A missing check bug in __sys_accept4_file()

[Paul Moore]

On Wed, May 12, 2021 at 3:43 AM Jinmeng Zhou <jjjinmeng.zhou at gmail.com> wrote:
> Dear maintainers,
>
> hi, our team has found and reported a missing check bug on Linux
> kernel v5.10.7 using static analysis.
> We are looking forward to having more experts' eyes on this. Thank you!

Creating a new socket, not associated with a connection (e.g. via
sock_create_lite()), is a different operation than creating a new
socket in response to an incoming connection as is done in
__sys_accept4_file().  This is why the sock_create_lite() uses the
security_socket_create() LSM hook and __sys_accept4_file() uses the
security_socket_accept() LSM hook.

> > On Fri, May 7, 2021 at 1:59 AM Jakub Kicinski <kuba at kernel.org> wrote:
> > > On Thu, 6 May 2021 15:44:36 +0800 Jinmeng Zhou wrote:
> > > hi, our team has found a missing check bug on Linux kernel v5.10.7 using static analysis. There is a path calls sock_alloc() after checking LSM function security_socket_create(), while another path calls it without checking.
> > > We think there is a missing check bug in __sys_accept4_file() before calling sock_alloc().
> >
> > Perhaps the semantics for listening sockets is that only the parent
> > sockets get the LSM check. Could you please circulate the report more
> > widely? I'd be good to have LSM experts' eyes on this at least.
> > CCing the mailing list should help get more opinions. Thank you!
> >
> > > Function sock_create_lite() uses security_socket_create() to check.
> > > 1.    // check security_socket_create() ///////////////////////
> > > 2.    int sock_create_lite(int family, int type, int protocol, struct socket **res)
> > > 3.    {
> > > 4.      int err;
> > > 5.      struct socket *sock = NULL;
> > > 6.      err = security_socket_create(family, type, protocol, 1);
> > > 7.      if (err)
> > > 8.        goto out;
> > > 9.      sock = sock_alloc();
> > > 10.   ...
> > > 11.   }
> > >
> > > However, __sys_accept4_file() directly calls sock_alloc() without the security check.
> > > 1.    // no check ////////////////////////////////////
> > > 2.    int __sys_accept4_file(struct file *file, unsigned file_flags,
> > > 3.          struct sockaddr __user *upeer_sockaddr,
> > > 4.          int __user *upeer_addrlen, int flags,
> > > 5.          unsigned long nofile)
> > > 6.    {
> > > 7.      struct socket *sock, *newsock;
> > > 8.      struct file *newfile;
> > > 9.      int err, len, newfd;
> > > 10.     struct sockaddr_storage address;
> > > 11.     if (flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
> > > 12.       return -EINVAL;
> > > 13.     if (SOCK_NONBLOCK != O_NONBLOCK && (flags & SOCK_NONBLOCK))
> > > 14.       flags = (flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
> > > 15.     sock = sock_from_file(file, &err);
> > > 16.     if (!sock)
> > > 17.       goto out;
> > > 18.     err = -ENFILE;
> > > 19.     newsock = sock_alloc();
> > > 20.   ...
> > > 21.   }
> > >
> > > This no-check function can be reached through syscall.
> > > syscall => __sys_accept4 => __sys_accept4_file
> > >
> > > SYSCALL_DEFINE4(accept4, int, fd, struct sockaddr __user *, upeer_sockaddr,
> > > int __user *, upeer_addrlen, int, flags)
> > > {
> > > return __sys_accept4(fd, upeer_sockaddr, upeer_addrlen, flags);
> > > }

-- 
paul moore
www.paul-moore.com