[PATCH] arm64: account for sparsemem section alignment when choosing vmemmap offset

[Ard Biesheuvel]

On 8 March 2016 at 22:12, Catalin Marinas <catalin.marinas at arm.com> wrote:
> On Tue, Mar 08, 2016 at 09:09:29PM +0700, Ard Biesheuvel wrote:
>> Commit dfd55ad85e4a ("arm64: vmemmap: use virtual projection of linear
>> region") fixed an issue where the struct page array would overflow into the
>> adjacent virtual memory region if system RAM was placed so high up in
>> physical memory that its addresses were not representable in the build time
>> configured virtual address size.
>>
>> However, the fix failed to take into account that the vmemmap region needs
>> to be relatively aligned with respect to the sparsemem section size, so that
>> a sequence of page structs corresponding with a sparsemem section in the
>> linear region appears naturally aligned in the vmemmap region.
>>
>> So round up vmemmap to sparsemem section size. Since this essentially moves
>> the projection of the linear region up in memory, also revert the reduction
>> of the size of the vmemmap region.
>>
>> Fixes: dfd55ad85e4a ("arm64: vmemmap: use virtual projection of linear region")
>> Tested-by: Mark Langsdorf <mlangsdo at redhat.com>
>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
>> ---
>>  arch/arm64/include/asm/pgtable.h | 5 +++--
>>  1 file changed, 3 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
>> index f50608674580..819aff5d593f 100644
>> --- a/arch/arm64/include/asm/pgtable.h
>> +++ b/arch/arm64/include/asm/pgtable.h
>> @@ -40,7 +40,7 @@
>>   * VMALLOC_END: extends to the available space below vmmemmap, PCI I/O space,
>>   *   fixed mappings and modules
>>   */
>> -#define VMEMMAP_SIZE         ALIGN((1UL << (VA_BITS - PAGE_SHIFT - 1)) * sizeof(struct page), PUD_SIZE)
>> +#define VMEMMAP_SIZE         ALIGN((1UL << (VA_BITS - PAGE_SHIFT)) * sizeof(struct page), PUD_SIZE)
>
> I think we could have extended the existing halved VMEMMAP_SIZE by
> PAGES_PER_SECTION * sizeof(struct page) to cope with the alignment but I
> don't think it's worth.
>

Indeed. But it is only temporary anyway, since the problem this patch
solves does not exist anymore in for-next/core, considering that
memstart_addr itself should be sufficiently aligned by construction.
So I intend to propose a revert of this change, after -rc1 perhaps?

>>
>>  #ifndef CONFIG_KASAN
>>  #define VMALLOC_START                (VA_START)
>> @@ -52,7 +52,8 @@
>>  #define VMALLOC_END          (PAGE_OFFSET - PUD_SIZE - VMEMMAP_SIZE - SZ_64K)
>>
>>  #define VMEMMAP_START                (VMALLOC_END + SZ_64K)
>> -#define vmemmap                      ((struct page *)VMEMMAP_START - (memstart_addr >> PAGE_SHIFT))
>> +#define vmemmap                      ((struct page *)VMEMMAP_START - \
>> +                              SECTION_ALIGN_DOWN(memstart_addr >> PAGE_SHIFT))
>
> It looks fine to me:
>
> Acked-by: Catalin Marinas <catalin.marinas at arm.com>
>
> Will would probably pick it up tomorrow (and add a cc stable as well).
>

Thanks