mm/hugetlb: fix uffd-wp bit lost when unsharing happens

When we try to unshare a pinned page for a private hugetlb, uffd-wp bit
can get lost during unsharing.

When above condition met, one can lose uffd-wp bit on the privately mapped
hugetlb page.  It allows the page to be writable even if it should still be
wr-protected.  I assume it can mean data loss.

This should be very rare, only if an unsharing happened on a private
hugetlb page with uffd-wp protected (e.g.  in a child which shares the
same page with parent with UFFD_FEATURE_EVENT_FORK enabled).

When I wrote the reproducer (provided in the last patch) I needed to
use the newest gup_test cmd introduced by David to trigger it because I
don't even know another way to do a proper RO longerm pin.

Besides that, it needs a bunch of other conditions all met:

        (1) hugetlb being mapped privately,
        (2) userfaultfd registered with WP and EVENT_FORK,
        (3) the user app fork()s, then,
        (4) RO longterm pin onto a wr-protected anonymous page.

If it's not impossible to hit in production I'd say extremely rare.

Link: https://lkml.kernel.org/r/[email protected]
Fixes: 166f3ec ("mm/hugetlb: hook page faults for uffd write protection")
Signed-off-by: Peter Xu <[email protected]>
Reported-by: Mike Kravetz <[email protected]>
Reviewed-by: David Hildenbrand <[email protected]>
Reviewed-by: Mike Kravetz <[email protected]>
Cc: Andrea Arcangeli <[email protected]>
Cc: Axel Rasmussen <[email protected]>
Cc: Mika Penttilä <[email protected]>
Cc: Nadav Amit <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

mm/hugetlb.c

@@ -5644,13 +5644,16 @@ static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
 	spin_lock(ptl);
 	ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
 	if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
+		pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
+
 		/* Break COW or unshare */
 		huge_ptep_clear_flush(vma, haddr, ptep);
 		mmu_notifier_invalidate_range(mm, range.start, range.end);
 		page_remove_rmap(old_page, vma, true);
 		hugepage_add_new_anon_rmap(new_folio, vma, haddr);
-		set_huge_pte_at(mm, haddr, ptep,
-				make_huge_pte(vma, &new_folio->page, !unshare));
+		if (huge_pte_uffd_wp(pte))
+			newpte = huge_pte_mkuffd_wp(newpte);
+		set_huge_pte_at(mm, haddr, ptep, newpte);
 		folio_set_hugetlb_migratable(new_folio);
 		/* Make the old page be freed below */
 		new_folio = page_folio(old_page);