Re: [PATCH] x86/mm/pat: Fix BUG_ON in mmap_mem on QEMU/i386

[Toshi Kani <toshi.kani@hpe.com>]

On Tue, 2016-04-05 at 09:24 -0600, Toshi Kani wrote:
> +xen-devl
> 
> On Tue, 2016-04-05 at 13:09 +0200, Borislav Petkov wrote:
> > On Fri, Apr 01, 2016 at 04:19:45PM -0600, Toshi Kani wrote:
> > > 
 :
> > > 
> > > When the system does not have much memory, 'high_memory' points to
> >
> > What does "much memory" mean, exactly?
>
> I meant to say when a 32-bit system does not have ZONE_HIGHMEM,
> __pa(high_memory) points to the maximum memory address + 1.
> 
> I will remove this sentence since it is irrelevant to this BUG_ON.  Even
> if a 32-bit system does have ZONE_HIGHMEM, slow_virt_to_phys() still
> returns 0 for high_memory because it is set to the maximum direct mapped
> address + 1 in this case.  This address is not covered by page table,
> either.
> 
> But this made me realized that this high_memory check can be harmful in
> such case, ie. __pa(high_memory) is not the maximum memory address when
> ZONE_HIGHMEM is present.
> 
> I assume when this code block was originally added, legacy systems
> without MTRRs did not have ZONE_HIGHMEM.  However, MTRRs are also
> disabled on Xen. Reactivating this code may cause an issue on Xen 32-bit
> guests with ZONE_HIGHMEM.
> 
> Question to Xen folks: Does Xen support 32-bit guests with ZONE_HIGHMEM?
> 
> If yes, a safer fix may be to remove this code block since it was
> deadcode anyway...

I have not heard a confirmation from Xen folks, but I believe ZONE_HIGHMEM
is supported on 32-bit Xen guests.  So, unless someone has an objection, I
am going to remove this code block in the next version of this patch.

Thanks,
-Toshi