From: "David Hildenbrand (Arm)" <david@kernel.org>
To: Chris Down <chris@chrisdown.name>,
Andrew Morton <akpm@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>,
kernel-team@fb.com, linux-mm@kvack.org,
linux-kernel@vger.kernel.org, stable@vger.kernel.org
Subject: Re: [PATCH v2 3/3] selftests/mm: Add UFFDIO_MOVE huge zeropage PMD regression test
Date: Thu, 26 Feb 2026 17:29:51 +0100 [thread overview]
Message-ID: <cb38e860-bd1a-409e-9685-de8dbe9ee54e@kernel.org> (raw)
In-Reply-To: <aaBWG4fajXXbjpVN@chrisdown.name>
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@chrisdown.name>
> ---
> 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
prev parent reply other threads:[~2026-02-26 16:30 UTC|newest]
Thread overview: 2+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-02-26 14:18 Chris Down
2026-02-26 16:29 ` David Hildenbrand (Arm) [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=cb38e860-bd1a-409e-9685-de8dbe9ee54e@kernel.org \
--to=david@kernel.org \
--cc=akpm@linux-foundation.org \
--cc=chris@chrisdown.name \
--cc=kernel-team@fb.com \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=stable@vger.kernel.org \
--cc=willy@infradead.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox