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

[Catalin Marinas <catalin.marinas@arm.com>]

On Mon, Sep 22, 2025 at 07:59:00PM +0200, David Hildenbrand wrote:
> On 22.09.25 19:24, Catalin Marinas wrote:
> > On Mon, Sep 22, 2025 at 10:14:58AM +0800, Lance Yang wrote:
> > > 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.
> 
> We discussed something similar in the other thread (I suggested
> page_is_mergable()). I'd prefer to use pages_identical() for now, so we have
> the same logic here and in ksm code.
> 
> (this patch here almost looks like a cleanup :) )
> 
> If this becomes a problem, what we could do is in pages_identical() would be
> simply doing the memchr_inv() in case is_zero_pfn(). KSM might benefit from
> that as well when merging with the shared zeropage through
> try_to_merge_with_zero_page().

Yes, we can always optimise it later.

I just realised that on arm64 with MTE we won't get any merging with the
zero page even if the user page isn't mapped with PROT_MTE. In
cpu_enable_mte() we zero the tags in the zero page and set
PG_mte_tagged. The reason is that we want to use the zero page with
PROT_MTE mappings (until tag setting causes CoW). Hmm, the arm64
memcmp_pages() messed up KSM merging with the zero page even before this
patch.

The MTE tag setting evolved a bit over time with some locking using PG_*
flags to avoid a set_pte_at() race trying to initialise the tags on the
same page. We also moved the swap restoring to arch_swap_restore()
rather than the set_pte_at() path. So it is safe now to merge with the
zero page if the other page isn't tagged. A subsequent set_pte_at()
attempting to clear the tags would notice that the zero page is already
tagged.

We could go a step further and add tag comparison (I had some code
around) but I think the quick fix is to just not treat the zero page as
tagged. Not fully tested yet:

diff --git a/arch/arm64/kernel/mte.c b/arch/arm64/kernel/mte.c
index e5e773844889..72a1dfc54659 100644
--- a/arch/arm64/kernel/mte.c
+++ b/arch/arm64/kernel/mte.c
@@ -73,6 +73,8 @@ int memcmp_pages(struct page *page1, struct page *page2)
 {
 	char *addr1, *addr2;
 	int ret;
+	bool page1_tagged = page_mte_tagged(page1) && !is_zero_page(page1);
+	bool page2_tagged = page_mte_tagged(page2) && !is_zero_page(page2);
 
 	addr1 = page_address(page1);
 	addr2 = page_address(page2);
@@ -83,11 +85,10 @@ int memcmp_pages(struct page *page1, struct page *page2)
 
 	/*
 	 * If the page content is identical but at least one of the pages is
-	 * tagged, return non-zero to avoid KSM merging. If only one of the
-	 * pages is tagged, __set_ptes() may zero or change the tags of the
-	 * other page via mte_sync_tags().
+	 * tagged, return non-zero to avoid KSM merging. Ignore the zero page
+	 * since it is always tagged with the tags cleared.
 	 */
-	if (page_mte_tagged(page1) || page_mte_tagged(page2))
+	if (page1_tagged || page2_tagged)
 		return addr1 != addr2;
 
 	return ret;

-- 
Catalin