Re: [PATCH] fs: Additional checks on new and old dir

[Al Viro]

On Wed, May 14, 2025 at 06:39:40AM +0800, Edward Adam Davis wrote:
> In the reproducer, when calling renameat2(), olddirfd and newdirfd passed
> are the same value r0, see [1]. This situation should be avoided.
> 
> [1]
> renameat2(r0, &(0x7f0000000240)='./bus/file0\x00', r0, &(0x7f00000001c0)='./file0\x00', 0x0)
> 
> Reported-by: syzbot+321477fad98ea6dd35b7@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://syzkaller.appspot.com/bug?extid=321477fad98ea6dd35b7
> Tested-by: syzbot+321477fad98ea6dd35b7@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Edward Adam Davis <eadavis@xxxxxx>
> ---
>  fs/namei.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/namei.c b/fs/namei.c
> index 84a0e0b0111c..ff843007ca94 100644
> --- a/fs/namei.c
> +++ b/fs/namei.c
> @@ -5013,7 +5013,7 @@ int vfs_rename(struct renamedata *rd)
>  	struct name_snapshot old_name;
>  	bool lock_old_subdir, lock_new_subdir;
>  
> -	if (source == target)
> +	if (source == target || old_dir == target)
>  		return 0;

What the hell?

1) olddirfd and newdirfd have nothing to do with vfs_rename() - they are
bloody well gone by the time we get there.

2) there's nothing wrong with having the same value passed in both -
and it's certainly not a "quietly do nothing".

3) the check added in this patch is... odd.  You are checking essentically
for rename("foo/bar", "foo").  It should fail (-ENOTEMPTY or -EINVAL, depending
upon RENAME_EXCHANGE in flags) without having reached vfs_rename().