Re: [cxl:pending 13/21] drivers/cxl/pci.c:511:8-33: WARNING: Threaded IRQ with no primary handler requested without IRQF_ONESHOT (unless it is nested IRQ)

[Davidlohr Bueso <dave@stgolabs.net>]

On Tue, 31 Jan 2023, Dan Williams wrote:

>Ira Weiny wrote:
>[..]
>> > Considering MSIs are edge triggered interrupts, this should be benign.
>>
>> I'm trying to understand this a bit more.  So I looked in more detail.  I
>> believe Dan suggested that I could use irq_default_primary_handler() in
>> this case by passing NULL.  And this would handle threading the irq
>> processing for me.
>>
>> It looks from a software standpoint this should work.  But the kdocs
>> really imply that not masking the irq in the handler may be an issue.
>> Thus the 'requirement' of using IRQF_ONESHOT.
>>
>> Is this somehow different with MSI?  Is this a case of the kdocs being a
>> bit out of date for MSI?
>
>My reading, may be from the changelog and not the kdocs, was that at
>request_irq() time it is difficult to discern level triggered vs
>edge-triggered. So for the default primary handler, that does nothing to
>quiet a screaming level-triggered interrupt, just require oneshot mode.

Right and it's only level-triggered that are the problem after done with
the hardirq handler. So in the msi/cxl case we already know that all this
is edge (sent once and forgotten by the hardware), which is why I was saying
it's harmless. It's still probably a good idea to have it explicitly oneshot
semantics as Dan's patch.

Thanks,
Davidlohr