Re: Information leak in llcp_sock_bind/llcp_raw_sock_bind

On Tue, Dec 15, 2015 at 9:48 PM, David Miller [off-list ref] wrote:

From: Dmitry Vyukov <dvyukov@google.com>
Date: Tue, 15 Dec 2015 21:45:16 +0100

quoted

On Tue, Dec 15, 2015 at 9:36 PM, David Miller [off-list ref] wrote:

quoted

From: Dmitry Vyukov <dvyukov@google.com>
Date: Tue, 15 Dec 2015 21:00:20 +0100

quoted

The problem is that llcp_sock_bind/llcp_raw_sock_bind do not check
sockaddr_len passed in, so they copy stack garbage from stack into the
socket and then return it in getsockname.
This can defeat ASLR, leak crypto keys, etc.

That's actually the first thing these functions do.

They completely clear out the on-stack llcp_addr, then they copy only
as much as the user gave them, being careful not to use more than
sizeof(llcp_addr).

        memset(&llcp_addr, 0, sizeof(llcp_addr));
        len = min_t(unsigned int, sizeof(llcp_addr), alen);
        memcpy(&llcp_addr, addr, len);

I don't see what the problem is, you'll need to be more specific.

You are right. Sorry.

There still seems to be a minor leak here:

  if (!addr || addr->sa_family != AF_NFC)
      return -EINVAL;

addr->sa_family can be uninit.

That shouldn't matter at all, that can't cause socket state corruption.

I want to ask you if you are actually seeing kernel stack in that hexdump
you are posting?  If so, how do you actually account for it?  Nothing you
have shown so far make that clear.

I've seen a kernel address at least in pptp_bind, it was a return pc
in SyS_socket call that was executed just before bind.
Exact contents of the leaked info depend on kernel config, compiler
and a previous executed syscall (there are thousands of them if we
count ioctls and friends). So it is almost impossible to prove that a
PC cannot be leaked.