Re: Netlink mmap tx security?

[Andy Lutomirski <luto@amacapital.net>]

On Tue, Oct 14, 2014 at 1:00 PM, David Miller <davem@davemloft.net> wrote:
> From: Andy Lutomirski <luto@amacapital.net>
> Date: Tue, 14 Oct 2014 12:33:43 -0700
>
>> For full honesty, there is now the machinery developed for memfd
>> sealing, but I can't imagine that this is ever faster than just
>> copying the buffer.
>
> I don't have much motivation to even check if it's a worthwhile
> pursuit at this point.
>
> Someone who wants to can :-)
>
>> I think that the NETLINK_SKB_TX declaration in include/linux/netlink.h
>> should probably go, too.  And there's the last parameter to
>> netlink_set_ring, too, and possibly even the nlk->tx_ring struct
>> itself.
>
> Agreed on all counts, here is the new patch:
>
> ====================
> [PATCH] netlink: Remove TX mmap support.
>
> There is no reasonable manner in which to absolutely make sure that another
> thread of control cannot write to the pages in the mmap()'d area and thus
> make sure that netlink messages do not change underneath us after we've
> performed verifications.
>
> Reported-by: Andy Lutomirski <luto@amacapital.net>
> Signed-off-by: David S. Miller <davem@davemloft.net>

Looks sensible to me, but I have no idea how to test it.

It's at least remotely possible that there's something that assumes
that assumes that the availability of NETLINK_RX_RING implies
NETLINK_TX_RING, which would be unfortunate.

--Andy