On Mon, 2009-07-27 at 14:44 -0700, Andrew Morton wrote:

(cc netdev)

On Thu, 23 Jul 2009 14:39:28 -0700
Chia-chi Yeh (_________) [off-list ref] wrote:

quoted

In IPv4, sending a packet with ipi_ifindex = 0 via an interface-bound
socket will unbind the packet to that interface. This behavior is
different from IPv6 which treats ipi6_ifindex = 0 as unspecified.
Furthermore, IPv6 does not allow sending packets to other interface
than the bound one, but I am not sure if it is necessary in IPv4. The
following patch keeps the interface binding when ipi_ifindex = 0.

Thanks,
Chia-chi

--- a/net/ipv4/ip_sockglue.c	2009-07-24 04:50:45.000000000 +0800
+++ b/net/ipv4/ip_sockglue.c	2009-07-24 04:51:09.000000000 +0800

@@ -213,7 +213,8 @@
 			if (cmsg->cmsg_len != CMSG_LEN(sizeof(struct in_pktinfo)))
 				return -EINVAL;
 			info = (struct in_pktinfo *)CMSG_DATA(cmsg);
-			ipc->oif = info->ipi_ifindex;
+			if (info->ipi_ifindex)
+				ipc->oif = info->ipi_ifindex;
 			ipc->addr = info->ipi_spec_dst.s_addr;
 			break;
 		}

A convenient copy of the man page says "When ipi_ifindex is not zero the
primary local address of the interface specified by the index overwrites
ipi_spec_dst for the routing table lookup."

But this is changing API behavior that's been this way since at least
the beginning of git history.

I guess Dave's letting this stand.  I'm posting this just to make sure
this is an explicit decision.

--
John