Re: [PATCH] commoncap: check return value to avoid null pointer dereference

[Al Viro <viro@zeniv.linux.org.uk>]

On Mon, May 16, 2022 at 10:40:02AM -0700, Yongzhi Liu wrote:
> The pointer inode is dereferenced before a null pointer
> check on inode, hence if inode is actually null we will
> get a null pointer dereference. Fix this by only dereferencing
> inode after the null pointer check on inode.
> 
> Fixes: c6f493d631c ("VFS: security/: d_backing_inode() annotations")
> Fixes: 8db6c34 ("Introduce v3 namespaced file capabilities")
> 
> Signed-off-by: Yongzhi Liu <lyz_cs@pku.edu.cn>
> ---
>  security/commoncap.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/security/commoncap.c b/security/commoncap.c
> index 5fc8986..978f011 100644
> --- a/security/commoncap.c
> +++ b/security/commoncap.c
> @@ -298,6 +298,8 @@ int cap_inode_need_killpriv(struct dentry *dentry)
>  	struct inode *inode = d_backing_inode(dentry);
>  	int error;
>  
> +	if (!inode)
> +		return 0;
>  	error = __vfs_getxattr(dentry, inode, XATTR_NAME_CAPS, NULL, 0);
>  	return error > 0;
>  }


	Huh?  AFAICS, that has all of two callers - notify_change() and
dentry_needs_remove_privs().  Both of them should never be called on
negative dentries and both would have blown up well before the call of
security_inode_need_killpriv().

> @@ -545,11 +547,13 @@ int cap_convert_nscap(struct user_namespace *mnt_userns, struct dentry *dentry,
>  	const struct vfs_cap_data *cap = *ivalue;
>  	__u32 magic, nsmagic;
>  	struct inode *inode = d_backing_inode(dentry);
> -	struct user_namespace *task_ns = current_user_ns(),
> -		*fs_ns = inode->i_sb->s_user_ns;
> +	struct user_namespace *task_ns = current_user_ns(), *fs_ns;
>  	kuid_t rootid;
>  	size_t newsize;
>  
> +	if (!inode)
> +		return -EINVAL;
> +	fs_ns = inode->i_sb->s_user_ns;
>  	if (!*ivalue)
>  		return -EINVAL;
>  	if (!validheader(size, cap))

... and that one come from vfs_setxattr().  Again, calling that on a negative
dentry is an oopsable bug.  If you have any reproducers, please post the stack
traces so it would be possible to deal with the underlying problem.