kselftest: mm: fix some failure of split_huge_page_test

There're some random failure for split_huge_page_test when khugepaged collapses pages into pmd again which had split by the test.

Prevent the khugepaged's collapses for split page by setting the mapped pmd-huge-page with MADV_NOHUGEPAGE before split.

Reported-by Kevin Brodsky <[email protected]>

Reviewed-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

---
 tools/testing/selftests/mm/split_huge_page_test.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

tools/testing/selftests/mm/split_huge_page_test.c +0 -0

index 86a603692826..c718269a291f 100644

tools/testing/selftests/mm/split_huge_page_test.c +4 -0

 	/* split all THPs */
 	write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
 		(uint64_t)one_page + len, order);

@@ -313,6 +321,10 @@ static void split_pte_mapped_thp(void)
 		goto out;
 	}
 
+	/* Prevent khugepaged from collapsing the pages. */
+	if (madvise(thp_area, thp_area_size, MADV_NOHUGEPAGE))
+		ksft_print_msg("madvise(MADV_NOHUGEPAGE) failed\n");
+
 	/* Split all THPs through the remapped pages. */
 	write_debugfs(PID_FMT, getpid(), (uint64_t)page_area,
 		      (uint64_t)page_area + page_area_size, 0);

tools/testing/selftests/mm/split_huge_page_test.c +0 -1

 	return 0;
 err_out_close:
 	close(*fd);
 
--

2.43.0

tools/testing/selftests/mm/split_huge_page_test.c static void verify_rss_anon_split_huge_page_all_zeroes(char *one_page, int nr_hp · line 185

Quoted 6 lines

tools/testing/selftests/mm/split_huge_page_test.c +3 -0

 	if (!rss_anon_before)
 		ksft_exit_fail_msg("No RssAnon is allocated before split\n");
 
+	/* Prevent khugepaged from collapsing the pages. */
+	if (madvise(one_page, len, MADV_NOHUGEPAGE))
+		ksft_print_msg("madvise(MADV_NOHUGEPAGE) failed\n");

This should probably terminate the test no? There's no reason to expect this to fail and it's better to fail then to risk a flake :)

tools/testing/selftests/mm/split_huge_page_test.c static void split_pmd_thp_to_order(int order) · line 237

Quoted 13 lines

tools/testing/selftests/mm/split_huge_page_test.c +5 -0

+
 	/* split all THPs */
 	write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
 		      (uint64_t)one_page + len, 0);

@@ -227,6 +231,10 @@ static void split_pmd_thp_to_order(int order)
 	if (!check_huge_anon(one_page, 4 * pmd_pagesize, 4, pmd_pagesize))
 		ksft_exit_fail_msg("No THP is allocated\n");
 
+	/* Prevent khugepaged from collapsing the pages. */
+	if (madvise(one_page, len, MADV_NOHUGEPAGE))
+		ksft_print_msg("madvise(MADV_NOHUGEPAGE) failed\n");
+

Same comment as above, also since this is a repeated pattern, I think it's worth abstracting it like:

	static void madv_nohuge(char *ptr, size_t len)
	{
		if (!madvise(ptr, len, MADV_NOHUGEPAGE))
			return;

		ksft_exit_fail_msg("MADV_NOHUGEPAGE failed, err=%d\n", errno);
	}

Quoted 8 lines

tools/testing/selftests/mm/split_huge_page_test.c +0 -0

 	/* split all THPs */
 	write_debugfs(PID_FMT, getpid(), (uint64_t)one_page,
 		(uint64_t)one_page + len, order);

@@ -313,6 +321,10 @@ static void split_pte_mapped_thp(void)
 		goto out;
 	}
 

Quoted 8 lines

tools/testing/selftests/mm/split_huge_page_test.c +4 -0

+	/* Prevent khugepaged from collapsing the pages. */
+	if (madvise(thp_area, thp_area_size, MADV_NOHUGEPAGE))
+		ksft_print_msg("madvise(MADV_NOHUGEPAGE) failed\n");
+
 	/* Split all THPs through the remapped pages. */
 	write_debugfs(PID_FMT, getpid(), (uint64_t)page_area,
 		      (uint64_t)page_area + page_area_size, 0);

tools/testing/selftests/mm/split_huge_page_test.c static int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size, · line 559

