Re: [PATCH v2 3/3] selftests/mm: Add UFFDIO_MOVE huge zeropage PMD regression test

[David Hildenbrand (Arm)]

On 2/26/26 15:18, Chris Down wrote:
> The existing uffd-unit-tests move-pmd coverage exercises PMD-sized
> UFFDIO_MOVE on anonymous THPs, but it does not force the huge zeropage
> PMD path in move_pages_huge_pmd().
> 
> Add a dedicated anonymous UFFDIO_MOVE PMD test that exercises this
> relatively comprehensively.
> 
> Signed-off-by: Chris Down <chris@xxxxxxxxxxxxxx>
> ---
>  tools/testing/selftests/mm/uffd-unit-tests.c | 176 +++++++++++++++++++
>  1 file changed, 176 insertions(+)
> 
> diff --git a/tools/testing/selftests/mm/uffd-unit-tests.c b/tools/testing/selftests/mm/uffd-unit-tests.c
> index 6f5e404a446c..372619e3906d 100644
> --- a/tools/testing/selftests/mm/uffd-unit-tests.c
> +++ b/tools/testing/selftests/mm/uffd-unit-tests.c
> @@ -203,6 +203,62 @@ static int pagemap_open(void)
>  	return fd;
>  }
>  

Thanks for implementing this!

[...]

> +static bool uffd_pagemap_scan_get_categories(int fd, char *start, uint64_t *categories)
> +{
> +	struct page_region r = { 0 };
> +	long ret;
> +
> +	if (!uffd_pagemap_scan_supported(fd, start))
> +		return false;
> +
> +	ret = uffd_pagemap_scan_get_categories_raw(fd, start, &r);
> +	if (ret < 0)
> +		err("PAGEMAP_SCAN failed: %s", strerror(errno));
> +
> +	*categories = ret ? r.categories : 0;
> +	return true;
> +}

Can we unify (deduplicate) that code with what we have in vm_util.c, and
then only export a function like

bool pagemap_is_huge_zero(int fd, char *start)

similar to pagemap_is_softdirty() etc to be consumed by the test here?

It can simply return "false" in case the interface is not available.

> +
>  /* This macro let __LINE__ works in err() */
>  #define  pagemap_check_wp(value, wp) do {				\
>  		if (!!(value & PM_UFFD_WP) != wp)			\
> @@ -1227,6 +1283,119 @@ static void uffd_move_pmd_test(uffd_global_test_opts_t *gopts, uffd_test_args_t
>  			      uffd_move_pmd_handle_fault);
>  }
>  
> +static void uffd_move_pmd_huge_zeropage_test(uffd_global_test_opts_t *gopts,
> +					     uffd_test_args_t *targs)
> +{
> +	unsigned long pmd_size = read_pmd_pagesize();
> +	unsigned long pmd_pages;
> +	unsigned long bytes = gopts->nr_pages * gopts->page_size;
> +	char *orig_area_src = gopts->area_src, *orig_area_dst = gopts->area_dst;
> +	char *aligned_src, *aligned_dst;
> +	unsigned long src_offs, dst_offs, max_offs;
> +	pthread_t uffd_mon;
> +	struct uffd_args args = { 0 };
> +	char c = '\0';
> +	int pagemap_fd;
> +	uint64_t categories;
> +	unsigned long i;
> +
> +	if (pmd_size <= gopts->page_size) {
> +		uffd_test_skip("huge page size is 0, feature missing?");
> +		return;
> +	}
> +	if (!detect_huge_zeropage()) {
> +		uffd_test_skip("transparent huge zeropage disabled");
> +		return;
> +	}
> +
> +	pmd_pages = pmd_size / gopts->page_size;
> +	if (bytes < pmd_size) {
> +		uffd_test_skip("not enough pages for one PMD-sized move");
> +		return;
> +	}
> +
> +	aligned_src = ALIGN_UP(orig_area_src, pmd_size);
> +	aligned_dst = ALIGN_UP(orig_area_dst, pmd_size);
> +	src_offs = (aligned_src - orig_area_src) / gopts->page_size;
> +	dst_offs = (aligned_dst - orig_area_dst) / gopts->page_size;
> +	max_offs = src_offs > dst_offs ? src_offs : dst_offs;
> +	if (max_offs + pmd_pages > gopts->nr_pages) {
> +		uffd_test_skip("could not find aligned PMD-sized src/dst window");
> +		return;
> +	}
> +
> +	if (madvise(orig_area_dst, bytes, MADV_HUGEPAGE))
> +		err("madvise(MADV_HUGEPAGE) failure");
> +	if (madvise(orig_area_src, bytes, MADV_DONTFORK))
> +		err("madvise(MADV_DONTFORK) failure");
> +	if (madvise(aligned_src, pmd_size, MADV_DONTNEED))
> +		err("madvise(MADV_DONTNEED) failure");
> +
> +	/* Materialise a PMD-sized huge zeropage mapping in the source. */

"Fault in the huge zeropage"

> +	force_read_pages(aligned_src, pmd_pages, gopts->page_size);
> +
> +	pagemap_fd = pagemap_open();
> +	if (!uffd_pagemap_scan_get_categories(pagemap_fd, aligned_src, &categories)) {
> +		close(pagemap_fd);
> +		uffd_test_skip("PAGEMAP_SCAN unsupported");
> +		return;
> +	}
> +	if ((categories & (PAGE_IS_PRESENT | PAGE_IS_PFNZERO | PAGE_IS_HUGE)) !=
> +	    (PAGE_IS_PRESENT | PAGE_IS_PFNZERO | PAGE_IS_HUGE)) {
> +		close(pagemap_fd);
> +		uffd_test_skip("could not materialise a huge zeropage PMD mapping");

"could not fault in the huge zeropage" ?

> +		return;
> +	}
> +	gopts->area_src = aligned_src;
> +	gopts->area_dst = aligned_dst;
> +
> +	if (uffd_register(gopts->uffd, gopts->area_dst, pmd_size, true, false, false))
> +		err("register failure");
> +
> +	args.gopts = gopts;
> +	args.handle_fault = uffd_move_pmd_handle_fault;
> +	if (pthread_create(&uffd_mon, NULL, uffd_poll_thread, &args))
> +		err("uffd_poll_thread create");
> +
> +	/*
> +	 * One fault on dst should trigger a single PMD-sized UFFDIO_MOVE from
> +	 * the huge zeropage PMD we populated in the source.
> +	 */
> +	force_read_pages(gopts->area_dst, pmd_pages, gopts->page_size);
> +
> +	if (write(gopts->pipefd[1], &c, sizeof(c)) != sizeof(c))
> +		err("pipe write");
> +	if (pthread_join(uffd_mon, NULL))
> +		err("join() failed");
> +
> +	if (args.missing_faults != 1 || args.minor_faults != 0) {
> +		uffd_test_fail("stats check error");
> +	} else if (!uffd_pagemap_scan_get_categories(pagemap_fd, gopts->area_dst,
> +						     &categories)) {
> +		uffd_test_fail("PAGEMAP_SCAN unsupported");

Why should it suddenly be unsupported when it worked before?

With the pagemap_is_huge_zero() helper, this will boil down to

} else if (pagemap_is_huge_zero(...)) {
	uffd_test_fail("moved destination is not a huge zeropage PMD");
}

> +	} else if ((categories & (PAGE_IS_PRESENT | PAGE_IS_PFNZERO |
> +				  PAGE_IS_HUGE)) !=
> +			(PAGE_IS_PRESENT | PAGE_IS_PFNZERO | PAGE_IS_HUGE)) {
> +		uffd_test_fail("moved destination is not a huge zeropage PMD");
> +	} else if (!check_huge_anon(gopts->area_dst, 0, pmd_size)) {
> +		/* vm_normal_page_pmd() must continue to treat the moved PMD as special. */
> +		uffd_test_fail("moved huge zeropage PMD counted as AnonHugePages");
> +	} else {
> +		for (i = 0; i < pmd_size; i++) {
> +			if (gopts->area_dst[i]) {
> +				uffd_test_fail("moved huge zeropage PMD data is not zero");
> +				goto out_restore;
> +			}
> +		}

That's a bit excessive given that pagemap told us that it is indeed the
huge zeropage. So I would just drop this.


-- 
Cheers,

David