Re: [WTF?] random test in netlink_sendmsg()

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

On Fri, Dec 12, 2014 at 03:34:33PM -0500, David Miller wrote:
> > 	What is that check trying to do?  Is that simply missing
> > "(msg->msg_iovlen > 0) &&"?  And why on the Earth didn't it simply use
> > zero msg_iovlen as the indicator, instead of messing with iovec contents?
> > Obviously too late to change, but... ouch.
> 
> I think it's simply trying to say: if nothing in the given iovec, use
> the mmap() netlink area for the data.
> 
> I cannot vouch for the correctness of this test.
> 
> If we take the netlink_mmap_sendmsg() path, msg->msg_iov is not
> accessed at all, so it cannot be a huge problem.

Yes, but decision whether to take that path or not is random in case of
msg_iovlen being 0...

What do we want sendmsg(fd, &msg, 0) to do when fd is AF_NETLINK socket
that had setsockopt(fd, SOL_NETLINK, NETLINK_TX_RING, ...) successfully done
to it and msg.msg_iovlen is 0?  Userland ABI question - not a "do we have
a security hole there" one...  As it is, it might decide to do
netlink_mmap_sendmsg() or it might decide to act as if we hadn't done
NETLINK_TX_RING at all.  In _both_ cases it won't try to dereference
kernel-side ->msg_iov[0].iov_base - the former won't look at msg_iov at all,
the latter will ask to copy 0 bytes from it, which will succeed and do nothing
whatsoever.  Which path is taken depends on whether an uninitialized array
on stack frame of ___sys_sendmsg() happens to begin with NULL.

IMO that's bogus - behaviour ought to be at least deterministic...