On Wed, 22 Sept 2021 at 15:21, Chengguang Xu [off-list ref] wrote:

在 2021/9/22 16:24, Huang Jianan 写道:

quoted

在 2021/9/22 16:06, Chengguang Xu 写道:

quoted

在 2021/9/22 15:23, Huang Jianan 写道:

quoted

From: Huang Jianan <redacted>

At present, overlayfs provides overlayfs inode to users. Overlayfs
inode provides ovl_aops with noop_direct_IO to avoid open failure
with O_DIRECT. But some compressed filesystems, such as erofs and
squashfs, don't support direct_IO.

Users who use f_mapping->a_ops->direct_IO to check O_DIRECT support,
will read file through this way. This will cause overlayfs to access
a non-existent direct_IO function and cause panic due to null pointer:

I just looked around the code more closely, in open_with_fake_path(),

do_dentry_open() has already checked O_DIRECT open flag and
a_ops->direct_IO of underlying real address_space.

Am I missing something?

It seems that loop_update_dio will set lo->use_dio after open file
without set O_DIRECT.
loop_update_dio will check f_mapping->a_ops->direct_IO but it deal
with ovl_aops with
noop _direct_IO.

So I think we still need a new aops?

It means we should only set ->direct_IO for overlayfs inodes whose
underlying fs has DIRECT IO ability.

First let's fix the oops: ovl_read_iter()/ovl_write_iter() must check
real file's ->direct_IO if IOCB_DIRECT is set in iocb->ki_flags and
return -EINVAL if not.

To fix the loop -> overlay -> squashfs case your suggestion of having
separate aops depending on the real inode's ->direct_IO sounds good.

Thanks,
Miklos