[PATCH v3 7/8] exfat: complete direct writes synchronously when they update metadata

[Chi Zhiling <chizhiling@163.com>]

From: Chi Zhiling <chizhiling@kylinos.cn>

The inode metadata for a direct write is updated in ->end_io().  That
handler runs without holding i_rwsem whenever the I/O completes
asynchronously.  As a result, the i_size update can race with other
operations.

Whenever a write extends beyond the current valid_size, that is,
whenever ->end_io() will have metadata to update, pass
IOMAP_DIO_FORCE_WAIT to iomap_dio_rw().  This guarantees that the
metadata update happens under the inode lock held across the write.

Note that this forces extending O_DIRECT writes, the common append case
for AIO and io_uring, to complete synchronously.  Pure overwrites still
complete asynchronously because ->end_io() has nothing to publish for
them.

Fixes: 867b9c96dc83 ("exfat: add iomap direct I/O support")
Signed-off-by: Chi Zhiling <chizhiling@kylinos.cn>
---
 fs/exfat/file.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/fs/exfat/file.c b/fs/exfat/file.c
index 23b8a36acaf7..27050f0358ee 100644
--- a/fs/exfat/file.c
+++ b/fs/exfat/file.c
@@ -857,10 +857,16 @@ static ssize_t exfat_fallback_buffered_write(struct kiocb *iocb,
 
 static ssize_t exfat_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 {
+	struct inode *inode = file_inode(iocb->ki_filp);
+	loff_t end = iocb->ki_pos + iov_iter_count(from);
+	unsigned int dio_flags = 0;
 	ssize_t ret;
 
+	if (end > exfat_get_valid_size(EXFAT_I(inode)))
+		dio_flags |= IOMAP_DIO_FORCE_WAIT;
+
 	ret = iomap_dio_rw(iocb, from, &exfat_write_iomap_ops,
-			&exfat_write_dio_ops, 0, NULL, 0);
+			&exfat_write_dio_ops, dio_flags, NULL, 0);
 	if (ret == -ENOTBLK)
 		ret = 0;
 	else if (ret < 0)
-- 
2.53.0