Re: [PATCH v5] staging: gpib: Fix i386 build issue

[Guenter Roeck]

On 12/10/24 01:02, Geert Uytterhoeven wrote:
> Hi Greg,
>
> On Tue, Dec 10, 2024 at 9:39 AM Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> > On Tue, Dec 10, 2024 at 08:52:08AM +0100, Geert Uytterhoeven wrote:
> > > On Mon, Dec 9, 2024 at 5:18 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> > > > The underlying problem is that the code uses a pointer to store the physical
> > > > address. That doesn't work if sizeof(pointer) < sizeof(physical address),
> > > > which affects systems with X86_PAE enabled. I have not seen the problem
> > > > anywhere else.
> > >
> > > I could reproduce the build issue on ARM, with CONFIG_ARM_LPAE=y,
> > > which is not enabled by allmodconfig.

I think it may actually also affect mips (CPU_MIPS32_R5_XPA) and arc
(ARC_HAS_PAE40).

> > So does that mean this patch is incorrect?
>
> Purely from an arch-agnostic LPAE PoV, it should be:
>
>      depends on 64BIT || !PHYS_ADDR_T_64BIT

Dave, would you mind submitting another version of your patch with the above
dependencies ?

> However, that assumes the driver actually works on 64-bit or non-x86.
> Perhaps people keep an old i386 to control their GPIB gear?
>
> The drivers do not use ioremap(), but just cast the PCI resource
> addresses to void * pointers. No idea if that works on x86_64.

I did wonder about that as well, but unfortunately it isn't easy
to change, and would most definitely require testing on real hardware.

> It will probably crash spectacularly on non-x86...

I'd guess so. However, that isn't something we can address here.
All we can make sure is that the code compiles.

Thanks,
Guenter