Quoted 6 lines

tools/testing/selftests/mm/split_huge_page_test.c +3 -0

@@ -542,6 +554,9 @@ static int create_pagecache_thp_and_fd(const char *testfile, size_t fd_size,
 		ksft_test_result_skip("Pagecache folio split skipped\n");
 		return -2;
 	}
+	/* Prevent khugepaged from collapsing the pages. */
+	if (madvise(*addr, fd_size, MADV_NOHUGEPAGE))
+		ksft_print_msg("madvise(MADV_NOHUGEPAGE) failed\n");

Obviously same comments re: this and above

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

Reported-by David Hildenbrand (Arm) <[email protected]>

Suggested-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

---
 tools/testing/selftests/mm/vm_util.c | 145 ++++++++++++++++++++++-------------
 tools/testing/selftests/mm/vm_util.h |   1 +
 2 files changed, 92 insertions(+), 54 deletions(-)

tools/testing/selftests/mm/vm_util.c +6 -38

index 4821a3563036..083b72d92b14 100644
@@ -351,31 +351,13 @@ char *__get_smap_entry(void *addr, const char *pattern, char *buf, size_t len)
 	return entry;
 }
 
-static bool __check_pmd_huge(void *addr, char *pattern, int nr_hpages,
-		  uint64_t hpage_size)
-{
-	char buffer[MAX_LINE_LENGTH];
-	uint64_t thp = -1;
-	char *entry;
-
-	entry = __get_smap_entry(addr, pattern, buffer, sizeof(buffer));
-	if (!entry)
-		goto err_out;
-
-	if (sscanf(entry, "%9" SCNu64 " kB", &thp) != 1)
-		ksft_exit_fail_msg("Reading smap error\n");
-
-err_out:
-	return thp == (nr_hpages * (hpage_size >> 10));
-}
-
-static bool check_large_folios(void *addr, size_t len, int nr_hpages,
-		uint64_t hpage_size)
+static bool check_large_folios(int pagemap_fd, int kpageflags_fd,
+			       void *addr, size_t len, int nr_hpages,
+			       uint64_t hpage_size)
 {
 	int order = 0, pagesize = getpagesize();
 	unsigned int nr_pages = hpage_size / pagesize;
 	int orders[MAX_NR_ORDERS], status;
-	int pagemap_fd, kpageflags_fd;
 	bool ret = false;
 
 	if (!nr_pages)

@@ -386,15 +368,6 @@ static bool check_large_folios(void *addr, size_t len, int nr_hpages,
 		ksft_exit_fail_msg("invalid order\n");
 
 	memset(orders, 0, sizeof(int) * MAX_NR_ORDERS);
-	pagemap_fd = open(PAGEMAP_PATH, O_RDONLY);
-	if (pagemap_fd == -1)
-		ksft_exit_fail_msg("read pagemap fail\n");
-
-	kpageflags_fd = open(KPAGEFLAGS_PATH, O_RDONLY);
-	if (kpageflags_fd == -1) {
-		close(pagemap_fd);
-		ksft_exit_fail_msg("read kpageflags fail\n");
-	}
 
 	status = gather_folio_orders(addr, len, pagemap_fd,
 			kpageflags_fd, orders, MAX_NR_ORDERS);

@@ -405,48 +378,112 @@ static bool check_large_folios(void *addr, size_t len, int nr_hpages,
 		ret = true;
 
 out:
-	close(pagemap_fd);
-	close(kpageflags_fd);
 	return ret;
 }
 
-bool check_huge_anon(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
-{
-	uint64_t pmd_pagesize = read_pmd_pagesize();
-
-	if (!pmd_pagesize)
-		ksft_exit_fail_msg("reading PMD pagesize failed\n");
+enum check_huge_type {
+	CHECK_HUGE_ANON,
+	CHECK_HUGE_FILE,

tools/testing/selftests/mm/vm_util.c +78 -14

+
+	switch (type) {
+	case CHECK_HUGE_ANON:
+		return !file;
+	case CHECK_HUGE_FILE:
+		return file && !swapbacked;
+	case CHECK_HUGE_SHMEM:
+		return file & swapbacked;
+	}
 
