Re: [RFC PATCH net-next] sctp: fix src address selection if using secondary addresses

[Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>]

On Fri, Jul 10, 2015 at 03:27:02PM -0300, Marcelo Ricardo Leitner wrote:
> On Fri, Jul 10, 2015 at 01:14:21PM -0400, Vlad Yasevich wrote:
> > On 07/10/2015 12:17 PM, Marcelo Ricardo Leitner wrote:
> > > On Fri, Jul 10, 2015 at 11:35:28AM -0400, Vlad Yasevich wrote:
...
> > >> have been numerous times where I've seen weak host model in use on the wire
> > >> even with a BSD peer.
> > >>
> > >> This also puts a very big nail through many suggestions we've had over the years
> > >> to allow source based path multihoming in addition to destination based multihoming
> > >> we currently support.
> > >>
> > >> It might be a good idea to make rp-filter like behavior best effort, and have
> > >> the old behavior as fallback.  I am still trying to think up different scenarios
> > >> where rp-filter behavior will cause things to fail prematurely...
> > > 
> > > The old behavior is like "if we don't have a src yet and can't find a
> > > preferred src for this dst, use the 1st bound address". We can add it
> > > but as I said, I'm afraid it is just doing wrong and not worth. If such
> > > randomly src addressed packet is meant to be routed, the router will
> > > likely drop it as it is seen as a spoof. And if it reaches the peer, it
> > > will probably come back through a different path.
> > > 
> > > I'm tempted to say that current usual use cases are handled by the first
> > > check on this function, which returns the preferred/primary address for
> > > the interface and checks against bound addresses. Whenever you reach the
> > > second check, it just allows you to use that 1st bound address that is
> > > checked. I mean, I can't see use cases that we would be breaking with
> > > this change. 
> > 
> > Yes,  the secondary check didn't amount to much, but we've kept it since 2.5
> > days (when sctp was introduced).  I've made attempts over the years to
> > try to make it stricter, but that never amounted to anything that worked well.
> > 
> > > 
> > > But yeah, it impacts source based routing, and I'm not aware of previous
> > > discussions on it. I'll try to dig some up but if possible, please share
> > > some pointers on it.
> > 
> > It's been suggested a few times that we should support source based multihoming
> > particularly for the case where one peer has only 1 address.
> > We've always punted on this, but people still ask every now and then.
> 
> Ah okay, now I see it.
>  
> > I do have a question about the code though.. Have you tried with mutlipath routing
> > enabled.  I see rp_filter checks have special code to handle that.  Seem like we
> > might get false negatives in sctp.
> 
> In the sense of CONFIG_IP_ROUTE_MULTIPATH=y, yes, but just that. My
> routes were simple ones, either 2 peers attaches to 2 local subnets, or
> with a gateway in the middle (with 2 subnets on each side, but mapped
> 1-1, no crossing. Aka subnet1<->subnet2 and subnet3<->subnet4 while not
> (subnet1<->subnet4 or subnet3<->subnet2).
> 
> Note that this is not rp_filter strictly speaking, as it's mirrored.
> rp_filter needs to calculate all possible output routes (actually until
> it finds a valid one) for finding one that would match the one used for
> incoming. 
> 
> This check already has an output path, and it's calculating if such
> input would be acceptable. We can't really expect/check for other hits
> because it invalidates the chosen output path.
> 
> Hmmm... but we could support multipath in the output selection, ie in
> the outputs of ip_route_output_key(), probably in another patch then?

Thinking further.. we could just compare it with the addresses assigned to the
interface instead of doing a whole new routing. Cheaper/faster, provides the
results I'm looking for and the consequences are easier to see.

Something like (not tested, just illustrating the idea):

--- a/net/sctp/protocol.c
+++ b/net/sctp/protocol.c
@@ -489,22 +489,33 @@ static void sctp_v4_get_dst(struct sctp_transport *t, union sctp_addr *saddr,
        list_for_each_entry_rcu(laddr, &bp->address_list, list) {
                if (!laddr->valid)
                        continue;
                if ((laddr->state == SCTP_ADDR_SRC) &&
                    (AF_INET == laddr->a.sa.sa_family)) {
+                       struct net_device *odev;
+
                        fl4->fl4_sport = laddr->a.v4.sin_port;
                        flowi4_update_output(fl4,
                                             asoc->base.sk->sk_bound_dev_if,
                                             RT_CONN_FLAGS(asoc->base.sk),
                                             daddr->v4.sin_addr.s_addr,
                                             laddr->a.v4.sin_addr.s_addr);
 
                        rt = ip_route_output_key(sock_net(sk), fl4);
-                       if (!IS_ERR(rt)) {
-                               dst = &rt->dst;
-                               goto out_unlock;
-                       }
+                       if (IS_ERR(rt))
+                               continue;
+
+                       /* Ensure the src address belongs to the output
+                        * interface.
+                        */
+                       odev = __ip_dev_find(net, laddr->a.v4.sin_addr.s_addr,
+                                            false);
+                       if (odev->if_index != fl4->flowi4_oif)
+                               continue;
+
+                       dst = &rt->dst;
+                       goto out_unlock;
                }
        }
 
 out_unlock:
        rcu_read_unlock();


I like this better than my 1st attempt. What do you think?

I'll split the refactoring from this fix on v2, so it's easier to review.

  Marcelo