Re: [PATCH 3/3] [RFC] x86/devmem: remove low 1MB hack for x86-64

["Arnd Bergmann" <arnd@arndb.de>]

On Thu, May 22, 2025, at 00:14, Dan Williams wrote:
> Arnd Bergmann wrote:
>>
>> Unlike the other two patches in this series, this one is expected to
>> change the behavior on x86-64 kernels, which has the risk of
>> regressions, but seems worthwhile to me.
>> 
>> Are there any reasons left for keeping these hacks?
>
> Kees did this search which seems to suggest that there is still code out
> there that may not be prepared for a behavior change here:
>
> http://lore.kernel.org/202504101926.0F8FB73@keescook
>
> Maybe those paths fallback to dmi sysfs or other mechanisms for digging through
> BIOS data, but I do not think we can know for sure that this removal is
> regression free ahead of time.

I looked at those three and from what I can tell, two attempt
to access PCI BAR areas, which should not change with my patch:
if they are busy or exclusive, they are already disallowed, otherwise
they can still be mapped.

The third one maps the BIOS area at 0xf0000, and as far as I can tell
the hack explicitly allowed mapping that even though it is marked
busy on x86-64 since 5d94e81f69d4 ("x86: Introduce pci_map_biosrom()").

Is there any downside to marking this one non-busy and still allowing
the ROM to be mapped? Would that bring back the issue of conflicting
mapping flags between kernel and userspace?

--- a/arch/x86/kernel/probe_roms.c
+++ b/arch/x86/kernel/probe_roms.c
@@ -27,7 +27,7 @@ static struct resource system_rom_resource = {
        .name   = "System ROM",
        .start  = 0xf0000,
        .end    = 0xfffff,
-       .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
+       .flags  = IORESOURCE_READONLY | IORESOURCE_MEM
 };
 
 static struct resource extension_rom_resource = {

> I would be interested to see though, so:
>
> Acked-by: Dan Williams <dan.j.williams@intel.com>

Thanks,

      ARnd