-	return check_large_folios(addr, len, nr_hpages, hpage_size);
+	return false;
 }
 
-bool check_huge_file(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+static bool __check_huge(void *addr, size_t len, int nr_hpages,
+			 uint64_t hpage_size, enum check_huge_type type)
 {
-	uint64_t pmd_pagesize = read_pmd_pagesize();
+	bool ret = false;
+	int pagemap_fd, kpageflags_fd;
+	int nr_pmd_mappings = 0;
+	uint64_t pmd_pagesize, scan_mapping_size;
+	uint64_t categories, kpf;
+	unsigned long pfn;
+	bool check_pmd_mapping, allow_nonpresent;
+	char *start = addr;
+	char *end = start + len;
 
+	pmd_pagesize = read_pmd_pagesize();
 	if (!pmd_pagesize)
 		ksft_exit_fail_msg("reading PMD pagesize failed\n");
 
-	if (hpage_size == pmd_pagesize)
-		return __check_pmd_huge(addr, "FilePmdMapped:", nr_hpages, hpage_size);
+	check_pmd_mapping = (hpage_size == pmd_pagesize);
+	scan_mapping_size = (nr_hpages > 0) ? hpage_size : psize();
+	/* Some mTHP tests check a partially populated PMD-sized range. */
+	allow_nonpresent = (uint64_t)nr_hpages * hpage_size < len;
 
-	return check_large_folios(addr, len, nr_hpages, hpage_size);
+	pagemap_fd = open(PAGEMAP_PATH, O_RDONLY);
+	if (pagemap_fd < 0)
+		ksft_exit_fail_msg("open pagemap fail\n");
+
+	kpageflags_fd = open(KPAGEFLAGS_PATH, O_RDONLY);
+	if (kpageflags_fd < 0) {
+		close(pagemap_fd);
+		ksft_exit_fail_msg("open kpageflags fail\n");
+	}
+
+	if (!check_pmd_mapping &&
+	    !check_large_folios(pagemap_fd, kpageflags_fd,
+				addr, len, nr_hpages, hpage_size))
+		goto out;
+
+	for (; start < end; start += scan_mapping_size) {
+		categories = pagemap_scan_get_categories(pagemap_fd, start);
+		pfn = pagemap_get_pfn(pagemap_fd, start);
+		if (pfn == -1UL) {
+			if (!allow_nonpresent)
+				goto out;
+			else
+				continue;
+		}
+		if (pageflags_get(pfn, kpageflags_fd, &kpf))
+			ksft_exit_fail_msg("read kpageflags: %s\n", strerror(errno));
+		if (check_pmd_mapping && (categories & PAGE_IS_HUGE))
+			nr_pmd_mappings++;
+		if (kpf & KPF_COMPOUND_TAIL)
+			continue;
+		if (!check_huge_type(categories, kpf, type))
+			goto out;
+	}
+
+	if (check_pmd_mapping && (nr_pmd_mappings != nr_hpages))
+		goto out;
+	ret = true;
+
+out:
+	close(pagemap_fd);
+	close(kpageflags_fd);
+	return ret;
 }
 
-bool check_huge_shmem(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+bool check_huge_anon(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
 {
-	uint64_t pmd_pagesize = read_pmd_pagesize();
-
-	if (!pmd_pagesize)
-		ksft_exit_fail_msg("reading PMD pagesize failed\n");
+	return __check_huge(addr, len, nr_hpages, hpage_size, CHECK_HUGE_ANON);
+}
 
-	if (hpage_size == pmd_pagesize)
-		return __check_pmd_huge(addr, "ShmemPmdMapped:", nr_hpages, hpage_size);
+bool check_huge_file(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+{
+	return __check_huge(addr, len, nr_hpages, hpage_size, CHECK_HUGE_FILE);
+}
 
