Kernel memory corruption in CIPSO labeled TCP packets processing.

[Paul Moore]

On Tue, Jan 15, 2019 at 2:52 PM Paul Moore <paul at paul-moore.com> wrote:
> On Tue, Jan 15, 2019 at 12:55 PM Casey Schaufler <casey at schaufler-ca.com> wrote:
> > On 1/15/2019 9:06 AM, Nazarov Sergey wrote:
> > > Hello!
> > > Security modules (selinux, smack) use icmp_send for discarded incorrectly labeled network packets.
> > > This could be on TCP level too (security_sock_rcv_skb -> cipso_v4_error for INET stream connection, for example).
> > > icmp_send calls ip_option_echo, which uses IPCB to take compiled IP options.
> > > After moving IP header data to the end of the struct tcp_skb_cb (since 3.18 kernel), this could lead to
> > > kernel memory corruption when IP options copying.
> >
> > Can you explain how that corruption might occur?
> > Do you have a test case?
>
> Thanks for pointing this out Nazarov.
>
> Presumably we are going to hit a problem whenever icmp_send is called
> from outside the IP layer in the stack.  We fixed a similar problem a
> few years back with 04f81f0154e4 ("cipso: don't use IPCB() to locate
> the CIPSO IP option").
>
> I've CC'd netdev, as I'm guessing they will have some thoughts on
> this, but my initial reaction is that your proposed patch isn't as
> generic as it should be for code that lives in icmp_send().  I suspect
> the safe thing to do would be to call ip_options_compile() again on
> skb_in and build a local copy of the ip_options struct that could then
> be used in the call to __ip_options_echo(); the code could either live
> in icmp_send() or some new ip_options_echo() variant
> (ip_options_echo_safe()?  I dunno).  Unfortunately, calling
> ip_options_compile() is going to add some overhead, and may be a
> non-starter for the netdev folks, but this is error path code, so it
> might be acceptable.  Hopefully the netdev folks will have some
> better, more clever suggestions.

It's been a few days now with no comments from the netdev folks, so I
think it's probably best to start putting together a RFC patch for
review/comment.  Nazarov, would you like to do that?  If not, that's
okay, just let me know.

I'm still concerned about calling ip_options_compile() in icmp_send()
and I'm thinking we might be better off to add a new ip_options
parameter to icmp_send(); if the parameter is NULL we behave as we do
today, but if it is non-NULL we use the given ip_options parameter in
place of calling ip_options_echo().  With that change in place, we
would need to update cipso_v4_error() to do the extra work of calling
ip_options_compile() and __ip_options_echo().  There looks to be maybe
a dozen (or two?) existing icmp_send() callers, but it should be
pretty trivial to update them to pass NULL for the new parameter.

What do you think?

> > > This patch fix a bug, but I'm not sure, that this is a best solution. Perhaps someone more familiar with the
> > > linux TCP/IP stack will offer a better one.
> > > Thanks.
> > >
> > > --- a/net/ipv4/icmp.c
> > > +++ b/net/ipv4/icmp.c
> > > @@ -679,7 +679,8 @@ void icmp_send(struct sk_buff *skb_in, i
> > >                                         iph->tos;
> > >       mark = IP4_REPLY_MARK(net, skb_in->mark);
> > >
> > > -     if (ip_options_echo(&icmp_param->replyopts.opt.opt, skb_in))
> > > +     if (__ip_options_echo(&icmp_param->replyopts.opt.opt, skb_in,
> > > +                     ip_hdr(skb_in)->protocol == IPPROTO_TCP ? &TCP_SKB_CB(skb_in)->header.h4.opt : &IPCB(skb_in)->opt))
> > >               goto out_unlock;
> > >
> > >

-- 
paul moore
www.paul-moore.com