Re: [PATCH] arm64, mm: avoid always making PTE dirty in pte_mkwrite()

[Catalin Marinas]

On Wed, Oct 15, 2025 at 10:37:12AM +0800, Huang Ying wrote:
> Current pte_mkwrite_novma() makes PTE dirty unconditionally.  This may
> mark some pages that are never written dirty wrongly.  For example,
> do_swap_page() may map the exclusive pages with writable and clean PTEs
> if the VMA is writable and the page fault is for read access.
> However, current pte_mkwrite_novma() implementation always dirties the
> PTE.  This may cause unnecessary disk writing if the pages are
> never written before being reclaimed.
> 
> So, change pte_mkwrite_novma() to clear the PTE_RDONLY bit only if the
> PTE_DIRTY bit is set to make it possible to make the PTE writable and
> clean.
> 
> The current behavior was introduced in commit 73e86cb03cf2 ("arm64:
> Move PTE_RDONLY bit handling out of set_pte_at()").  Before that,
> pte_mkwrite() only sets the PTE_WRITE bit, while set_pte_at() only
> clears the PTE_RDONLY bit if both the PTE_WRITE and the PTE_DIRTY bits
> are set.
> 
> To test the performance impact of the patch, on an arm64 server
> machine, run 16 redis-server processes on socket 1 and 16
> memtier_benchmark processes on socket 0 with mostly get
> transactions (that is, redis-server will mostly read memory only).
> The memory footprint of redis-server is larger than the available
> memory, so swap out/in will be triggered.  Test results show that the
> patch can avoid most swapping out because the pages are mostly clean.
> And the benchmark throughput improves ~23.9% in the test.
> 
> Fixes: 73e86cb03cf2 ("arm64: Move PTE_RDONLY bit handling out of set_pte_at()")
> Signed-off-by: Huang Ying <ying.huang@xxxxxxxxxxxxxxxxx>
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> Cc: Anshuman Khandual <anshuman.khandual@xxxxxxx>
> Cc: Ryan Roberts <ryan.roberts@xxxxxxx>
> Cc: Gavin Shan <gshan@xxxxxxxxxx>
> Cc: Ard Biesheuvel <ardb@xxxxxxxxxx>
> Cc: "Matthew Wilcox (Oracle)" <willy@xxxxxxxxxxxxx>
> Cc: Yicong Yang <yangyicong@xxxxxxxxxxxxx>
> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> ---
>  arch/arm64/include/asm/pgtable.h | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> index aa89c2e67ebc..0944e296dd4a 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -293,7 +293,8 @@ static inline pmd_t set_pmd_bit(pmd_t pmd, pgprot_t prot)
>  static inline pte_t pte_mkwrite_novma(pte_t pte)
>  {
>  	pte = set_pte_bit(pte, __pgprot(PTE_WRITE));
> -	pte = clear_pte_bit(pte, __pgprot(PTE_RDONLY));
> +	if (pte_sw_dirty(pte))
> +		pte = clear_pte_bit(pte, __pgprot(PTE_RDONLY));
>  	return pte;
>  }

This seems to be the right thing. I recall years ago I grep'ed
(obviously not hard enough) and most pte_mkwrite() places had a
pte_mkdirty(). But I missed do_swap_page() and possibly others.

For this patch:

Reviewed-by: Catalin Marinas <catalin.marinas@xxxxxxx>

I wonder whether we should also add (as a separate patch):

diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
index 830107b6dd08..df1c552ef11c 100644
--- a/mm/debug_vm_pgtable.c
+++ b/mm/debug_vm_pgtable.c
@@ -101,6 +101,7 @@ static void __init pte_basic_tests(struct pgtable_debug_args *args, int idx)
 	WARN_ON(pte_dirty(pte_mkclean(pte_mkdirty(pte))));
 	WARN_ON(pte_write(pte_wrprotect(pte_mkwrite(pte, args->vma))));
 	WARN_ON(pte_dirty(pte_wrprotect(pte_mkclean(pte))));
+	WARN_ON(pte_dirty(pte_mkwrite_novma(pte_mkclean(pte))));
 	WARN_ON(!pte_dirty(pte_wrprotect(pte_mkdirty(pte))));
 }

For completeness, also (and maybe other combinations):

	WARN_ON(!pte_write(pte_mkdirty(pte_mkwrite_novma(pte))));

I cc'ed linux-mm in case we missed anything. If nothing raised, I'll
queue it next week.

Thanks.

-- 
Catalin