[PATCH 0/3] fs: drain in-flight DIO before buffered write fallback

[PATCH 0/3] fs: drain in-flight DIO before buffered write fallback

[Jiale Yao]

An asynchronous direct write can remain in flight after its submitting
thread releases the inode lock.  If another direct write falls back to
buffered I/O in the meantime, the buffered write can dirty page cache
before the first write completes post-I/O invalidation.  The invalidation
then finds dirty pages, emits a page cache invalidation failure warning,
and records -EIO in the mapping error sequence.  A later fsync() returns
-EIO.

Commit 15cdefd0c0522f9d5e12d947fa04f4c11649b699 ("ext4: drain
in-flight DIO before buffered write fallback") fixed this race in ext4.
The same ordering is missing from the buffered fallback paths in ext2,
NTFS, and exFAT.

This series adds inode_dio_wait() before each fallback dirties page cache.
Each patch fixes one filesystem and remains independently buildable.

A reproducer using concurrent AIO direct writes and buffered fallback
triggered the following warning on all three filesystems and made a
subsequent fsync() return -EIO:

  Page cache invalidation failure on direct I/O.  Possible data corruption
  due to collision with buffered I/O!

Jiale Yao (3):
  ext2: drain in-flight DIO before buffered write fallback
  ntfs: drain in-flight DIO before buffered write fallback
  exfat: drain in-flight DIO before buffered write fallback

 fs/exfat/file.c | 6 ++++++
 fs/ext2/file.c  | 7 +++++++
 fs/ntfs/file.c  | 7 +++++++
 3 files changed, 20 insertions(+)

-- 
2.34.1

[PATCH 1/3] ext2: drain in-flight DIO before buffered write fallback

[Jiale Yao]

An asynchronous direct write can remain in flight after the inode lock is
released.  If another direct write falls back to buffered I/O while the
first write is still pending, generic_perform_write() can dirty pages
before the first write completes its post-I/O page cache invalidation.
The invalidation then finds dirty pages, reports a page cache invalidation
failure, and records -EIO in the mapping error sequence.  A later fsync()
therefore returns -EIO.

Commit 15cdefd0c0522f9d5e12d947fa04f4c11649b699 ("ext4: drain
in-flight DIO before buffered write fallback") fixed the same race in
ext4.  Ext2 has an equivalent fallback after iomap_dio_rw() returns
-ENOTBLK or a short write, but does not drain other in-flight DIO before
dirtying the page cache.

Wait for in-flight DIO before calling generic_perform_write() in the
fallback path.

A reproducer using concurrent AIO direct writes and buffered fallback
triggered the following warning and made a subsequent fsync() return
-EIO:

  Page cache invalidation failure on direct I/O.  Possible data corruption
  due to collision with buffered I/O!

Fixes: fb5de4358e1a ("ext2: Move direct-io to use iomap")
Link: https://lore.kernel.org/r/20260629113827.4074335-3-libaokun@linux.alibaba.com
Signed-off-by: Jiale Yao <yaojiale02@163.com>
---
 fs/ext2/file.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/ext2/file.c b/fs/ext2/file.c
index b9020df7d89e..67fe423c3828 100644
--- a/fs/ext2/file.c
+++ b/fs/ext2/file.c
@@ -135,6 +135,13 @@ static ssize_t ext2_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 		int ret2;
 
 		iocb->ki_flags &= ~IOCB_DIRECT;
+
+		/*
+		 * Prevent concurrent direct I/O and buffered I/O to the same file
+		 * range. Wait for in-flight DIO to finish before dirtying pages.
+		 */
+		inode_dio_wait(inode);
+
 		pos = iocb->ki_pos;
 		status = generic_perform_write(iocb, from);
 		if (unlikely(status < 0)) {
-- 
2.34.1

[PATCH 2/3] ntfs: drain in-flight DIO before buffered write fallback

[Jiale Yao]

An asynchronous direct write can remain in flight after the inode lock is
released.  If another direct write falls back to buffered I/O while the
first write is still pending, iomap_file_buffered_write() can dirty pages
before the first write completes its post-I/O page cache invalidation.
The invalidation then finds dirty pages, reports a page cache invalidation
failure, and records -EIO in the mapping error sequence.  A later fsync()
therefore returns -EIO.

Commit 15cdefd0c0522f9d5e12d947fa04f4c11649b699 ("ext4: drain
in-flight DIO before buffered write fallback") fixed the same race in
ext4.  NTFS has an equivalent fallback after iomap_dio_rw() returns
-ENOTBLK or a short write, but does not drain other in-flight DIO before
dirtying the page cache.

Wait for in-flight DIO before calling iomap_file_buffered_write() in the
fallback path.