-	return check_large_folios(addr, len, nr_hpages, hpage_size);
+bool check_huge_shmem(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+{
+	return __check_huge(addr, len, nr_hpages, hpage_size, CHECK_HUGE_SHMEM);
 }
 
 int64_t allocate_transhuge(void *ptr, int pagemap_fd)

tools/testing/selftests/mm/vm_util.h +1 -1

index 9a49af88702e..f19bd17817e5 100644
@@ -18,6 +18,7 @@
 #define PM_SWAP                       BIT_ULL(62)
 #define PM_PRESENT                    BIT_ULL(63)
 
+#define KPF_SWAPBACKED                BIT_ULL(14)
 #define KPF_COMPOUND_HEAD             BIT_ULL(15)
 #define KPF_COMPOUND_TAIL             BIT_ULL(16)
 #define KPF_HWPOISON                  BIT_ULL(19)
 
--

2.43.0

Quoted 9 lines

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

Quoted 3 lines

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Hmm, is the test correctly putting sentinels either side of the VMA? Any test that doesn't risks flaking due to unwanted VMA merges.

Quoted 4 lines

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

You should probably call out the fact you're doing some refactoring here also!

Quoted 3 lines

Reported-by David Hildenbrand (Arm) <[email protected]>

Should always have a Closes: tag if Reported-by: ideally.

Quoted 3 lines

Suggested-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

Generally looks reasonable to me, some nits above and below though :)

Acked-by Lorenzo Stoakes (ARM) <[email protected]>

Quoted 78 lines

---
 tools/testing/selftests/mm/vm_util.c | 145 ++++++++++++++++++++++-------------
 tools/testing/selftests/mm/vm_util.h |   1 +
 2 files changed, 92 insertions(+), 54 deletions(-)

tools/testing/selftests/mm/vm_util.c +6 -38

index 4821a3563036..083b72d92b14 100644
@@ -351,31 +351,13 @@ char *__get_smap_entry(void *addr, const char *pattern, char *buf, size_t len)
 	return entry;
 }
 
-static bool __check_pmd_huge(void *addr, char *pattern, int nr_hpages,
-		  uint64_t hpage_size)
-{
-	char buffer[MAX_LINE_LENGTH];
-	uint64_t thp = -1;
-	char *entry;
-
-	entry = __get_smap_entry(addr, pattern, buffer, sizeof(buffer));
-	if (!entry)
-		goto err_out;
-
-	if (sscanf(entry, "%9" SCNu64 " kB", &thp) != 1)
-		ksft_exit_fail_msg("Reading smap error\n");
-
-err_out:
-	return thp == (nr_hpages * (hpage_size >> 10));
-}
-
-static bool check_large_folios(void *addr, size_t len, int nr_hpages,
-		uint64_t hpage_size)
+static bool check_large_folios(int pagemap_fd, int kpageflags_fd,
+			       void *addr, size_t len, int nr_hpages,
+			       uint64_t hpage_size)
 {
 	int order = 0, pagesize = getpagesize();
 	unsigned int nr_pages = hpage_size / pagesize;
 	int orders[MAX_NR_ORDERS], status;
-	int pagemap_fd, kpageflags_fd;
 	bool ret = false;
 
 	if (!nr_pages)

@@ -386,15 +368,6 @@ static bool check_large_folios(void *addr, size_t len, int nr_hpages,
 		ksft_exit_fail_msg("invalid order\n");
 
 	memset(orders, 0, sizeof(int) * MAX_NR_ORDERS);
-	pagemap_fd = open(PAGEMAP_PATH, O_RDONLY);
-	if (pagemap_fd == -1)
-		ksft_exit_fail_msg("read pagemap fail\n");
-
-	kpageflags_fd = open(KPAGEFLAGS_PATH, O_RDONLY);
-	if (kpageflags_fd == -1) {
-		close(pagemap_fd);
-		ksft_exit_fail_msg("read kpageflags fail\n");
-	}
 
 	status = gather_folio_orders(addr, len, pagemap_fd,
 			kpageflags_fd, orders, MAX_NR_ORDERS);

