Skip to contentSkip to conversation

selftests/mm: skip hugetlb_dio tests when DIO alignment is incompatible

PATCH Superseded Li Wang 2 participants 4 messages Mar 27 – Mar 30

Subsystems (4)

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Li Wang

author
#0 lore raw

PATCH v2 1/1 selftests/mm: skip hugetlb_dio tests when DIO alignment is incompatible

hugetlb_dio test uses sub-page offsets (pagesize / 2) to verify that hugepages used as DIO user buffers are correctly unpinned at completion.

However, on filesystems with a logical block size larger than half the page size (e.g., 4K-sector block devices), these unaligned DIO writes are rejected with -EINVAL, causing the test to fail unexpectedly.

Add get_dio_alignment() to query the filesystem's required DIO alignment via statx(STATX_DIOALIGN) and pass it to run_dio_using_hugetlb(). Skip individual test cases whose write length is not a multiple of the alignment, so that aligned cases are still tested.

=== Reproduce Steps ===

  # dd if=/dev/zero of=/tmp/test.img bs=1M count=512
  # losetup --sector-size 4096 /dev/loop0 /tmp/test.img
  # mkfs.xfs /dev/loop0
  # mkdir -p /mnt/dio_test
  # mount /dev/loop0 /mnt/dio_test

// Modify test to open /mnt/dio_test and rebuild it:

  -       fd = open("/tmp", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
  +       fd = open("/mnt/dio_test", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
  # getconf PAGESIZE
  4096
  # echo 100 >/proc/sys/vm/nr_hugepages
  # ./hugetlb_dio
  TAP version 13
  1..4
  # No. Free pages before allocation : 100
  # No. Free pages after munmap : 100
  ok 1 free huge pages from 0-12288
  Bail out! Error writing to file
  : Invalid argument (22)
  # Planned tests != run tests (4 != 1)
  # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0

---

Notes:
v2:
- Pass dio_align as a parameter to run_dio_using_hugetlb() instead of generally page_size/2 alignment check.

 tools/testing/selftests/mm/hugetlb_dio.c | 56 +++++++++++++++++-------
 1 file changed, 40 insertions(+), 16 deletions(-)
tools/testing/selftests/mm/hugetlb_dio.c +40 -16
index 9ac62eb4c97d..638a8abe3ae2 100644
@@ -20,7 +20,35 @@
 #include "vm_util.h"
 #include "kselftest.h"
 
-void run_dio_using_hugetlb(unsigned int start_off, unsigned int end_off)
+#ifndef STATX_DIOALIGN
+#define STATX_DIOALIGN		0x00002000U
+#endif
+
+unsigned int get_dio_alignment(void)
+{
+	int fd, ret;
+	struct statx stx;
+	unsigned int dio_align = 1;
+
+	fd = open("/tmp", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
+	if (fd < 0)
+		ksft_exit_skip("Unable to allocate file: %s\n", strerror(errno));
+
+	ret = statx(fd, "", AT_EMPTY_PATH, STATX_DIOALIGN, &stx);
+	if (ret < 0) {
+		ksft_perror("statx() failed");
+	} else if ((stx.stx_mask & STATX_DIOALIGN) &&
+			stx.stx_dio_offset_align) {
+		dio_align = stx.stx_dio_offset_align;
+	}
+
+	close(fd);
+
+	return dio_align;
+}
+
+void run_dio_using_hugetlb(unsigned int start_off, unsigned int end_off,
+			unsigned int dio_align)
 {
 	int fd;
 	char *buffer =  NULL;
@@ -33,6 +61,11 @@ void run_dio_using_hugetlb(unsigned int start_off, unsigned int end_off)
 	const int mmap_prot  = PROT_READ | PROT_WRITE;
 
 	writesize = end_off - start_off;
+	if (writesize % dio_align != 0) {
+		ksft_test_result_skip("DIO alignment (%u) incompatible with offset %zu\n",
+				dio_align, writesize);
+		return;
+	}
 
 	/* Get the default huge page size */
 	h_pagesize = default_huge_page_size();
@@ -89,37 +122,28 @@ void run_dio_using_hugetlb(unsigned int start_off, unsigned int end_off)
 
 int main(void)
 {
-	size_t pagesize = 0;
-	int fd;
+	size_t pagesize = psize();
+	unsigned int dio_align = get_dio_alignment();
 
 	ksft_print_header();
 
-	/* Open the file to DIO */
-	fd = open("/tmp", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
-	if (fd < 0)
-		ksft_exit_skip("Unable to allocate file: %s\n", strerror(errno));
-	close(fd);
-
 	/* Check if huge pages are free */
 	if (!get_free_hugepages())
 		ksft_exit_skip("No free hugepage, exiting\n");
 
 	ksft_set_plan(4);
 
-	/* Get base page size */
-	pagesize  = psize();
-
 	/* start and end is aligned to pagesize */
-	run_dio_using_hugetlb(0, (pagesize * 3));
+	run_dio_using_hugetlb(0, (pagesize * 3), dio_align);
 
