Skip to contentSkip to conversation

Directory is not persisted after writing to the file within directory if system crashes

DISCUSSION Vyacheslav Kovalevsky 4 participants 6 messages Oct 24, 2025 – Oct 27, 2025 Btrfs file system

lore
Share & tools
lore
#0 lore raw

Under some circumstances, directory entry is not persisted after writing to the file inside the directory that was opened with O_SYNC flag if system crashes.

Detailed description
====================

Hello, we have found another issue with btrfs crash behavior.

In short, empty file is created and synced. Then, a new directory is created, old file is opened with O_SYNC flag and some data is written. After this, a new hard link is created inside the directory and the root is fsynced (directory should persist). However, after a crash, the directory entry is missing even though data written to the old file was persisted.

System info
===========

Linux version 6.18.0-rc2, also tested on 6.14.11.

How to reproduce
================

```
#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>

int main() {
int status;
int file_fd;
int root_fd;

status = creat("file1", S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
printf("CREAT: %d\n", status);

// persist `file1`
sync();

status = mkdir("dir", S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
printf("MKDIR: %d\n", status);

status = open("file1", O_WRONLY | O_SYNC);
printf("OPEN: %d\n", status);
file_fd = status;

status = write(file_fd, "Test data!", 10);
printf("WRITE: %d\n", status);

status = link("file1", "dir/file2");
printf("LINK: %d\n", status);

status = open(".", O_RDONLY | O_DIRECTORY);
printf("OPEN: %d\n", status);
root_fd = status;

// persist `dir`
status = fsync(root_fd);
printf("FSYNC: %d\n", status);
}
```

Steps:

1. Create and mount new btrfs file system in default configuration.
2. Change directory to root of the file system and run the compiled test.
3. Cause hard system crash (e.g. QEMU system_reset command).
4. Remount file system after crash.
5. Observe that dir directory is missing.

Notes:

- ext4 does persist dir and dir/file2 even though it was not synced.
- xfs does persist dir but does not persist dir/file2.

P.S. Want to apologize for formatting in previous report, first time using Thunderbird and plain text.

#1 lore raw
Quoted 71 lines · Vyacheslav Kovalevsky

Under some circumstances, directory entry is not persisted after writing to the file inside the directory that was opened with O_SYNC flag if system crashes.

Detailed description
====================

Hello, we have found another issue with btrfs crash behavior.

In short, empty file is created and synced. Then, a new directory is created, old file is opened with O_SYNC flag and some data is written. After this, a new hard link is created inside the directory and the root is fsynced (directory should persist). However, after a crash, the directory entry is missing even though data written to the old file was persisted.

System info
===========

Linux version 6.18.0-rc2, also tested on 6.14.11.

How to reproduce
================

```
#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>

int main() {
int status;
int file_fd;
int root_fd;

status = creat("file1", S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
printf("CREAT: %d\n", status);

// persist `file1`
sync();

status = mkdir("dir", S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH);
printf("MKDIR: %d\n", status);

status = open("file1", O_WRONLY | O_SYNC);
printf("OPEN: %d\n", status);
file_fd = status;

status = write(file_fd, "Test data!", 10);
printf("WRITE: %d\n", status);

status = link("file1", "dir/file2");
printf("LINK: %d\n", status);

status = open(".", O_RDONLY | O_DIRECTORY);
printf("OPEN: %d\n", status);
root_fd = status;

// persist `dir`
status = fsync(root_fd);
printf("FSYNC: %d\n", status);
}
```
Quoted 7 lines

Steps:

1. Create and mount new btrfs file system in default configuration.
2. Change directory to root of the file system and run the compiled test.
3. Cause hard system crash (e.g. QEMU system_reset command).
4. Remount file system after crash.
5. Observe that dir directory is missing.

I converted that to a test case for fstests and couldn't reproduce, "dir", "file1" and "dir/file2" exist after the power failure.

