Re: [PATCH] tools/testing/vma: cover hole filling through __mmap_region()

[Lorenzo Stoakes (ARM)]

On Sun, Sep 06, 2026 at 10:41:00PM +0800, Tianyi Chen wrote:
> The mmap tests extend existing mappings one neighbor at a time, while
> merge tests construct merge state directly. Neither exercises filling
> a hole between compatible mappings through the mmap setup and completion
> path.
>
> Fill a gap through __mmap_region() and require both neighbors to merge
> into one VMA. Repeat with only the new mapping's execute permission set
> and require three separate VMAs. Check boundaries, permissions, page
> offsets, map_count and tree lookups across the mapped pages.
>
> The full VMA test suite passes all 28 tests with ASan and UBSan enabled.
>
> Assisted-by: LLM

Thanks for adding the tag, always much appreciated!

I may actually set my own LLM loose on these tests to expand some more. Is
a good area for such work I think.

Though it still needs massaging to get good code :)

> Signed-off-by: Tianyi Chen <hi@xxxxxxxxx>

General idea seems reasonable to me, and never any harm in adding more
tests :)

A bunch of stuff to fix below, but with those addressed patch should be
good.

Also please rebase this on the mm-unstable branch of Andrew's tree:

https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/

As there is a minor conflict with my upcoming work :)

> ---
>  tools/testing/vma/tests/mmap.c | 75 ++++++++++++++++++++++++++++++++++
>  1 file changed, 75 insertions(+)
>
> diff --git a/tools/testing/vma/tests/mmap.c b/tools/testing/vma/tests/mmap.c
> index c85bc000d1c..b66ec932a75 100644
> --- a/tools/testing/vma/tests/mmap.c
> +++ b/tools/testing/vma/tests/mmap.c
> @@ -45,7 +45,82 @@ static bool test_mmap_region_basic(void)
>  	return true;
>  }
>
> +static bool mmap_region_fill_hole(bool merge)
> +{
> +	const vma_flags_t vma_flags = mk_vma_flags(VMA_READ_BIT, VMA_WRITE_BIT,
> +			VMA_MAYREAD_BIT, VMA_MAYWRITE_BIT, VMA_MAYEXEC_BIT);
> +	vma_flags_t hole_flags = vma_flags;

Hole is the wrong word, maybe 'middle_flags'?

> +	struct mm_struct mm = {};
> +	struct vm_area_struct *vma;
> +	unsigned long addr;
> +	int count = 0;
> +	VMA_ITERATOR(vmi, &mm, 0);
> +
> +	current->mm = &mm;
> +	if (!merge)
> +		vma_flags_set(&hole_flags, VMA_EXEC_BIT);
> +
> +	/* Leave a hole between two otherwise mergeable mappings. */
> +	addr = __mmap_region(NULL, 0x300000, 0x3000, vma_flags, 0x300, NULL);
> +	ASSERT_EQ(addr, 0x300000);
> +	addr = __mmap_region(NULL, 0x306000, 0x3000, vma_flags, 0x306, NULL);
> +	ASSERT_EQ(addr, 0x306000);
> +	ASSERT_EQ(mm.map_count, 2);

Everything here is reasonable but can you please add comments like other tests
like:

	/* Map at 0x306000, length 0x3000. */

etc.

> +	vma_iter_set(&vmi, 0x303000);
> +	ASSERT_EQ(vma_iter_load(&vmi), NULL);
> +	vma_iter_set(&vmi, 0x305fff);
> +	ASSERT_EQ(vma_iter_load(&vmi), NULL);

Let's drop these 4 lines they're a bit useless I think.

> +
> +	/* A single flag difference must prevent merging with either neighbor. */

Similar to above re: comment.

Also worth saying

> +	addr = __mmap_region(NULL, 0x303000, 0x3000, hole_flags, 0x303, NULL);
> +	ASSERT_EQ(addr, 0x303000);
> +	ASSERT_EQ(mm.map_count, merge ? 1 : 3);
> +
> +	vma_iter_set(&vmi, 0);
> +	for_each_vma(vmi, vma) {
> +		unsigned long start = 0x300000 + count * 0x3000;
> +		unsigned long end = merge ? 0x309000 : start + 0x3000;

NIT: Can we make these const please?

> +		VMA_ITERATOR(lookup, &mm, start);
> +
> +		ASSERT_EQ(vma->vm_start, start);
> +		ASSERT_EQ(vma->vm_end, end);
> +		ASSERT_EQ(vma_start_pgoff(vma), start >> PAGE_SHIFT);
> +		ASSERT_EQ(vma_start_anon_pgoff(vma), start >> PAGE_SHIFT);

Newline here maybe as basic stuff above.

> +		ASSERT_TRUE(vma_test_all(vma, VMA_READ_BIT, VMA_WRITE_BIT,
> +					VMA_MAYREAD_BIT, VMA_MAYWRITE_BIT,
> +					VMA_MAYEXEC_BIT));
> +		ASSERT_EQ(vma_test(vma, VMA_EXEC_BIT), !merge && count == 1);

Also can we separate out the !merge && count bit? So put this at the start
of the for_each_vma() block:

	/* If testing the non-merge case, middle VMA will be set VMA_EXEC. */
	const bool is_middle_vma = count == 1;
	const expect_exec_vma = is_middle_vma && !merge;

Then here:

	ASSERT_EQ(vma_test(vma, VMA_EXEC_BIT), expect_exec_vma);

Also a newline here would be nice.

> +		for (addr = start; addr < end; addr += PAGE_SIZE) {
> +			vma_iter_set(&lookup, addr);
> +			ASSERT_EQ(vma_iter_load(&lookup), vma);
> +			vma_iter_set(&lookup, addr + PAGE_SIZE - 1);
> +			ASSERT_EQ(vma_iter_load(&lookup), vma);
> +		}

Let's drop this entire block please I don't think it's achieving anything
useful.

> +		count++;
> +	}

Newline here please.

> +	ASSERT_EQ(count, mm.map_count);
> +	vma_iter_set(&vmi, 0x2fffff);
> +	ASSERT_EQ(vma_iter_load(&vmi), NULL);
> +	vma_iter_set(&vmi, 0x309000);
> +	ASSERT_EQ(vma_iter_load(&vmi), NULL);

Again let's drop this block, it's not useful I don't think.

> +
> +	ASSERT_EQ(cleanup_mm(&mm, &vmi), count);
> +	return true;
> +}
> +
> +static bool test_mmap_region_fill_hole_merge(void)
> +{
> +	return mmap_region_fill_hole(true);
> +}
> +
> +static bool test_mmap_region_fill_hole_flags_mismatch(void)
> +{
> +	return mmap_region_fill_hole(false);
> +}
> +
>  static void run_mmap_tests(int *num_tests, int *num_fail)
>  {
>  	TEST(mmap_region_basic);
> +	TEST(mmap_region_fill_hole_merge);
> +	TEST(mmap_region_fill_hole_flags_mismatch);
>  }
> --
> 2.55.0
>

--
Cheers, Lorenzo