 	/* start is aligned but end is not aligned */
-	run_dio_using_hugetlb(0, (pagesize * 3) - (pagesize / 2));
+	run_dio_using_hugetlb(0, (pagesize * 3) - (pagesize / 2), dio_align);
 
 	/* start is unaligned and end is aligned */
-	run_dio_using_hugetlb(pagesize / 2, (pagesize * 3));
+	run_dio_using_hugetlb(pagesize / 2, (pagesize * 3), dio_align);
 
 	/* both start and end are unaligned */
-	run_dio_using_hugetlb(pagesize / 2, (pagesize * 3) + (pagesize / 2));
+	run_dio_using_hugetlb(pagesize / 2, (pagesize * 3) + (pagesize / 2), dio_align);
 
 	ksft_finished();
 }
Signature · 1 line
2.53.0
maintainer
#1 lore raw
Quoted 40 lines · On Fri, 27 Mar 2026 20:03:05 +0800 Li Wang <[email protected]> wrote:

hugetlb_dio test uses sub-page offsets (pagesize / 2) to verify that hugepages used as DIO user buffers are correctly unpinned at completion.

However, on filesystems with a logical block size larger than half the page size (e.g., 4K-sector block devices), these unaligned DIO writes are rejected with -EINVAL, causing the test to fail unexpectedly.

Add get_dio_alignment() to query the filesystem's required DIO alignment via statx(STATX_DIOALIGN) and pass it to run_dio_using_hugetlb(). Skip individual test cases whose write length is not a multiple of the alignment, so that aligned cases are still tested.

=== Reproduce Steps ===

  # dd if=/dev/zero of=/tmp/test.img bs=1M count=512
  # losetup --sector-size 4096 /dev/loop0 /tmp/test.img
  # mkfs.xfs /dev/loop0
  # mkdir -p /mnt/dio_test
  # mount /dev/loop0 /mnt/dio_test

// Modify test to open /mnt/dio_test and rebuild it:

  -       fd = open("/tmp", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
  +       fd = open("/mnt/dio_test", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
  # getconf PAGESIZE
  4096
  # echo 100 >/proc/sys/vm/nr_hugepages
  # ./hugetlb_dio
  TAP version 13
  1..4
  # No. Free pages before allocation : 100
  # No. Free pages after munmap : 100
  ok 1 free huge pages from 0-12288
  Bail out! Error writing to file
  : Invalid argument (22)
  # Planned tests != run tests (4 != 1)
  # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
Quoted 6 lines

...

v2:
- Pass dio_align as a parameter to run_dio_using_hugetlb() instead of generally page_size/2 alignment check.

Whee, AI review has decided that pre glibc-2.36 is a problem (last time it was pre-2.37).

And it's forgotten the previous fs-doesnt-support-DIO issues. Did you alter that?

The get_dio_alignment-before-ksft_print_header thing seems legit.

https://sashiko.dev/#/patchset/[email protected]

author
#2 lore raw
Quoted 44 lines · Andrew Morton
Quoted 42 lines · On Fri, 27 Mar 2026 20:03:05 +0800 Li Wang <[email protected]> wrote:

hugetlb_dio test uses sub-page offsets (pagesize / 2) to verify that hugepages used as DIO user buffers are correctly unpinned at completion.

However, on filesystems with a logical block size larger than half the page size (e.g., 4K-sector block devices), these unaligned DIO writes are rejected with -EINVAL, causing the test to fail unexpectedly.

Add get_dio_alignment() to query the filesystem's required DIO alignment via statx(STATX_DIOALIGN) and pass it to run_dio_using_hugetlb(). Skip individual test cases whose write length is not a multiple of the alignment, so that aligned cases are still tested.

=== Reproduce Steps ===

  # dd if=/dev/zero of=/tmp/test.img bs=1M count=512
  # losetup --sector-size 4096 /dev/loop0 /tmp/test.img
  # mkfs.xfs /dev/loop0
  # mkdir -p /mnt/dio_test
  # mount /dev/loop0 /mnt/dio_test

// Modify test to open /mnt/dio_test and rebuild it:

  -       fd = open("/tmp", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
  +       fd = open("/mnt/dio_test", O_TMPFILE | O_RDWR | O_DIRECT, 0664);
  # getconf PAGESIZE
  4096
  # echo 100 >/proc/sys/vm/nr_hugepages
  # ./hugetlb_dio
  TAP version 13
  1..4
  # No. Free pages before allocation : 100
  # No. Free pages after munmap : 100
  ok 1 free huge pages from 0-12288
  Bail out! Error writing to file
  : Invalid argument (22)
  # Planned tests != run tests (4 != 1)
  # Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0

...

Quoted 7 lines
Quoted 4 lines

v2:
- Pass dio_align as a parameter to run_dio_using_hugetlb() instead of generally page_size/2 alignment check.

