Re: [PATCH] /dev/zero: try to align PMD_SIZE for private mapping

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Tue, Jul 29, 2025 at 09:49:41PM +0800, Zhang Qilong wrote:
> By default, THP are usually enabled. Mapping /dev/zero with a size

Err... we can't rely on this.

As per below comments on code, I'd update this to say something about fallback
if it's not.

> larger than 2MB could achieve performance gains by allocating aligned
> address. The mprot_tw4m in libMicro average execution time on arm64:
>   - Test case:        mprot_tw4m
>   - Before the patch:   22 us
>   - After the patch:    17 us
>
> Signed-off-by: Zhang Qilong <zhangqilong3@huawei.com>

This looks ok to me because there's a precedent for using
thp_get_unmapped_area() directly as a file_operations->get_unmapped_area e.g. in
ext4.

We also simply (amusingly, or perhaps not hugely amusingly, rather 'uniquely')
establish an anonymous mapping on f_op->mmap via mmap_zero() using
vma_set_anonymous(), so we can rely on the standard anon page memory faulting
logic to sort out the actual allocation/mapping of the huge page via:

__handle_mm_fault() -> create_huge_pmd() -> do_huge_pmd_anonymous_page() etc.

So everything should 'just work', and fallback if not permitted.

So in general seems fine.

> ---
>  drivers/char/mem.c | 5 +++++
>  1 file changed, 5 insertions(+)
>
> diff --git a/drivers/char/mem.c b/drivers/char/mem.c
> index 48839958b0b1..c57327ca9dd6 100644
> --- a/drivers/char/mem.c
> +++ b/drivers/char/mem.c
> @@ -515,10 +515,12 @@ static int mmap_zero(struct file *file, struct vm_area_struct *vma)
>  static unsigned long get_unmapped_area_zero(struct file *file,
>  				unsigned long addr, unsigned long len,
>  				unsigned long pgoff, unsigned long flags)
>  {
>  #ifdef CONFIG_MMU
> +	unsigned long ret;
> +
>  	if (flags & MAP_SHARED) {
>  		/*
>  		 * mmap_zero() will call shmem_zero_setup() to create a file,
>  		 * so use shmem's get_unmapped_area in case it can be huge;
>  		 * and pass NULL for file as in mmap.c's get_unmapped_area(),
> @@ -526,10 +528,13 @@ static unsigned long get_unmapped_area_zero(struct file *file,
>  		 */
>  		return shmem_get_unmapped_area(NULL, addr, len, pgoff, flags);
>  	}
>
>  	/* Otherwise flags & MAP_PRIVATE: with no shmem object beneath it */

Let's add a comment here like:

	/*
	 * Attempt to map aligned to huge page size if possible, otherwise we
         * fall back to system page size mappings. If THP is not enabled, this
         * returns NULL and we always fallback.
	 */

I think it'd be sensible to have an #ifdef CONFIG_TRANSPARENT_HUGEPAGE here,
because thp_get_unmapped_area() does the fallback for you, and then otherwise
we'd be trying it twice which is weird.

E.g.:

#ifdef CONFIG_TRANSPARENT_HUGEPAGE
	return thp_get_unmapped_area(file, addr, len, pgoff, flags);
#else
	return mm_get_unmapped_area(current->mm, file, addr, len, pgoff, flags);
#endif

> +	ret = thp_get_unmapped_area(file, addr, len, pgoff, flags);
> +	if (ret)
> +		return ret;
>  	return mm_get_unmapped_area(current->mm, file, addr, len, pgoff, flags);
>  #else
>  	return -ENOSYS;
>  #endif
>  }
> --
> 2.43.0
>

In _theory_ we should do the thing in mmap() where we check the size is
PMD-aligned (see __get_unmapped_area()), but I don't think anybody's mapping a
bunch of /dev/zero mappings next to each other or using them in any way where
that'd matter... So yeah let's not :)