Re: [PATCH v2] arm64: Restrict pagetable teardown to avoid false warning

[Ryan Roberts <ryan.roberts@arm.com>]

On 18/05/2025 10:54, Dev Jain wrote:
> Commit 9c006972c3fe removes the pxd_present() checks because the caller

nit: please use the standard format for describing commits: Commit 9c006972c3fe
("arm64: mmu: drop pXd_present() checks from pXd_free_pYd_table()")

> checks pxd_present(). But, in case of vmap_try_huge_pud(), the caller only
> checks pud_present(); pud_free_pmd_page() recurses on each pmd through
> pmd_free_pte_page(), wherein the pmd may be none. Thus it is possible to
> hit a warning in the latter, since pmd_none => !pmd_table(). Thus, add
> a pmd_present() check in pud_free_pmd_page().
> 
> This problem was found by code inspection.
> 
> This patch is based on 6.15-rc6.

nit: please remove this to below the "---", its not part of the commit log.

> 
> Fixes: 9c006972c3fe (arm64: mmu: drop pXd_present() checks from pXd_free_pYd_table())
> 

nit: remove empty line; the tags should all be in a single block with no empty
lines.

> Cc: <stable@vger.kernel.org>
> Reported-by: Ryan Roberts <ryan.roberts@arm.com> 
> Signed-off-by: Dev Jain <dev.jain@arm.com>
> ---
> v1->v2:
>  - Enforce check in caller
> 
>  arch/arm64/mm/mmu.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index ea6695d53fb9..5b1f4cd238ca 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -1286,7 +1286,8 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
>  	next = addr;
>  	end = addr + PUD_SIZE;
>  	do {
> -		pmd_free_pte_page(pmdp, next);
> +		if (pmd_present(*pmdp))

pmd_free_pte_page() is using READ_ONCE() to access the *pmdp to ensure it can't
be torn. I suspect we don't technically need that in these functions because
there can be no race with a writer. But the arm64 arch code always uses
READ_ONCE() for dereferencing pgtable entries for safely. Perhaps we should be
consistent here?

		if (pmd_present(READ_ONCE(*pmdp)))

> +			pmd_free_pte_page(pmdp, next);
>  	} while (pmdp++, next += PMD_SIZE, next != end);
>  
>  	pud_clear(pudp);