Re: [PATCH for-mm 1/2] mm/internal: Implement no-op mlock_page_drain() for !CONFIG_MMU

[Hugh Dickins]

On Wed, 9 Feb 2022, SeongJae Park wrote:
> Commit 4b3b8bd6c8287 ("mm/munlock: mlock_page() munlock_page() batch by
> pagevec") in -mm tree[1] implements 'mlock_page_drain()' under
> CONFIG_MMU only, but the function is used by 'lru_add_drain_cpu()',
> which defined outside of CONFIG_MMU.  As a result, below build error
> occurs.
> 
>     /linux/mm/swap.c: In function 'lru_add_drain_cpu':
>     /linux/mm/swap.c:637:2: error: implicit declaration of function 'mlock_page_drain' [-Werror=implicit-function-declaration]
>       637 |  mlock_page_drain(cpu);
>           |  ^~~~~~~~~~~~~~~~
>     cc1: some warnings being treated as errors
>     /linux/scripts/Makefile.build:289: recipe for target 'mm/swap.o' failed
> 
> This commit fixes it by implementing no-op 'mlock_page_drain()' for
> !CONFIG_MMU case, similar to 'mlock_new_page()'.
> 
> [1] https://www.ozlabs.org/~akpm/mmotm/broken-out/mm-munlock-mlock_page-munlock_page-batch-by-pagevec.patch
> 
> Signed-off-by: SeongJae Park <sj@xxxxxxxxxx>
> ---
>  mm/internal.h | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/mm/internal.h b/mm/internal.h
> index 0d240e876831..248224369b34 100644
> --- a/mm/internal.h
> +++ b/mm/internal.h
> @@ -508,6 +508,7 @@ static inline void mlock_vma_page(struct page *page,
>  static inline void munlock_vma_page(struct page *page,
>  			struct vm_area_struct *vma, bool compound) { }
>  static inline void mlock_new_page(struct page *page) { }
> +static inline void mlock_page_drain(int cpu) { }
>  static inline void vunmap_range_noflush(unsigned long start, unsigned long end)
>  {
>  }

Thank you, SeongJae, and thank you Geert for reporting.

This patch is good as far as it goes, but Andrew, please don't add
it right now: I need to think a bit more, and will send another
(or Ack SeongJae's) later in the day.

The thing is, SeongJae's patch makes me wonder, why did it not need a
!CONFIG_MMU definition for need_mlock_page_drain() too?  That's because
mm/swap.c's call to it is under an #ifdef CONFIG_SMP, and I imagine that
CONFIG_MMU=n usually goes along with (but does not necessarily imply?)
CONFIG_SMP=n.  It'll be safer to add a need_mlock_page_drain() stub too.

But more seriously, is the mlock_page_drain() going to be called when
it's wanted, when CONFIG_SMP is not set?  I had forgotten how mm/swap.c
goes in different directions (for some things but not for others)
according to CONFIG_SMP.  I'll look into it and come back later.

Thanks,
Hugh