Re: [PATCH] irq: Remove unnecessary warning with affinity_hint

[Seiichi Ikarashi <s.ikarashi@jp.fujitsu.com>]

On 2015-04-08 16:39, Ingo Molnar wrote:
> 
> * Seiichi Ikarashi <s.ikarashi@jp.fujitsu.com> wrote:
> 
>> Hi,
>>
>> On 2015-04-08 15:28, Ingo Molnar wrote:
>>>
>>> * Seiichi Ikarashi <s.ikarashi@jp.fujitsu.com> wrote:
>>>
>>>> Hi,
>>>>
>>>> If you turn off a PCI device whose driver has set affinity_hint,
>>>> you will get warning message which does _not_ explain the reason
>>>> why it appeared from the user's point of view.
>>>>
>>>>   # echo 0 > /sys/bus/pci/slots/65/power
>>>>
>>>>   Apr 28 20:29:39 localhost kernel: ------------[ cut here ]------------
>>>>   Apr 28 20:29:39 localhost kernel: WARNING: at kernel/irq/manage.c:1002 __free_irq+0x22d/0x250() (Tainted: P           ---------------   )
>>>>   (snip)
>>>>
>>>> Users will misunderstand some problem has happened
>>>> even though he or she succeeded to turn off the device.
>>>> I suppose this warning was originally for a debug purpose
>>>> for driver developers and has incidentally been left.
>>>>
>>>> Just remove the warning is good and enough.
>>>>
>>>> Signed-off-by: Seiichi Ikarashi <s.ikarashi@jp.fujitsu.com>
>>>>
>>>> --- a/kernel/irq/manage.c
>>>> +++ b/kernel/irq/manage.c
>>>> @@ -1335,7 +1335,7 @@ static struct irqaction *__free_irq(unsi
>>>>  
>>>>  #ifdef CONFIG_SMP
>>>>  	/* make sure affinity_hint is cleaned up */
>>>> -	if (WARN_ON_ONCE(desc->affinity_hint))
>>>> +	if (desc->affinity_hint)
>>>>  		desc->affinity_hint = NULL;
>>>
>>> Well, drivers that are using irq_set_affinity_hint() are expected to 
>>> call:
>>>
>>> 	irq_set_affinity_hint(irq, NULL);
>>>
>>> to clear the affinity mask, before releasing the irq. This warning 
>>> flags drivers that forgot to do that and which might thus leak a 
>>> dynamically allocated CPU mask (and/or other resources).
>>
>> Calling irq_set_affinity_hint(irq, NULL) does not guarantee that the 
>> driver does not forget to deallocate a dynamically allocated CPU 
>> mask and/or other resources. [...]
> 
> I said 'might leak', not 'guaranteed to leak'.

Yes, I know.
I wrote it because I was not sure about the primary purpose of
showing the warning message.

> 
> Calling irq_set_affinity_hint(irq, NULL) is the way this kernel API is 
> specified to be used. Forgetting to do it is a kernel driver bug and 
> triggers a warning message from the kernel's IRQ subsystem.
> 
>> [...] But if calling it with NULL 2nd-arg before releasing the irq 
>> is a virtual rule of using irq_set_affinity_hint() interface, I 
>> understand it.
>>
>>> Feel free to turn the warning message into a more informative 
>>> WARN() that will blame the driver that triggered it, if the stack 
>>> dump into the driver wasn't a clue enough ...
>>
>> Still, I do not know leaving the warning message is effective to 
>> prevent drivers from potentially leaking resource... considering a 
>> kind of cost-effectivenss. Business users (not developers) hate such 
>> kind of messages for developers.
> 
> it's a warning message pointing out a kernel bug: that 
> irq_set_affinity_hint(irq, NULL) was not called properly.
> 
> Messages pointing out kernel bugs should be fixed, not hidden.

OK, the conclusion is that a kernel bug on using irq_set_affinity_hint().

Thanks, Ingo.

Seiichi