@@ -405,48 +378,112 @@ static bool check_large_folios(void *addr, size_t len, int nr_hpages,
 		ret = true;
 
 out:
-	close(pagemap_fd);
-	close(kpageflags_fd);
 	return ret;
 }
 
-bool check_huge_anon(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
-{
-	uint64_t pmd_pagesize = read_pmd_pagesize();
-
-	if (!pmd_pagesize)
-		ksft_exit_fail_msg("reading PMD pagesize failed\n");
+enum check_huge_type {
+	CHECK_HUGE_ANON,
+	CHECK_HUGE_FILE,

tools/testing/selftests/mm/vm_util.c static bool check_large_folios(void *addr, size_t len, int nr_hpages, · line 394

Quoted 10 lines

tools/testing/selftests/mm/vm_util.c +7 -2

+	CHECK_HUGE_SHMEM,
+};
 
-	if (hpage_size == pmd_pagesize)
-		return __check_pmd_huge(addr, "AnonHugePages: ", nr_hpages, hpage_size);
+static bool check_huge_type(uint64_t categories, uint64_t kpageflags,
+			    enum check_huge_type type)
+{
+	bool file = categories & PAGE_IS_FILE;
+	bool swapbacked = kpageflags & KPF_SWAPBACKED;

NIT: nice to const these :)

Quoted 123 lines

tools/testing/selftests/mm/vm_util.c +78 -14

+
+	switch (type) {
+	case CHECK_HUGE_ANON:
+		return !file;
+	case CHECK_HUGE_FILE:
+		return file && !swapbacked;
+	case CHECK_HUGE_SHMEM:
+		return file & swapbacked;
+	}
 
-	return check_large_folios(addr, len, nr_hpages, hpage_size);
+	return false;
 }
 
-bool check_huge_file(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+static bool __check_huge(void *addr, size_t len, int nr_hpages,
+			 uint64_t hpage_size, enum check_huge_type type)
 {
-	uint64_t pmd_pagesize = read_pmd_pagesize();
+	bool ret = false;
+	int pagemap_fd, kpageflags_fd;
+	int nr_pmd_mappings = 0;
+	uint64_t pmd_pagesize, scan_mapping_size;
+	uint64_t categories, kpf;
+	unsigned long pfn;
+	bool check_pmd_mapping, allow_nonpresent;
+	char *start = addr;
+	char *end = start + len;
 
+	pmd_pagesize = read_pmd_pagesize();
 	if (!pmd_pagesize)
 		ksft_exit_fail_msg("reading PMD pagesize failed\n");
 
-	if (hpage_size == pmd_pagesize)
-		return __check_pmd_huge(addr, "FilePmdMapped:", nr_hpages, hpage_size);
+	check_pmd_mapping = (hpage_size == pmd_pagesize);
+	scan_mapping_size = (nr_hpages > 0) ? hpage_size : psize();
+	/* Some mTHP tests check a partially populated PMD-sized range. */
+	allow_nonpresent = (uint64_t)nr_hpages * hpage_size < len;
 
-	return check_large_folios(addr, len, nr_hpages, hpage_size);
+	pagemap_fd = open(PAGEMAP_PATH, O_RDONLY);
+	if (pagemap_fd < 0)
+		ksft_exit_fail_msg("open pagemap fail\n");
+
+	kpageflags_fd = open(KPAGEFLAGS_PATH, O_RDONLY);
+	if (kpageflags_fd < 0) {
+		close(pagemap_fd);
+		ksft_exit_fail_msg("open kpageflags fail\n");
+	}
+
+	if (!check_pmd_mapping &&
+	    !check_large_folios(pagemap_fd, kpageflags_fd,
+				addr, len, nr_hpages, hpage_size))
+		goto out;
+
+	for (; start < end; start += scan_mapping_size) {
+		categories = pagemap_scan_get_categories(pagemap_fd, start);
+		pfn = pagemap_get_pfn(pagemap_fd, start);
+		if (pfn == -1UL) {
+			if (!allow_nonpresent)
+				goto out;
+			else
+				continue;
+		}
+		if (pageflags_get(pfn, kpageflags_fd, &kpf))
+			ksft_exit_fail_msg("read kpageflags: %s\n", strerror(errno));
+		if (check_pmd_mapping && (categories & PAGE_IS_HUGE))
+			nr_pmd_mappings++;
+		if (kpf & KPF_COMPOUND_TAIL)
+			continue;
+		if (!check_huge_type(categories, kpf, type))
+			goto out;
+	}
+
+	if (check_pmd_mapping && (nr_pmd_mappings != nr_hpages))
+		goto out;
+	ret = true;
+
+out:
+	close(pagemap_fd);
+	close(kpageflags_fd);
+	return ret;
 }
 
