[PATCH 1/1] mm/thp: fix MTE tag mismatch when replacing zero-filled subpages

[Catalin Marinas]

On Mon, Sep 22, 2025 at 10:14:58AM +0800, Lance Yang wrote:
> From: Lance Yang <lance.yang at linux.dev>
> 
> When both THP and MTE are enabled, splitting a THP and replacing its
> zero-filled subpages with the shared zeropage can cause MTE tag mismatch
> faults in userspace.
> 
> Remapping zero-filled subpages to the shared zeropage is unsafe, as the
> zeropage has a fixed tag of zero, which may not match the tag expected by
> the userspace pointer.
> 
> KSM already avoids this problem by using memcmp_pages(), which on arm64
> intentionally reports MTE-tagged pages as non-identical to prevent unsafe
> merging.
> 
> As suggested by David[1], this patch adopts the same pattern, replacing the
> memchr_inv() byte-level check with a call to pages_identical(). This
> leverages existing architecture-specific logic to determine if a page is
> truly identical to the shared zeropage.
> 
> Having both the THP shrinker and KSM rely on pages_identical() makes the
> design more future-proof, IMO. Instead of handling quirks in generic code,
> we just let the architecture decide what makes two pages identical.
> 
> [1] https://lore.kernel.org/all/ca2106a3-4bb2-4457-81af-301fd99fbef4@redhat.com
> 
> Cc: <stable at vger.kernel.org>
> Reported-by: Qun-wei Lin <Qun-wei.Lin at mediatek.com>
> Closes: https://lore.kernel.org/all/a7944523fcc3634607691c35311a5d59d1a3f8d4.camel@mediatek.com
> Fixes: b1f202060afe ("mm: remap unused subpages to shared zeropage when splitting isolated thp")
> Suggested-by: David Hildenbrand <david at redhat.com>
> Signed-off-by: Lance Yang <lance.yang at linux.dev>

Functionally, the patch looks fine, both with and without MTE.

Reviewed-by: Catalin Marinas <catalin.marinas at arm.com>

> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index 32e0ec2dde36..28d4b02a1aa5 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -4104,29 +4104,20 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
>  static bool thp_underused(struct folio *folio)
>  {
>  	int num_zero_pages = 0, num_filled_pages = 0;
> -	void *kaddr;
>  	int i;
>  
>  	for (i = 0; i < folio_nr_pages(folio); i++) {
> -		kaddr = kmap_local_folio(folio, i * PAGE_SIZE);
> -		if (!memchr_inv(kaddr, 0, PAGE_SIZE)) {
> -			num_zero_pages++;
> -			if (num_zero_pages > khugepaged_max_ptes_none) {
> -				kunmap_local(kaddr);
> +		if (pages_identical(folio_page(folio, i), ZERO_PAGE(0))) {
> +			if (++num_zero_pages > khugepaged_max_ptes_none)
>  				return true;

I wonder what the overhead of doing a memcmp() vs memchr_inv() is. The
former will need to read from two places. If it's noticeable, it would
affect architectures that don't have an MTE equivalent.

Alternatively we could introduce something like folio_has_metadata()
which on arm64 simply checks PG_mte_tagged.

-- 
Catalin