Re: [PATCH v1 1/2] ata: libahci_platform: Get rid of dup message when IRQ can't be retrieved

[Sergey Shtylyov]

On 12/10/21 11:47 AM, Andy Shevchenko wrote:

>>> platform_get_irq() will print a message when it fails.
>>> No need to repeat this.
>>>
>>> While at it, drop redundant check for 0 as platform_get_irq() spills
>>> out a big WARN() in such case.
>>
>> The reason you should be able to remove the "if (!irq)" test is that
>> platform_get_irq() never returns 0. At least, that is what the function kdoc
>> says. But looking at platform_get_irq_optional(), which is called by
>> platform_get_irq(), the out label is:
>>
>>         WARN(ret == 0, "0 is an invalid IRQ number\n");
>>         return ret;
>>
>> So 0 will be returned as-is. That is rather weird. That should be fixed to
>> return -ENXIO:
>>
>>         if (WARN(ret == 0, "0 is an invalid IRQ number\n"))
>>                 return -ENXIO;

   -ENXIO seems to me more fitting indeed (than -EINVAL that I used).

> 
> No, this is wrong for the same reasons I explained to Sergey.

   I fail to understand you, sorry. We're going in circles, it seems... :-/

> The problem is that this is _optional API and it has been misdesigned.
> Replacing things like above will increase the mess.

   What's wrong with replacing IRQ0 with -ENXIO now? platform_get_irq_optional()
(as in your patch) could then happily return 0 ISO -ENXIO. Contrarywise, if we don't
replace IRQ0 with -ENXIO, platform_get_irq_optional() will return 0 for both IRQ0
and missing IRQ! Am I clear enough? If you don't understand me now, I don't know what
to say... :-/

> 
>>         return ret;
>>
>> Otherwise, I do not think that removing the "if (!irq)" hunk is safe. no ?
> 
> No. This is not a business of the caller to workaround implementation
> details (bugs) of the core APIs.
> If something goes wrong, then it's platform_get_irq() to blame, and
> not the libahci_platform.

   I'm repeating myself already: we don't work around the bug in platform_get_irq(),
we're working around the driver subsystems that treat 0 specially (and so don't
support IRQ0); libata treats 0 as an indication of the polling mode (moreover,
it will curse if you pass to it both IRQ == 0 and a pointer to an interrupt handler!
Am I clear enough this time? :-)

MBR, Sergey