[PATCH v3] mm/hugetlb: preserve mremap address delta when skipping page tables

[PATCH v3] mm/hugetlb: preserve mremap address delta when skipping page tables

[Jaewook You]

move_hugetlb_page_tables() optimizes mremap() by advancing to the last
entry in the page table when the source page table does not exist, either
initially or after unsharing a PMD table.  The common loop increment then
steps to the first entry in the next page table.

However, the code advances both the source and destination addresses to
the last entries in their respective page tables, which is wrong.  The
destination address must be advanced only by the same amount as the source
address.

If the source and destination offsets within their page tables differ, the
destination address can be advanced too far, causing follow-up issues.  Fix
this by advancing the destination address by the source advance distance.

With a reproducer, we were able to trigger a kernel panic on x86-64.  With
this fix in place, we can no longer reproduce the issue.

Fixes: e95a9851787b ("hugetlb: skip to end of PT page mapping when pte not present")
Fixes: 4ddb4d91b82f ("hugetlb: do not update address in huge_pmd_unshare")
Cc: stable@vger.kernel.org
Assisted-by: LLM
Signed-off-by: Jaewook You <jaewook376@gmail.com>
---
Changes in v3:
- Clarify that the optimization advances to the last entry in the current
  page table before the common loop increment steps to the next entry.
- Rename remaining_size to offset_to_last_entry as suggested by David.

v2: https://lore.kernel.org/20260911182408.75821-1-jaewook376@gmail.com/

 mm/hugetlb.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 4f6f58bf3db6c..5749f6270fb1f 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -5161,18 +5161,21 @@ int move_hugetlb_page_tables(struct vm_area_struct *vma,
 	hugetlb_vma_lock_write(vma);
 	i_mmap_lock_write(mapping);
 	for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
+		const unsigned long offset_to_last_entry =
+			(old_addr | last_addr_mask) - old_addr;
+
 		src_pte = hugetlb_walk(vma, old_addr, sz);
 		if (!src_pte) {
-			old_addr |= last_addr_mask;
-			new_addr |= last_addr_mask;
+			old_addr += offset_to_last_entry;
+			new_addr += offset_to_last_entry;
 			continue;
 		}
 		if (huge_pte_none(huge_ptep_get(mm, old_addr, src_pte)))
 			continue;
 
 		if (huge_pmd_unshare(&tlb, vma, old_addr, src_pte)) {
-			old_addr |= last_addr_mask;
-			new_addr |= last_addr_mask;
+			old_addr += offset_to_last_entry;
+			new_addr += offset_to_last_entry;
 			continue;
 		}
 

base-commit: 08df884136f1c1197bab2a27814404fd329d9aac
-- 
2.43.0

Re: [PATCH v3] mm/hugetlb: preserve mremap address delta when skipping page tables

[David Hildenbrand (Arm)]

On 9/14/26 15:23, Jaewook You wrote:
> move_hugetlb_page_tables() optimizes mremap() by advancing to the last
> entry in the page table when the source page table does not exist, either
> initially or after unsharing a PMD table.  The common loop increment then
> steps to the first entry in the next page table.
> 
> However, the code advances both the source and destination addresses to
> the last entries in their respective page tables, which is wrong.  The
> destination address must be advanced only by the same amount as the source
> address.
> 
> If the source and destination offsets within their page tables differ, the
> destination address can be advanced too far, causing follow-up issues.  Fix
> this by advancing the destination address by the source advance distance.
> 
> With a reproducer, we were able to trigger a kernel panic on x86-64.  With
> this fix in place, we can no longer reproduce the issue.
> 
> Fixes: e95a9851787b ("hugetlb: skip to end of PT page mapping when pte not present")
> Fixes: 4ddb4d91b82f ("hugetlb: do not update address in huge_pmd_unshare")
> Cc: stable@vger.kernel.org
> Assisted-by: LLM
> Signed-off-by: Jaewook You <jaewook376@gmail.com>
> ---
> Changes in v3:
> - Clarify that the optimization advances to the last entry in the current
>   page table before the common loop increment steps to the next entry.
> - Rename remaining_size to offset_to_last_entry as suggested by David.
> 
> v2: https://lore.kernel.org/20260911182408.75821-1-jaewook376@gmail.com/
> 

v2 does not exist publicly ;)

This would have been better send as a v1 upstream (I mentioned this in private
but you might have missed it).

>  mm/hugetlb.c | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 4f6f58bf3db6c..5749f6270fb1f 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -5161,18 +5161,21 @@ int move_hugetlb_page_tables(struct vm_area_struct *vma,
>  	hugetlb_vma_lock_write(vma);
>  	i_mmap_lock_write(mapping);
>  	for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
> +		const unsigned long offset_to_last_entry =
> +			(old_addr | last_addr_mask) - old_addr;
> +
>  		src_pte = hugetlb_walk(vma, old_addr, sz);
>  		if (!src_pte) {
> -			old_addr |= last_addr_mask;
> -			new_addr |= last_addr_mask;
> +			old_addr += offset_to_last_entry;
> +			new_addr += offset_to_last_entry;
>  			continue;
>  		}
>  		if (huge_pte_none(huge_ptep_get(mm, old_addr, src_pte)))
>  			continue;
>  
>  		if (huge_pmd_unshare(&tlb, vma, old_addr, src_pte)) {
> -			old_addr |= last_addr_mask;
> -			new_addr |= last_addr_mask;
> +			old_addr += offset_to_last_entry;
> +			new_addr += offset_to_last_entry;
>  			continue;

As Oscar expressed, he might prefer some comment about the situation. But maybe
that can be deferred to some proper cleanups here.

Acked-by: David Hildenbrand (Arm) <david@kernel.org>

-- 
Cheers,

David