On Tue, Jun 14, 2022 at 4:46 AM Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
>
> Below is v2 of the patch. I was just going to change the type of start/end
> hugetlbfs_zero_partial_page arguments. However, Matthew also convinced me
> to do the simple folio conversion. Tested with new libhugetlbfs code.
>
>
> From 36a18e0b07c2e189092cc2d516e8cfedcb57d191 Mon Sep 17 00:00:00 2001
> From: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> Date: Mon, 13 Jun 2022 13:36:48 -0700
> Subject: [Patch v2] hugetlbfs: zero partial pages during fallocate hole punch
>
> hugetlbfs fallocate support was originally added with commit 70c3547e36f5
> ("hugetlbfs: add hugetlbfs_fallocate()"). Initial support only operated
> on whole hugetlb pages. This makes sense for populating files as other
> interfaces such as mmap and truncate require hugetlb page size alignment.
> Only operating on whole hugetlb pages for the hole punch case was a
> simplification and there was no compelling use case to zero partial pages.
>
> In a recent discussion[1] it was assumed that hugetlbfs hole punch would
> zero partial hugetlb pages as that is in line with the man page
> description saying 'partial filesystem blocks are zeroed'. However,
> the hugetlbfs hole punch code actually does this:
>
> hole_start = round_up(offset, hpage_size);
> hole_end = round_down(offset + len, hpage_size);
>
> Modify code to zero partial hugetlb pages in hole punch range. It is
> possible that application code could note a change in behavior. However,
> that would imply the code is passing in an unaligned range and expecting
> only whole pages be removed. This is unlikely as the fallocate
> documentation states the opposite.
>
> The current hugetlbfs fallocate hole punch behavior is tested with the
> libhugetlbfs test fallocate_align[2]. This test will be updated to
> validate partial page zeroing.
>
> [1] https://lore.kernel.org/linux-mm/20571829-9d3d-0b48-817c-b6b15565f651@xxxxxxxxxx/
> [2] https://github.com/libhugetlbfs/libhugetlbfs/blob/master/tests/fallocate_align.c
>
> Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
LGTM.
Reviewed-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
Thanks.
