Re: [PATCH] vduse: Use fixed 4KB bounce pages for arm64 64KB page size

[Jason Wang]

On Mon, Sep 15, 2025 at 3:34 PM <sheng.zhao@xxxxxxxxxxxxx> wrote:
>
> From: Sheng Zhao <sheng.zhao@xxxxxxxxxxxxx>
>
> The allocation granularity of bounce pages is PAGE_SIZE. This may cause
> even small IO requests to occupy an entire bounce page exclusively. The
> kind of memory waste will be more significant on arm64 with 64KB pages.

Let's tweak the title as there are archs that are using non 4KB pages
other than arm.

>
> So, optimize it by using fixed 4KB bounce pages.
>
> Signed-off-by: Sheng Zhao <sheng.zhao@xxxxxxxxxxxxx>
> ---
>  drivers/vdpa/vdpa_user/iova_domain.c | 120 +++++++++++++++++----------
>  drivers/vdpa/vdpa_user/iova_domain.h |   5 ++
>  2 files changed, 83 insertions(+), 42 deletions(-)
>
> diff --git a/drivers/vdpa/vdpa_user/iova_domain.c b/drivers/vdpa/vdpa_user/iova_domain.c
> index 58116f89d8da..768313c80b62 100644
> --- a/drivers/vdpa/vdpa_user/iova_domain.c
> +++ b/drivers/vdpa/vdpa_user/iova_domain.c
> @@ -103,19 +103,26 @@ void vduse_domain_clear_map(struct vduse_iova_domain *domain,
>  static int vduse_domain_map_bounce_page(struct vduse_iova_domain *domain,
>                                          u64 iova, u64 size, u64 paddr)
>  {
> -       struct vduse_bounce_map *map;
> +       struct vduse_bounce_map *map, *head_map;
> +       struct page *tmp_page;
>         u64 last = iova + size - 1;
>
>         while (iova <= last) {
> -               map = &domain->bounce_maps[iova >> PAGE_SHIFT];
> +               map = &domain->bounce_maps[iova >> BOUNCE_PAGE_SHIFT];

BOUNCE_PAGE_SIZE is kind of confusing as it's not the size of any page
at all when PAGE_SIZE is not 4K.

>                 if (!map->bounce_page) {
> -                       map->bounce_page = alloc_page(GFP_ATOMIC);
> -                       if (!map->bounce_page)
> -                               return -ENOMEM;
> +                       head_map = &domain->bounce_maps[(iova & PAGE_MASK) >> BOUNCE_PAGE_SHIFT];
> +                       if (!head_map->bounce_page) {
> +                               tmp_page = alloc_page(GFP_ATOMIC);
> +                               if (!tmp_page)
> +                                       return -ENOMEM;
> +                               if (cmpxchg(&head_map->bounce_page, NULL, tmp_page))
> +                                       __free_page(tmp_page);

I don't understand why we need cmpxchg() logic.

Btw, it looks like you want to make multiple bounce_map to point to
the same 64KB page? I wonder what's the advantages of doing this. Can
we simply keep the 64KB page in bounce_map?

Thanks