-bool check_huge_shmem(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+bool check_huge_anon(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
 {
-	uint64_t pmd_pagesize = read_pmd_pagesize();
-
-	if (!pmd_pagesize)
-		ksft_exit_fail_msg("reading PMD pagesize failed\n");
+	return __check_huge(addr, len, nr_hpages, hpage_size, CHECK_HUGE_ANON);
+}
 
-	if (hpage_size == pmd_pagesize)
-		return __check_pmd_huge(addr, "ShmemPmdMapped:", nr_hpages, hpage_size);
+bool check_huge_file(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+{
+	return __check_huge(addr, len, nr_hpages, hpage_size, CHECK_HUGE_FILE);
+}
 
-	return check_large_folios(addr, len, nr_hpages, hpage_size);
+bool check_huge_shmem(void *addr, size_t len, int nr_hpages, uint64_t hpage_size)
+{
+	return __check_huge(addr, len, nr_hpages, hpage_size, CHECK_HUGE_SHMEM);
 }
 
 int64_t allocate_transhuge(void *ptr, int pagemap_fd)

tools/testing/selftests/mm/vm_util.h +1 -1

index 9a49af88702e..f19bd17817e5 100644
@@ -18,6 +18,7 @@
 #define PM_SWAP                       BIT_ULL(62)
 #define PM_PRESENT                    BIT_ULL(63)
 
+#define KPF_SWAPBACKED                BIT_ULL(14)
 #define KPF_COMPOUND_HEAD             BIT_ULL(15)
 #define KPF_COMPOUND_TAIL             BIT_ULL(16)
 #define KPF_HWPOISON                  BIT_ULL(19)
 
--

Quoted 2 lines

2.43.0

--
Cheers, Lorenzo

Quoted 14 lines · Yeoreum Yun

Quoted 12 lines

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Quoted 2 lines

Hmm, is the test correctly putting sentinels either side of the VMA? Any test that doesn't risks flaking due to unwanted VMA merges.

I believe that with this change, we don’t need to worry about unwanted VMA merges when checking for huge pages, since the test no longer relies on VMA sentinels but directly checks whether the mapping is huge or not.

Also, this flaky failure was not caused by a VMA merge, but by a change in glibc’s behavior that sets HUGEPAGE for sufficiently large areas.

Might for the *NO_HUGEPAGE* setup, there would be a chance to merge VMA area, But since it seraches the mapping directly, it's fine.

Quoted 8 lines

Quoted 5 lines

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

You should probably call out the fact you're doing some refactoring here also!

Okay. I'll spell out with some detail. Thanks!

Quoted 6 lines

Quoted 4 lines

Reported-by David Hildenbrand (Arm) <[email protected]>

Should always have a Closes: tag if Reported-by: ideally.

Yes. but talked with personally nothing to close. So Reported-by tag only. Would it be better to remove?

Quoted 14 lines

Quoted 14 lines · Yeoreum Yun

Quoted 12 lines

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Quoted 6 lines

Quoted 2 lines

Hmm, is the test correctly putting sentinels either side of the VMA? Any test that doesn't risks flaking due to unwanted VMA merges.

I believe that with this change, we don’t need to worry about unwanted VMA merges when checking for huge pages, since the test no longer relies on VMA sentinels but directly checks whether the mapping is huge or not.

We always need to worry about that though, anything that asserts stuff on a VMA level must prevent unwanted merges first.

I guess I'm asking you in a roundabout way to go check and if there isn't isolation like that (i.e. mapping PROT_NONE VMAs around the VMA you want) then to go do that :)

I think glibc going off and mapping a bigger VMA is possibly a symptom of this not being done.

Quoted 4 lines

Also, this flaky failure was not caused by a VMA merge, but by a change in glibc’s behavior that sets HUGEPAGE for sufficiently large areas.

