On Sun, Jun 10, 2018 at 7:31 AM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > On Tue, May 29, 2018 at 04:43:23PM +0200, Miklos Szeredi wrote: >> Signed-off-by: Miklos Szeredi <mszeredi@xxxxxxxxxx> >> --- >> fs/overlayfs/file.c | 3 +++ >> 1 file changed, 3 insertions(+) >> >> diff --git a/fs/overlayfs/file.c b/fs/overlayfs/file.c >> index 3f610a5b38e4..e5e7ccaaf9ec 100644 >> --- a/fs/overlayfs/file.c >> +++ b/fs/overlayfs/file.c >> @@ -110,6 +110,9 @@ static int ovl_open(struct inode *inode, struct file *file) >> if (IS_ERR(realfile)) >> return PTR_ERR(realfile); >> >> + /* For O_DIRECT dentry_open() checks f_mapping->a_ops->direct_IO */ >> + file->f_mapping = realfile->f_mapping; > > Umm... What happens if upper layer doesn't allow O_DIRECT, while the lower one does? Will get EINVAL on read(2) after copy up. Not sure if it can be called a regression, since it's a corner case of a corner case. I think proper solution is to support O_DIRECT unconditionally on upper (and for the likes of shmfs, just fall back to "cached" I/O). Thanks, Miklos