On ARM32 with HIGHMEM/HIGHPTE, break_ksm_pmd_entry() triggers a BUG
during KSM unmerging because pte_unmap_unlock() is passed a pointer
that may be beyond the mapped PTE page.

The issue occurs when the PTE iteration loop completes without finding
a KSM page. After the loop, 'ptep' has been incremented past the last
PTE entry. On ARM32 LPAE with 512 PTEs per page (512 * 8 = 4096 bytes),
this means ptep points to the next page, outside the kmap'd region.

When pte_unmap_unlock(ptep, ptl) calls kunmap_local(ptep), it unmaps
the wrong page address, leaving the original kmap slot still mapped.
The next kmap_local then finds this slot unexpectedly occupied:

  WARNING: mm/highmem.c:622 kunmap_local_indexed  (address mismatch)
  kernel BUG at mm/highmem.c:564  __kmap_local_pfn_prot  (slot not empty)

Fix this by passing start_ptep to pte_unmap_unlock(), which always
points within the originally mapped PTE page.

Reproducer: Run LTP ksm03 test on ARM32 with HIGHMEM enabled. The test
triggers KSM merging followed by unmerging (writing 0 then 2 to
/sys/kernel/mm/ksm/run), which exercises break_ksm_pmd_entry().

Fixes: 5d4939fc2258 ("ksm: perform a range-walk in break_ksm")
Assisted-by: claude-opus-4-5-20251101
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 mm/ksm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/ksm.c b/mm/ksm.c
index cfc182255c7b..2d89a7c8b4eb 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -650,7 +650,7 @@ static int break_ksm_pmd_entry(pmd_t *pmdp, unsigned long addr, unsigned long en
 		}
 	}
 out_unlock:
-	pte_unmap_unlock(ptep, ptl);
+	pte_unmap_unlock(start_ptep, ptl);
 	return found;
 }
 
-- 
2.51.0

On 2025/12/21 04:29, Sasha Levin wrote:
> On ARM32 with HIGHMEM/HIGHPTE, break_ksm_pmd_entry() triggers a BUG
> during KSM unmerging because pte_unmap_unlock() is passed a pointer
> that may be beyond the mapped PTE page.
> 
> The issue occurs when the PTE iteration loop completes without finding
> a KSM page. After the loop, 'ptep' has been incremented past the last
> PTE entry. On ARM32 LPAE with 512 PTEs per page (512 * 8 = 4096 bytes),
> this means ptep points to the next page, outside the kmap'd region.
> 
> When pte_unmap_unlock(ptep, ptl) calls kunmap_local(ptep), it unmaps
> the wrong page address, leaving the original kmap slot still mapped.
> The next kmap_local then finds this slot unexpectedly occupied:
> 
>    WARNING: mm/highmem.c:622 kunmap_local_indexed  (address mismatch)
>    kernel BUG at mm/highmem.c:564  __kmap_local_pfn_prot  (slot not empty)
> 
> Fix this by passing start_ptep to pte_unmap_unlock(), which always
> points within the originally mapped PTE page.
> 
> Reproducer: Run LTP ksm03 test on ARM32 with HIGHMEM enabled. The test
> triggers KSM merging followed by unmerging (writing 0 then 2 to
> /sys/kernel/mm/ksm/run), which exercises break_ksm_pmd_entry().
> 
> Fixes: 5d4939fc2258 ("ksm: perform a range-walk in break_ksm")
> Assisted-by: claude-opus-4-5-20251101
> Signed-off-by: Sasha Levin <sashal@kernel.org>

Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>

Thanks.

> ---
>   mm/ksm.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/ksm.c b/mm/ksm.c
> index cfc182255c7b..2d89a7c8b4eb 100644
> --- a/mm/ksm.c
> +++ b/mm/ksm.c
> @@ -650,7 +650,7 @@ static int break_ksm_pmd_entry(pmd_t *pmdp, unsigned long addr, unsigned long en
>   		}
>   	}
>   out_unlock:
> -	pte_unmap_unlock(ptep, ptl);
> +	pte_unmap_unlock(start_ptep, ptl);
>   	return found;
>   }
>

On 12/20/25 21:29, Sasha Levin wrote:
> On ARM32 with HIGHMEM/HIGHPTE, break_ksm_pmd_entry() triggers a BUG
> during KSM unmerging because pte_unmap_unlock() is passed a pointer
> that may be beyond the mapped PTE page.
> 
> The issue occurs when the PTE iteration loop completes without finding
> a KSM page. After the loop, 'ptep' has been incremented past the last
> PTE entry. On ARM32 LPAE with 512 PTEs per page (512 * 8 = 4096 bytes),
> this means ptep points to the next page, outside the kmap'd region.
> 
> When pte_unmap_unlock(ptep, ptl) calls kunmap_local(ptep), it unmaps
> the wrong page address, leaving the original kmap slot still mapped.
> The next kmap_local then finds this slot unexpectedly occupied:
> 
>    WARNING: mm/highmem.c:622 kunmap_local_indexed  (address mismatch)
>    kernel BUG at mm/highmem.c:564  __kmap_local_pfn_prot  (slot not empty)
> 
> Fix this by passing start_ptep to pte_unmap_unlock(), which always
> points within the originally mapped PTE page.
> 
> Reproducer: Run LTP ksm03 test on ARM32 with HIGHMEM enabled. The test
> triggers KSM merging followed by unmerging (writing 0 then 2 to
> /sys/kernel/mm/ksm/run), which exercises break_ksm_pmd_entry().
> 
> Fixes: 5d4939fc2258 ("ksm: perform a range-walk in break_ksm")
> Assisted-by: claude-opus-4-5-20251101
> Signed-off-by: Sasha Levin <sashal@kernel.org>
> ---
>   mm/ksm.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/ksm.c b/mm/ksm.c
> index cfc182255c7b..2d89a7c8b4eb 100644
> --- a/mm/ksm.c
> +++ b/mm/ksm.c
> @@ -650,7 +650,7 @@ static int break_ksm_pmd_entry(pmd_t *pmdp, unsigned long addr, unsigned long en
>   		}
>   	}
>   out_unlock:
> -	pte_unmap_unlock(ptep, ptl);
> +	pte_unmap_unlock(start_ptep, ptl);
>   	return found;
>   }
>   

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

-- 
Cheers

David