Yep your change is worthwhile reagrdless! :)

Quoted 4 lines

Might for the *NO_HUGEPAGE* setup, there would be a chance to merge VMA area, But since it seraches the mapping directly, it's fine.

As above, generally we should put this protection in case, even if it's unlikely for a VMA with VMA_NOHUGEPAGE_BIT set!

Quoted 11 lines

Quoted 9 lines

Quoted 6 lines

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

You should probably call out the fact you're doing some refactoring here also!

Okay. I'll spell out with some detail. Thanks!

Thnks!

Quoted 5 lines

Quoted 5 lines

Quoted 4 lines

Reported-by David Hildenbrand (Arm) <[email protected]>

Quoted 4 lines

Quoted 1 line

Should always have a Closes: tag if Reported-by: ideally.

Yes. but talked with personally nothing to close. So Reported-by tag only. Would it be better to remove?

Yeah or Suggested-by maybe? If he said it to you on-list then a Closes:
<lore link> works also!

Quoted 22 lines · Yeoreum Yun

Quoted 14 lines

Quoted 14 lines · Yeoreum Yun

Quoted 12 lines

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Quoted 6 lines

Quoted 2 lines

Hmm, is the test correctly putting sentinels either side of the VMA? Any test that doesn't risks flaking due to unwanted VMA merges.

I believe that with this change, we don’t need to worry about unwanted VMA merges when checking for huge pages, since the test no longer relies on VMA sentinels but directly checks whether the mapping is huge or not.

Quoted 6 lines

We always need to worry about that though, anything that asserts stuff on a VMA level must prevent unwanted merges first.

I guess I'm asking you in a roundabout way to go check and if there isn't isolation like that (i.e. mapping PROT_NONE VMAs around the VMA you want) then to go do that :)

Oh. Sorry not to catch your intend. not for only this series but for the future usage, might it sound reasonable to add some mem alloca fucntion for the test which could replace memalign() with above concept.

Quoted 2 lines

I think glibc going off and mapping a bigger VMA is possibly a symptom of this not being done.

TBH, what user would generally care whether VMAs are merged or not? Anyway, it would be better to add a helper with some PROT_NONE guards to prevent unwanted VMA merges while testing.

Quoted 7 lines

Quoted 5 lines

Also, this flaky failure was not caused by a VMA merge, but by a change in glibc’s behavior that sets HUGEPAGE for sufficiently large areas.

Yep your change is worthwhile reagrdless! :)

Thanks :)

Quoted 2 lines

Quoted 2 lines

Quoted 5 lines

Quoted 2 lines

Might for the *NO_HUGEPAGE* setup, there would be a chance to merge VMA area, But since it seraches the mapping directly, it's fine.

As above, generally we should put this protection in case, even if it's unlikely for a VMA with VMA_NOHUGEPAGE_BIT set!

Agree. let me add one for this and replace memalign() used in khugepagd and split_huge_page_test at first.

Quoted 24 lines

Quoted 11 lines

Quoted 9 lines

Quoted 6 lines

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

You should probably call out the fact you're doing some refactoring here also!

Okay. I'll spell out with some detail. Thanks!

Thnks!

Quoted 9 lines

Quoted 6 lines

Quoted 4 lines

Reported-by David Hildenbrand (Arm) <[email protected]>

Should always have a Closes: tag if Reported-by: ideally.

Yes. but talked with personally nothing to close. So Reported-by tag only. Would it be better to remove?

Quoted 2 lines

Yeah or Suggested-by maybe? If he said it to you on-list then a Closes:
<lore link> works also!

Okay. I'll drop the Reported-by: and change to Suggseted-by.

Thanks!