A reproducer using concurrent AIO direct writes and buffered fallback
triggered the following warning and made a subsequent fsync() return
-EIO:

  Page cache invalidation failure on direct I/O.  Possible data corruption
  due to collision with buffered I/O!

Fixes: 9c87959601e8 ("ntfs: update file operations")
Link: https://lore.kernel.org/r/20260629113827.4074335-3-libaokun@linux.alibaba.com
Signed-off-by: Jiale Yao <yaojiale02@163.com>
---
 fs/ntfs/file.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/ntfs/file.c b/fs/ntfs/file.c
index 007d1614b9ac..2fc2ffde3846 100644
--- a/fs/ntfs/file.c
+++ b/fs/ntfs/file.c
@@ -527,6 +527,13 @@ static ssize_t ntfs_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 
 		offset = iocb->ki_pos;
 		iocb->ki_flags &= ~IOCB_DIRECT;
+
+		/*
+		 * Prevent concurrent direct I/O and buffered I/O to the same file
+		 * range. Wait for in-flight DIO to finish before dirtying pages.
+		 */
+		inode_dio_wait(file_inode(iocb->ki_filp));
+
 		written = iomap_file_buffered_write(iocb, from,
 				&ntfs_write_iomap_ops, &ntfs_iomap_folio_ops,
 				NULL);
-- 
2.34.1

Re: [PATCH 2/3] ntfs: drain in-flight DIO before buffered write fallback

[liubaolin]

在 2026/9/24 19:05, Jiale Yao 写道:
> An asynchronous direct write can remain in flight after the inode lock is
> released.  If another direct write falls back to buffered I/O while the
> first write is still pending, iomap_file_buffered_write() can dirty pages
> before the first write completes its post-I/O page cache invalidation.
> The invalidation then finds dirty pages, reports a page cache invalidation
> failure, and records -EIO in the mapping error sequence.  A later fsync()
> therefore returns -EIO.
> 
> Commit 15cdefd0c0522f9d5e12d947fa04f4c11649b699 ("ext4: drain
> in-flight DIO before buffered write fallback") fixed the same race in
> ext4.  NTFS has an equivalent fallback after iomap_dio_rw() returns
> -ENOTBLK or a short write, but does not drain other in-flight DIO before
> dirtying the page cache.
> 
> Wait for in-flight DIO before calling iomap_file_buffered_write() in the
> fallback path.
> 
> A reproducer using concurrent AIO direct writes and buffered fallback
> triggered the following warning and made a subsequent fsync() return
> -EIO:
> 
>    Page cache invalidation failure on direct I/O.  Possible data corruption
>    due to collision with buffered I/O!
> 
> Fixes: 9c87959601e8 ("ntfs: update file operations")
> Link: https://lore.kernel.org/r/20260629113827.4074335-3-libaokun@linux.alibaba.com
> Signed-off-by: Jiale Yao <yaojiale02@163.com>
> ---
>   fs/ntfs/file.c | 7 +++++++
>   1 file changed, 7 insertions(+)
> 
> diff --git a/fs/ntfs/file.c b/fs/ntfs/file.c
> index 007d1614b9ac..2fc2ffde3846 100644
> --- a/fs/ntfs/file.c
> +++ b/fs/ntfs/file.c
> @@ -527,6 +527,13 @@ static ssize_t ntfs_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
>   
>   		offset = iocb->ki_pos;
>   		iocb->ki_flags &= ~IOCB_DIRECT;
> +
> +		/*
> +		 * Prevent concurrent direct I/O and buffered I/O to the same file
> +		 * range. Wait for in-flight DIO to finish before dirtying pages.
> +		 */
> +		inode_dio_wait(file_inode(iocb->ki_filp));

Hi Jiale,
    This fallback path can still have IOCB_NOWAIT set, but 
inode_dio_wait() blocks when there is outstanding DIO. Clearing 
IOCB_DIRECT does not clear IOCB_NOWAIT.

   In comparison, ext4_buffered_write_iter() rejects NOWAIT requests 
before reaching the wait.

   Could we also handle IOCB_NOWAIT before entering this potentially 
blocking fallback?
   If no bytes have been written, we can return -EAGAIN; if some DIO has 
already completed, we should preserve the positive short-write result.

Thanks,
Baolin.


> +
>   		written = iomap_file_buffered_write(iocb, from,
>   				&ntfs_write_iomap_ops, &ntfs_iomap_folio_ops,
>   				NULL);

Re:Re: [PATCH 2/3] ntfs: drain in-flight DIO before buffered write fallback

[jiale yao]

Hi Baolin,

You're right. NTFS advertises FMODE_NOWAIT, and clearing IOCB_DIRECT
does not make it valid to block in inode_dio_wait().

