Zi Yan, I'm hoping you might have some input on the below...
On 30/04/2024 14:31, Ryan Roberts wrote:
> __split_huge_pmd_locked() can be called for a present THP, devmap or
> (non-present) migration entry. It calls pmdp_invalidate()
> unconditionally on the pmdp and only determines if it is present or not
> based on the returned old pmd.
>
> But arm64's pmd_mkinvalid(), called by pmdp_invalidate(),
> unconditionally sets the PMD_PRESENT_INVALID flag, which causes future
> pmd_present() calls to return true - even for a swap pmd. Therefore any
> lockless pgtable walker could see the migration entry pmd in this state
> and start interpretting the fields (e.g. pmd_pfn()) as if it were
> present, leading to BadThings (TM). GUP-fast appears to be one such
> lockless pgtable walker.
>
> While the obvious fix is for core-mm to avoid such calls for non-present
> pmds (pmdp_invalidate() will also issue TLBI which is not necessary for
> this case either), all other arches that implement pmd_mkinvalid() do it
> in such a way that it is robust to being called with a non-present pmd.
OK the plot thickens; The tests I wrote to check that pmd_mkinvalid() is safe for swap entries fails on x86_64. See below...
> So it is simpler and safer to make arm64 robust too. This approach means
> we can even add tests to debug_vm_pgtable.c to validate the required
> behaviour.
>
> This is a theoretical bug found during code review. I don't have any
> test case to trigger it in practice.
>
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: 53fa117bb33c ("arm64/mm: Enable THP migration")
> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx>
> ---
>
> Hi all,
>
> v1 of this fix [1] took the approach of fixing core-mm to never call
> pmdp_invalidate() on a non-present pmd. But Zi Yan highlighted that only arm64
> suffers this problem; all other arches are robust. So his suggestion was to
> instead make arm64 robust in the same way and add tests to validate it. Despite
> my stated reservations in the context of the v1 discussion, having thought on it
> for a bit, I now agree with Zi Yan. Hence this post.
>
> Andrew has v1 in mm-unstable at the moment, so probably the best thing to do is
> remove it from there and have this go in through the arm64 tree? Assuming there
> is agreement that this approach is right one.
>
> This applies on top of v6.9-rc5. Passes all the mm selftests on arm64.
>
> [1] https://lore.kernel.org/linux-mm/20240425170704.3379492-1-ryan.roberts@xxxxxxx/
>
> Thanks,
> Ryan
>
>
> arch/arm64/include/asm/pgtable.h | 12 +++++--
> mm/debug_vm_pgtable.c | 61 ++++++++++++++++++++++++++++++++
> 2 files changed, 71 insertions(+), 2 deletions(-)
>
> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> index afdd56d26ad7..7d580271a46d 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -511,8 +511,16 @@ static inline int pmd_trans_huge(pmd_t pmd)
>
> static inline pmd_t pmd_mkinvalid(pmd_t pmd)
> {
> - pmd = set_pmd_bit(pmd, __pgprot(PMD_PRESENT_INVALID));
> - pmd = clear_pmd_bit(pmd, __pgprot(PMD_SECT_VALID));
> + /*
> + * If not valid then either we are already present-invalid or we are
> + * not-present (i.e. none or swap entry). We must not convert
> + * not-present to present-invalid. Unbelievably, the core-mm may call
> + * pmd_mkinvalid() for a swap entry and all other arches can handle it.
> + */
> + if (pmd_valid(pmd)) {
> + pmd = set_pmd_bit(pmd, __pgprot(PMD_PRESENT_INVALID));
> + pmd = clear_pmd_bit(pmd, __pgprot(PMD_SECT_VALID));
> + }
>
> return pmd;
> }
> diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
> index 65c19025da3d..7e9c387d06b0 100644
> --- a/mm/debug_vm_pgtable.c
> +++ b/mm/debug_vm_pgtable.c
> @@ -956,6 +956,65 @@ static void __init hugetlb_basic_tests(struct pgtable_debug_args *args) { }
> #endif /* CONFIG_HUGETLB_PAGE */
>
> #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +#if !defined(__HAVE_ARCH_PMDP_INVALIDATE) && defined(CONFIG_ARCH_ENABLE_THP_MIGRATION)
> +static void __init swp_pmd_mkinvalid_tests(struct pgtable_debug_args *args)
> +{
Printing various values at different locations in this function for debug:
> + unsigned long max_swap_offset;
> + swp_entry_t swp_set, swp_clear, swp_convert;
> + pmd_t pmd_set, pmd_clear;
> +
> + /*
> + * See generic_max_swapfile_size(): probe the maximum offset, then
> + * create swap entry will all possible bits set and a swap entry will
> + * all bits clear.