The conversion for fstests:

#! /bin/bash
# SPDX-License-Identifier: GPL-2.0
# Copyright (c) 2025 SUSE S.A. All Rights Reserved. #
# FS QA Test 780
#

# what am I here for?
#
. ./common/preamble
_begin_fstest auto quick log

_cleanup()
{
_cleanup_flakey
cd /
rm -r -f $tmp.*
}

. ./common/filter
. ./common/dmflakey

_require_scratch
_require_dm_target flakey

rm -f $seqres.full

_scratch_mkfs >>$seqres.full 2>&1 || _fail "mkfs failed" _require_metadata_journaling $SCRATCH_DEV _init_flakey
_mount_flakey

touch $SCRATCH_MNT/file1

_scratch_sync

mkdir $SCRATCH_MNT/dir
echo -n "hello world" > $SCRATCH_MNT/file1
ln $SCRATCH_MNT/file1 $SCRATCH_MNT/dir/file2

$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/

# Simulate a power failure and then mount again the filesystem to replay the

# journal/log.
_flakey_drop_and_remount

ls -R $SCRATCH_MNT/ | _filter_scratch

_unmount_flakey

# success, all done
_exit 0

Quoted 13 lines

Notes:

- ext4 does persist dir and dir/file2 even though it was not synced.
- xfs does persist dir but does not persist dir/file2.

P.S. Want to apologize for formatting in previous report, first time using Thunderbird and plain text.

#2 lore raw
Quoted 50 lines · Filipe Manana

I converted that to a test case for fstests and couldn't reproduce, "dir", "file1" and "dir/file2" exist after the power failure.

The conversion for fstests:

#! /bin/bash
# SPDX-License-Identifier: GPL-2.0
# Copyright (c) 2025 SUSE S.A. All Rights Reserved. #
# FS QA Test 780
#

# what am I here for?
#
. ./common/preamble
_begin_fstest auto quick log

_cleanup()
{
_cleanup_flakey
cd /
rm -r -f $tmp.*
}

. ./common/filter
. ./common/dmflakey

_require_scratch
_require_dm_target flakey

rm -f $seqres.full
On 24/10/2025 19:17, Filipe Manana wrote:
_scratch_mkfs >>$seqres.full 2>&1 || _fail "mkfs failed" _require_metadata_journaling $SCRATCH_DEV _init_flakey
_mount_flakey

touch $SCRATCH_MNT/file1

_scratch_sync

mkdir $SCRATCH_MNT/dir
echo -n "hello world" > $SCRATCH_MNT/file1
ln $SCRATCH_MNT/file1 $SCRATCH_MNT/dir/file2

$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/

# Simulate a power failure and then mount again the filesystem to replay the

# journal/log.
_flakey_drop_and_remount
Quoted 6 lines

ls -R $SCRATCH_MNT/ | _filter_scratch

_unmount_flakey

# success, all done
_exit 0

I think the line with echo may not be the correct translation:

Quoted 1 line

echo -n "hello world" > $SCRATCH_MNT/file1

In the original test, the file was opened with O_SYNC flag, if you remove it, the directory will be there when the system crashes. I also forgot to close the file after the creat call in the original test, may be important as well.

The test itself is quite weird (why would dir be gone after seemingly unrelated operation?), any detail can matter.

Please run the original test with a real system crash. I will also double check everything on my side.

#3 lore raw
Quoted 8 lines · Vyacheslav Kovalevsky

I think the line with echo may not be the correct translation:

Quoted 1 line

echo -n "hello world" > $SCRATCH_MNT/file1

In the original test, the file was opened with O_SYNC flag, if you remove it, the directory will be there when the system crashes. I also forgot to close the file after the creat call in the original test, may be important as well.

Quoted 4 lines

The test itself is quite weird (why would dir be gone after seemingly unrelated operation?), any detail can matter.

Please run the original test with a real system crash.

