Catalin Marinas <catalin.mari...@arm.com> writes:
> On Wed, Apr 12, 2017 at 03:04:57PM +0100, Punit Agrawal wrote:
>> When memory failure is enabled, a poisoned hugepage pte is marked as a
>> swap entry. huge_pte_offset() does not return the poisoned page table
>> entries when it encounters PUD/PMD hugepages.
>>
>> This behaviour of huge_pte_offset() leads to error such as below when
>> munmap is called on poisoned hugepages.
>>
>> [ 344.165544] mm/pgtable-generic.c:33: bad pmd 000000083af00074.
>>
>> Fix huge_pte_offset() to return the poisoned pte which is then
>> appropriately handled by the generic layer code.
>>
>> Signed-off-by: Punit Agrawal <punit.agra...@arm.com>
>> Cc: Catalin Marinas <catalin.mari...@arm.com>
>> Cc: Steve Capper <steve.cap...@arm.com>
>> Cc: David Woods <dwo...@mellanox.com>
>> ---
>> arch/arm64/mm/hugetlbpage.c | 20 +++++++++++++++-----
>> 1 file changed, 15 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/arm64/mm/hugetlbpage.c b/arch/arm64/mm/hugetlbpage.c
>> index 7514a000e361..5f1832165d69 100644
>> --- a/arch/arm64/mm/hugetlbpage.c
>> +++ b/arch/arm64/mm/hugetlbpage.c
>> @@ -143,15 +143,24 @@ pte_t *huge_pte_offset(struct mm_struct *mm, unsigned
>> long addr)
>> pr_debug("%s: addr:0x%lx pgd:%p\n", __func__, addr, pgd);
>> if (!pgd_present(*pgd))
>> return NULL;
>> - pud = pud_offset(pgd, addr);
>> - if (!pud_present(*pud))
>> - return NULL;
>>
>> - if (pud_huge(*pud))
>> + pud = pud_offset(pgd, addr);
>> + /*
>> + * In case of HW Poisoning, a hugepage pud/pmd can contain
>> + * poisoned entries. Poisoned entries are marked as swap
>> + * entries.
>> + *
>> + * For puds/pmds that are not present, check to see if it
>> + * could be a swap entry (!present and !none).
>> + */
>> + if ((!pte_present(pud_pte(*pud)) && !pud_none(*pud)) || pud_huge(*pud))
>> return (pte_t *)pud;
>
> Since we use puds as huge pages, can we just change pud_present() to
> match the pmd_present()? I'd like to see similar checks for pud and pmd,
> it would be easier to follow. Something like (unchecked):
>
> if (pud_none(*pud))
> return NULL;
> /* swap or huge page */
> if (!pud_present(*pud) || pud_huge(*pud))
> return (pte_t *)pud;
> /* table; check the next level */
>
>> +
>> pmd = pmd_offset(pud, addr);
>> - if (!pmd_present(*pmd))
>> + if (pmd_none(*pmd))
>> return NULL;
>> + if (!pmd_present(*pmd) && !pmd_none(*pmd))
>> + return (pte_t *)pmd;
>
> At this point, we already know that pmd_none(*pmd) is false, no ned to
> check it again.
Indeed - I was avoiding changing the function to drop contiguous
hugepage support which follows this hunk.
I've made changes locally based on your suggestion and will post a
revised version after the merge window.
Thanks,
Punit