--
Sincerely,
Yeoreum Yun

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

Reported-by David Hildenbrand (Arm) <[email protected]>

Suggested-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

---
 tools/testing/selftests/mm/vm_util.c | 145 ++++++++++++++++++++++-------------
 tools/testing/selftests/mm/vm_util.h |   1 +
 2 files changed, 92 insertions(+), 54 deletions(-)

tools/testing/selftests/mm/vm_util.c +0 -0

index 4821a3563036..083b72d92b14 100644
@@ -351,31 +351,13 @@ char *__get_smap_entry(void *addr, const char *pattern, char *buf, size_t len)
 	return entry;
 }
 

tools/testing/selftests/mm/vm_util.c +9 -0

+	for (; start < end; start += scan_mapping_size) {
+		categories = pagemap_scan_get_categories(pagemap_fd, start);
+		pfn = pagemap_get_pfn(pagemap_fd, start);
+		if (pfn == -1UL) {
+			if (!allow_nonpresent)
+				goto out;
+			else
+				continue;
+		}

Quoted 32 lines · Yeoreum Yun

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

Reported-by David Hildenbrand (Arm) <[email protected]>

Suggested-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

---
 tools/testing/selftests/mm/vm_util.c | 145 ++++++++++++++++++++++-------------
 tools/testing/selftests/mm/vm_util.h |   1 +
 2 files changed, 92 insertions(+), 54 deletions(-)

tools/testing/selftests/mm/vm_util.c +0 -0

index 4821a3563036..083b72d92b14 100644
@@ -351,31 +351,13 @@ char *__get_smap_entry(void *addr, const char *pattern, char *buf, size_t len)
 	return entry;
 }
 

<snip>

Quoted 9 lines

tools/testing/selftests/mm/vm_util.c +9 -0

+	for (; start < end; start += scan_mapping_size) {
+		categories = pagemap_scan_get_categories(pagemap_fd, start);
+		pfn = pagemap_get_pfn(pagemap_fd, start);
+		if (pfn == -1UL) {
+			if (!allow_nonpresent)
+				goto out;
+			else
+				continue;
+		}

tools/testing/selftests/mm/vm_util.c static bool check_large_folios(void *addr, size_t len, int nr_hpages, · line 463

Quoted 10 lines

tools/testing/selftests/mm/vm_util.c +10 -0

+		if (pageflags_get(pfn, kpageflags_fd, &kpf))
+			ksft_exit_fail_msg("read kpageflags: %s\n", strerror(errno));
+		if (check_pmd_mapping && (categories & PAGE_IS_HUGE))
+			nr_pmd_mappings++;
+		if (kpf & KPF_COMPOUND_TAIL)
+			continue;
+		if (!check_huge_type(categories, kpf, type))
+			goto out;
+	}
+	

There is a trailing tab.

tools/testing/selftests/mm/vm_util.c static bool check_large_folios(void *addr, size_t len, int nr_hpages, · line 472

Quoted 10 lines

tools/testing/selftests/mm/vm_util.c +8 -0

+	if (check_pmd_mapping && (nr_pmd_mappings != nr_hpages))
+		goto out;
+	ret = true;
+
+out:
+	close(pagemap_fd);
+	close(kpageflags_fd);
+	return ret;
 }

Otherwise, LGTM. I assume the replacement of memalign(), from the discussion between you and Lorenzo, will be added in a separate patch. For this patch,

Reviewed-by Zi Yan <[email protected]>

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

Reported-by David Hildenbrand (Arm) <[email protected]>

Suggested-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

---

Quoted 13 lines · Yeoreum Yun

Since glibc commit 321e1fc73f (“malloc: Enable 2MB THP by default on AArch64”), glibc may call madvise(MADV_HUGEPAGE) for sufficiently large allocations made by memalign().

The underlying VMA may start at a different address from the aligned address returned by memalign(). Furthermore, a subsequent madvise(MADV_HUGEPAGE) call does not split the VMA because the flag is already set.

This causes split_huge_page_test to fail because the check_huge_xxx() helpers incorrectly require the address returned by memalign() to match the VMA start address reported in /proc/self/smaps.

Quoted 7 lines

Fix this by using /proc/self/pagemap and /proc/kpageflags instead of /proc/self/smaps to detect huge pages.

Reported-by David Hildenbrand (Arm) <[email protected]>

Suggested-by Zi Yan <[email protected]>

Signed-off-by Yeoreum Yun <[email protected]>

---

LGTM, with comments from Lorenzo and Zi addressed.

Reviewed-by Baolin Wang <[email protected]>

Tested-by Baolin Wang <[email protected]>