This would produce hardware specific results rather than determining whether the file system is behaving correctly. It's possible the hardware is acknowledging the metadata, flush, super, flush, but then it's still not really persisting on disk

Signature · 1 line
Chris Murphy
#4 lore raw
In reply to Vyacheslav Kovalevsky
Quoted 52 lines · Vyacheslav Kovalevsky
Quoted 50 lines · Filipe Manana

I converted that to a test case for fstests and couldn't reproduce, "dir", "file1" and "dir/file2" exist after the power failure.

The conversion for fstests:

#! /bin/bash
# SPDX-License-Identifier: GPL-2.0
# Copyright (c) 2025 SUSE S.A. All Rights Reserved. #
# FS QA Test 780
#

# what am I here for?
#
. ./common/preamble
_begin_fstest auto quick log

_cleanup()
{
_cleanup_flakey
cd /
rm -r -f $tmp.*
}

. ./common/filter
. ./common/dmflakey

_require_scratch
_require_dm_target flakey

rm -f $seqres.full
On 24/10/2025 19:17, Filipe Manana wrote:
_scratch_mkfs >>$seqres.full 2>&1 || _fail "mkfs failed" _require_metadata_journaling $SCRATCH_DEV _init_flakey
_mount_flakey

touch $SCRATCH_MNT/file1

_scratch_sync

mkdir $SCRATCH_MNT/dir
echo -n "hello world" > $SCRATCH_MNT/file1
ln $SCRATCH_MNT/file1 $SCRATCH_MNT/dir/file2

$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/

# Simulate a power failure and then mount again the filesystem to replay the

# journal/log.
_flakey_drop_and_remount
Quoted 9 lines
Quoted 6 lines

ls -R $SCRATCH_MNT/ | _filter_scratch

_unmount_flakey

# success, all done
_exit 0

I think the line with echo may not be the correct translation:

Quoted 1 line

echo -n "hello world" > $SCRATCH_MNT/file1

An echo is just a write...

Quoted 5 lines

In the original test, the file was opened with O_SYNC flag, if you remove it, the directory will be there when the system crashes. I also forgot to close the file after the creat call in the original test, may be important as well.

An O_SYNC, which is what I missed before, is essentially just an implicit fsync after every write on a file.
Adding an fsync after the echo:

$XFS_IO_PROG -c "fsync" $SCRATCH_MNT/file1

Triggers the problem of "dir" not being persisted.

Quoted 3 lines

The test itself is quite weird (why would dir be gone after seemingly unrelated operation?), any detail can matter.

"dir" should be persisted as well as "dir/file2", according to the SOMC (Strictly Ordered Metadata Consistency) that Dave Chinner discussed many times in the past in fstests and btrfs mailing lists.

You should also reach the xfs mailing list and mention that "dir/file2" is not persisted.

Quoted 3 lines

Please run the original test with a real system crash. I will also double check everything on my side.

I've said before in another thread: we don't need to trigger qemu crashes in order to test fsync.
Just use the dm flakey target with fstests - no need to do reboots, much more practical and way less time consuming.
In 12 years of fixing fsync stuff on btrfs, I haven't yet seen any case where dm flakey didn't do the job of reproducing issues.

Quoted 1 line
#5 lore raw
Quoted 10 lines · Filipe Manana
Quoted 3 lines

The test itself is quite weird (why would dir be gone after seemingly unrelated operation?), any detail can matter.

"dir" should be persisted as well as "dir/file2", according to the SOMC (Strictly Ordered Metadata Consistency) that Dave Chinner discussed many times in the past in fstests and btrfs mailing lists.

You should also reach the xfs mailing list and mention that "dir/file2" is not persisted.

No. The fsync is on the root directory. So the only thing that is needed to per persisted is transactions touching that. The transaction to created dir is persisted because it is an entry in the root directly. creating and writing to file2 has nothing to do with the root directly and absolutely should not be effected by an fsync on an unrelated inode.