Whee, AI review has decided that pre glibc-2.36 is a problem (last time it was pre-2.37).

Ok, will handle this build issue with old glibc.

Quoted 2 lines

And it's forgotten the previous fs-doesnt-support-DIO issues. Did you alter that?

[Sorry, forgot to mention in change logs]

Yes, I added O_DIRECT flag back to the first open(). Then it'd skip into a failure on filesystems that doesn't support O_DIRECT.

Also, statx on FS doesn't support DIO will set stx_dio_offset_align to 0, which may led a division by zero issue, I added a zero check.

	} else if ((stx.stx_mask & STATX_DIOALIGN) &&
			stx.stx_dio_offset_align) {
Quoted 1 line

The get_dio_alignment-before-ksft_print_header thing seems legit.

+1

And, before sending v3, I would leave this patch for more days to see if others comment.

Signature · 2 lines
Regards,
Li Wang
author
#3 lore raw
In reply to Andrew Morton

Reply Sashiko comment:

Quoted 12 lines

https://sashiko.dev/#/patchset/[email protected]

Quoted 1 line
tools/testing/selftests/mm/hugetlb_dio.c +1 -0
+	if (writesize % dio_align != 0) {
tools/testing/selftests/mm/hugetlb_dio.c +4 -0
+		ksft_test_result_skip("DIO alignment (%u) incompatible with offset %zu\n",
+				dio_align, writesize);
+		return;
+	}

Is this alignment check complete?

Direct I/O requires both the transfer length and the memory buffer address to be aligned. Later in this function, start_off is used as the buffer offset:

Quoted 8 lines
	buffer = orig_buffer;
	buffer += start_off;

If start_off is pagesize / 2 (e.g., 2048) and writesize is pagesize * 3 (e.g., 12288), writesize is a multiple of a 4096-byte alignment, so the test is not skipped.

However, the memory buffer itself is only 2048-byte aligned. Will the subsequent write() still fail with -EINVAL on 4K-sector devices?

TL;DR: Yes, we should do both buffer address and writesize alignment
checks to satisfy all FS types.

Looking at the kernel code: fs/iomap/direct-dio.c, the only alignment check there is at line#413, which checks file's pos and write length.

EXT4:

ext4_file_write_iter
ext4_dio_write_iter

    iomap_dio_rw
      __iomap_dio_rw
        iomap_dio_iter
          iomap_dio_bio_iter
  390	static int iomap_dio_bio_iter(struct iomap_iter *iter, struct iomap_dio *dio)
  391	{
  		...
  403
  404		/*
  405		 * File systems that write out of place and always allocate new blocks
  406		 * need each bio to be block aligned as that's the unit of allocation.
  407		 */
  408		if (dio->flags & IOMAP_DIO_FSBLOCK_ALIGNED)
  409			alignment = fs_block_size;
  410		else
  411			alignment = bdev_logical_block_size(iomap->bdev);
  412
  413		if ((pos | length) & (alignment - 1))
  414			return -EINVAL;
  415		...

Sashiko points out the buffer-address should do alignment check as well, I firstly suspect it based on the FS extra check before the iomap_dio_rw:

ext4_file_write_iter
ext4_dio_write_iter
ext4_should_use_dio
iov_iter_alignment <--- do buffer/writesize alignment check

  842 unsigned long iov_iter_alignment(const struct iov_iter *i)
  843 {
  		...
  853                 return iov_iter_alignment_iovec(i);
  		...
  865 }

799 static unsigned long iov_iter_alignment_iovec(const struct iov_iter *i) 800 {
...

  809                   res |= (unsigned long)iov->iov_base + skip;
  812                   res |= len;
		...

818 return res;
819 }

But eventually I found that this is only fallback to the buffer I/O when the direct I/O is unsupported (go to: ext4_buffered_write_iter). This wouldn't happen in the test as it open with O_DIRECT flag.

Then, I turned to look at Btrfs path:

btrfs_file_write_iter
btrfs_do_write_iter
btrfs_direct_write
check_direct_IO <--- do buffer alignment check
...
btrfs_dio_write
__iomap_dio_rw <--- do samething like ext4

  778 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
  779                                const struct iov_iter *iter, loff_t offset)
  780 {
  781         const u32 blocksize_mask = fs_info->sectorsize - 1;
  782
  783         if (offset & blocksize_mask)
  784                 return -EINVAL;
  785
  786         if (iov_iter_alignment(iter) & blocksize_mask)
  787                 return -EINVAL;
  788         return 0;
  789 }

Yes, here I found the evendice that iov_iter_alignment(iter) & blocksize_mask) do the alignment check.

Unlike ext4 which never reaches the check for normal files, btrfs always checks buffer alignment for every DIO operation. And it's a hard -EINVAL, not a silent fallback to buffered I/O.

Signature · 2 lines
Regards,
Li Wang