Re: [irqchip: irq/irqchip-next] irqdomain: Protect the linear revmap with RCU

[Guenter Roeck]

On 7/6/21 2:24 AM, Marc Zyngier wrote:
> On Mon, 05 Jul 2021 21:36:36 +0100,
> Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> > On 7/5/21 11:43 AM, Marc Zyngier wrote:
>
> > > It definitely helps, and confirms my hunch. With the patch below, I'm
> > > not getting the warnings anymore. I'm pretty sure a number of other
> > > MIPS systems suffer from similar issues, which I'll address similarly.
> > >
> > > Please let me know if that addresses the issue on your end.
> >
> > Yes, it does. Feel free to add
> >
> > Tested-by: Guenter Roeck <linux@xxxxxxxxxxxx>
> >
> > to the real patch.
>
> Thanks.
>
> > Now the big question: Why does this only affect 32-bit little endian
> > mips images, but not the matching big endian images, nor 64-bit images ?

Actually they do.

> Are you sure these images are using the exact same HW? A bunch (most?)
> of MIPS systems do not use irqdomains in their root interrupt
> handling, so this issue wouldn't be visible (irq_enter() will already
> have been done for the chained interrupt handling).
>
> FWIW, I can reproduce the problem by switching your mipsel config to
> BE, and adding this patch fixes it.

Yes, turns out I did not have the necessary debugging options enabled
in my other mips tests.

Thanks,
Guenter