On Wed, Jan 06, 2021 at 08:12:27PM +0000, Ionela Voinescu wrote:

> Initially I though it always only makes sense to have a __iomem pointer.
> That is, it only makes sense to have a pointer with a cookie attached
> specifying that it addresses a device memory space that should only be
> accessed using special functions.
> 
> But then you've got something like this in drivers/input/serio/apbps2.c:
> struct apbps2_regs {
> 	u32 __iomem data;	/* 0x00 */
> 	u32 __iomem status;	/* 0x04 */
> 	u32 __iomem ctrl;	/* 0x08 */
> 	u32 __iomem reload;	/* 0x0c */
> };
> struct apbps2_priv {
> 	struct serio		*io;
> 	struct apbps2_regs	*regs;
> };
> [..] (followed by)
> ioread32be(&priv->regs->status)
> 
> which I think is correct despite contradicting my assumption, but it's
> the only example I've found in the kernel.

Frankly, I would rather turn that into
	struct apbps2_regs	__iomem *regs;
and striped the individual field qualifiers...