Re: [PATCH v1 3/4] scsi - convert mvme146_scsi.c to platform device

[Michael Schmitz <schmitzmic@gmail.com>]

Hi Geert,

Am 11.07.2022 um 20:27 schrieb Geert Uytterhoeven:
> Hi Michael,
>
> On Mon, Jul 11, 2022 at 9:57 AM Michael Schmitz <schmitzmic@gmail.com> wrote:
>> Am 11.07.2022 um 19:16 schrieb Geert Uytterhoeven:
>>> On Sun, Jul 10, 2022 at 6:12 PM Arnd Bergmann <arnd@kernel.org> wrote:
>>>> A resource would pass a phys_addr_t token, but the driver expects a
>>>> virtual address that should be an __iomem pointer. The MMIO area
>>>> already gets mapped into virtual addresses in arch/m68k/kernel/head.S,
>>>> so it makes sense to skip the extra ioremap() and just use the address,
>>>> but then you can't pass it as an IORESOUCRE_MEM token and should
>>>> use platform_data with the pointer instead.
>>
>> OK, got it now (I had missed the physical/virtual mismatch entirely).
>
> And the __iomem is there so we can catch mistakes using sparse
> (make C=2 path/to/file/.o).

I'll need to make a habit of that ...

>>>> The alternative is to do it the normal way and pass the physical address
>>>> as a resource, that you can pass into devm_platform_ioremap_resource()
>>>> or a similar helper.
>>>
>>> I would prefer the latter.  While head.S sets up the mapping,
>>> __ioremap() does not have support for this on the mvme platform,
>>> so this has to be added first. Look at the amiga and virt platforms
>>> for examples.
>>
>> I see - doesn't look too hard to do, and should not affect any other
>> existing code.
>> Is it worth adding the same support for Atari as well?
>
> From a quick glance at arch/m68k/kernel/head.S, it seems that
> on Atari there is no identity mapping (the high I/O area is mapped
> to the virtual low area).  That means __ioremap() and iounmap()
> wouldn't be symmetrical, but it can be done.

As I read it, it's the other way: virtual 0xffxxxxxx is mapped to phys. 
0x00ffxxxx, and all hardware addresses are given in the upper window 
(Medusa/Hades use that window only, and have identity mappings there).

So returning identity mapping in the high window would work AFAICS. I'll 
give that a try sometime.

> Note that on Amiga we only use the identity shortcut for Zorro III
> memory (and only for the first half?), i.e. ioremap() on Zorro II I/O
> does add new mappings.  Hence most Zorro II drivers use ZTWO_VADDR()
> instead of ioremap().

I see ZTWO_VADDR() already returns void __iomem* ... Fixing this patch 
would sort out the m68k drivers, leaving SGI. I'll have to see whether I 
can dig up a SGI workstation with this driver; might be easier than 
getting hold of any Amgias here Dowh Under.

Cheers,

	Michael


>
> Gr{oetje,eeting}s,
>
>                         Geert
>
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org
>
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
>                                 -- Linus Torvalds
>