Denys Vlasenko has reported two bug about page table check on arm64.
On arm64, pmd_present() contains non-leaf pmd and invalid pmd too.
When collapse hugepage, the pmd is non-leaf and should skip the check.
Use pmd_leaf() instead of pmd_present().
When split hugepage, the pmd will be marked as invalid and then populate.
So we should decrease file_map_count when invalid pmd and then increase
when populate the pmd. Add pmd_valid() check.
v1->v2: Update comment and optimize the code by moving p?d_valid() at
first place suggested by Mark.
Liu Shixin (2):
arm64/mm: fix incorrect file_map_count for non-leaf pmd/pud
arm64/mm: fix incorrect file_map_count for invalid pmd/pud
arch/arm64/include/asm/pgtable.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
--
2.25.1
On Thu, 17 Nov 2022 15:56:00 +0800, Liu Shixin wrote:
> Denys Vlasenko has reported two bug about page table check on arm64.
> On arm64, pmd_present() contains non-leaf pmd and invalid pmd too.
>
> When collapse hugepage, the pmd is non-leaf and should skip the check.
> Use pmd_leaf() instead of pmd_present().
>
> When split hugepage, the pmd will be marked as invalid and then populate.
> So we should decrease file_map_count when invalid pmd and then increase
> when populate the pmd. Add pmd_valid() check.
>
> [...]
Applied to arm64 (for-next/fixes), thanks!
[1/2] arm64/mm: fix incorrect file_map_count for non-leaf pmd/pud
https://git.kernel.org/arm64/c/5b47348fc0b1
I only merged the first patch in this series as Will had some questions
on the second patch (it does seem weird that the pud and pmd functions
are different w.r.t. the p*d_user() checks).
--
Catalin