> + */
> + max_swap_offset = swp_offset(pmd_to_swp_entry(swp_entry_to_pmd(swp_entry(0, ~0UL))));
> + swp_set = swp_entry((1 << MAX_SWAPFILES_SHIFT) - 1, max_swap_offset);
> + swp_clear = swp_entry(0, 0);
> +
> + /* Convert to pmd. */
> + pmd_set = swp_entry_to_pmd(swp_set);
> + pmd_clear = swp_entry_to_pmd(swp_clear);
[ 0.702163] debug_vm_pgtable: [swp_pmd_mkinvalid_tests ]: valid: pmd_set=f800000000000000, pmd_clear=7fffffffffffe00
> +
> + /*
> + * Sanity check that the pmds are not-present, not-huge and swap entry
> + * is recoverable without corruption.
> + */
> + WARN_ON(pmd_present(pmd_set));
> + WARN_ON(pmd_trans_huge(pmd_set));
> + swp_convert = pmd_to_swp_entry(pmd_set);
> + WARN_ON(swp_type(swp_set) != swp_type(swp_convert));
> + WARN_ON(swp_offset(swp_set) != swp_offset(swp_convert));
> + WARN_ON(pmd_present(pmd_clear));
> + WARN_ON(pmd_trans_huge(pmd_clear));
> + swp_convert = pmd_to_swp_entry(pmd_clear);
> + WARN_ON(swp_type(swp_clear) != swp_type(swp_convert));
> + WARN_ON(swp_offset(swp_clear) != swp_offset(swp_convert));
> +
> + /* Now invalidate the pmd. */
> + pmd_set = pmd_mkinvalid(pmd_set);
> + pmd_clear = pmd_mkinvalid(pmd_clear);
[ 0.704452] debug_vm_pgtable: [swp_pmd_mkinvalid_tests ]: invalid: pmd_set=f800000000000000, pmd_clear=7ffffffffe00e00
> +
> + /*
> + * Since its a swap pmd, invalidation should effectively be a noop and
> + * the checks we already did should give the same answer. Check the
> + * invalidation didn't corrupt any fields.
> + */
> + WARN_ON(pmd_present(pmd_set));
> + WARN_ON(pmd_trans_huge(pmd_set));
> + swp_convert = pmd_to_swp_entry(pmd_set);
[ 0.706461] debug_vm_pgtable: [swp_pmd_mkinvalid_tests ]: set: swp=7c03ffffffffffff (1f, 3ffffffffffff), convert=7c03ffffffffffff (1f, 3ffffffffffff)
> + WARN_ON(swp_type(swp_set) != swp_type(swp_convert));
> + WARN_ON(swp_offset(swp_set) != swp_offset(swp_convert));
> + WARN_ON(pmd_present(pmd_clear));
> + WARN_ON(pmd_trans_huge(pmd_clear));
> + swp_convert = pmd_to_swp_entry(pmd_clear);
[ 0.708841] debug_vm_pgtable: [swp_pmd_mkinvalid_tests ]: clear: swp=0 (0, 0), convert=ff8 (0, ff8)
> + WARN_ON(swp_type(swp_clear) != swp_type(swp_convert));
> + WARN_ON(swp_offset(swp_clear) != swp_offset(swp_convert));
This line fails on x86_64.
The logs show that the offset is indeed being corrupted by pmd_mkinvalid(); 0 -> 0xff8.
I think this is due to x86's pmd_mkinvalid() assuming the pmd is present; pmd_flags() and pmd_pfn() do all sorts of weird and wonderful things.
So does this take us full circle? Are we now back to modifying the core-mm to never call pmd_mkinvalid() on a non-present entry? If so, then I guess we should remove the arm64 fix from for-next/fixes.
> +}
> +#else
> +static void __init swp_pmd_mkinvalid_tests(struct pgtable_debug_args *args) { }
> +#endif /* !__HAVE_ARCH_PMDP_INVALIDATE && CONFIG_ARCH_ENABLE_THP_MIGRATION */
> +
> static void __init pmd_thp_tests(struct pgtable_debug_args *args)
> {
> pmd_t pmd;
> @@ -982,6 +1041,8 @@ static void __init pmd_thp_tests(struct pgtable_debug_args *args)
> WARN_ON(!pmd_trans_huge(pmd_mkinvalid(pmd_mkhuge(pmd))));
> WARN_ON(!pmd_present(pmd_mkinvalid(pmd_mkhuge(pmd))));
> #endif /* __HAVE_ARCH_PMDP_INVALIDATE */
> +
> + swp_pmd_mkinvalid_tests(args);
> }
>
> #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
> --
> 2.25.1
>
