq_parse_pci() to _of_irq_parse_pci(), adding an argument and
> use it to report the case of PCI_INTERRUPT_PIN == 0.
>
> static int _of_irq_parse_pci(const struct pci_dev *pdev, struct of_phandle_args *out_irq, int *no_pin)
> {
>
> ...
> *no_pin = 0;
> ...
> /* No pin, exit */
> if (pin == 0) {
> *no_pin = 1;
> return -ENODEV;
> }
> ...
>
>
> int of_irq_parse_pci(const struct pci_dev *pdev, struct of_phandle_args *out_irq)
> {
> int no_pin;
> return _of_irq_parse_pci(pdev, out_irq, &no_pin)
> }
>
>
> (2) Then the fix to of_irq_parse_and_map_pci() would be:
>
> + int no_pin;
>> ret = of_irq_parse_pci(dev, &oirq, &no_pin);
>> if (ret) {
>> - dev_err(&dev->dev, "of_irq_parse_pci() failed with rc=%d\n", ret);
>> + if (!no_pin)
>> + dev_err(&dev->dev,
>> + "of_irq_parse_pci() failed with rc=%d\n", ret);
>> return 0; /* Proper return code 0 == NO_IRQ */
>> }
>
>
> I'm not sure I like my solution, there might be a better way.
I don't like it. That's way too complex for just silencing an
erroneous error message.
Perhaps just move the error message into of_irq_parse_pci and then you
can control the print more easily. Or just change to dev_dbg would be
okay by me.
> I also noticed another bug while looking at of_irq_parse_pci(). It returns
> the non-zero return value from pci_read_config_byte(). But that value is
> one of the PCI function error values from include/linux/pci.h, such as:
>
> #define PCIBIOS_BAD_REGISTER_NUMBER 0x87
>
> instead of a negative errno.
I was puzzled by why this is not standard error codes a while back. My
best guess is that that there is some legacy here. Changing error
values on widely used functions is impossible to audit. NO_IRQ being 0
or -1 is one such case.
Rob
--
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