I will check IOCB_NOWAIT before entering the buffered fallback. If
iomap_dio_rw() has not written any data, the path will return -EAGAIN.
If it completed a partial write, it will return the positive short-write
result without entering the potentially blocking fallback.

I will address this in v2, after 24h.

Thanks,
Jiale

[PATCH 3/3] exfat: drain in-flight DIO before buffered write fallback

[Jiale Yao]

An asynchronous direct write can remain in flight after the inode lock is
released.  If another direct write falls back to buffered I/O while the
first write is still pending, iomap_file_buffered_write() can dirty pages
before the first write completes its post-I/O page cache invalidation.
The invalidation then finds dirty pages, reports a page cache invalidation
failure, and records -EIO in the mapping error sequence.  A later fsync()
therefore returns -EIO.

Commit 15cdefd0c0522f9d5e12d947fa04f4c11649b699 ("ext4: drain
in-flight DIO before buffered write fallback") fixed the same race in
ext4.  ExFAT has an equivalent fallback after iomap_dio_rw() returns
-ENOTBLK or a short write, but does not drain other in-flight DIO before
dirtying the page cache.

Wait for in-flight DIO before calling iomap_file_buffered_write() in the
fallback path.

A reproducer using concurrent AIO direct writes and buffered fallback
triggered the following warning and made a subsequent fsync() return
-EIO:

  Page cache invalidation failure on direct I/O.  Possible data corruption
  due to collision with buffered I/O!

Fixes: 867b9c96dc83 ("exfat: add iomap direct I/O support")
Link: https://lore.kernel.org/r/20260629113827.4074335-3-libaokun@linux.alibaba.com
Signed-off-by: Jiale Yao <yaojiale02@163.com>
---
 fs/exfat/file.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/fs/exfat/file.c b/fs/exfat/file.c
index a2a9ee1a2004..cf5ccbd54823 100644
--- a/fs/exfat/file.c
+++ b/fs/exfat/file.c
@@ -807,6 +807,12 @@ static ssize_t exfat_fallback_buffered_write(struct kiocb *iocb,
 
 	iocb->ki_flags &= ~IOCB_DIRECT;
 
+	/*
+	 * Prevent concurrent direct I/O and buffered I/O to the same file
+	 * range. Wait for in-flight DIO to finish before dirtying pages.
+	 */
+	inode_dio_wait(file_inode(iocb->ki_filp));
+
 	written = iomap_file_buffered_write(iocb, from, &exfat_write_iomap_ops,
 			NULL, NULL);
 	if (written < 0)
-- 
2.34.1

[PATCH 2/3] ntfs: drain in-flight DIO before buffered write fallback

[Jiale Yao]

An asynchronous direct write can remain in flight after the inode lock is
released.  If another direct write falls back to buffered I/O while the
first write is still pending, iomap_file_buffered_write() can dirty pages
before the first write completes its post-I/O page cache invalidation.
The invalidation then finds dirty pages, reports a page cache invalidation
failure, and records -EIO in the mapping error sequence.  A later fsync()
therefore returns -EIO.

Commit 15cdefd0c0522f9d5e12d947fa04f4c11649b699 ("ext4: drain
in-flight DIO before buffered write fallback") fixed the same race in
ext4.  NTFS has an equivalent fallback after iomap_dio_rw() returns
-ENOTBLK or a short write, but does not drain other in-flight DIO before
dirtying the page cache.

Wait for in-flight DIO before calling iomap_file_buffered_write() in the
fallback path.

A reproducer using concurrent AIO direct writes and buffered fallback
triggered the following warning and made a subsequent fsync() return
-EIO:

  Page cache invalidation failure on direct I/O.  Possible data corruption
  due to collision with buffered I/O!

Fixes: 9c87959601e8 ("ntfs: update file operations")
Link: https://lore.kernel.org/r/20260629113827.4074335-3-libaokun@linux.alibaba.com
Signed-off-by: Jiale Yao <yaojiale02@163.com>
---
 fs/ntfs/file.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/ntfs/file.c b/fs/ntfs/file.c
index 007d1614b9ac..2fc2ffde3846 100644
--- a/fs/ntfs/file.c
+++ b/fs/ntfs/file.c
@@ -527,6 +527,13 @@ static ssize_t ntfs_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 
 		offset = iocb->ki_pos;
 		iocb->ki_flags &= ~IOCB_DIRECT;
+
+		/*
+		 * Prevent concurrent direct I/O and buffered I/O to the same file
+		 * range. Wait for in-flight DIO to finish before dirtying pages.
+		 */
+		inode_dio_wait(file_inode(iocb->ki_filp));
+
 		written = iomap_file_buffered_write(iocb, from,
 				&ntfs_write_iomap_ops, &ntfs_iomap_folio_ops,
 				NULL);
-- 
2.34.1