Implementing IP_SENDIF (like SO_BINDTODEVICE)

[Bruce M Simpson]

On Thu, Oct 28, 2004 at 03:41:34PM +0400, Gleb Smirnoff wrote:
> ng_device can be attached to "orphans" hook of ng_ether. /dev/ngdX opened
> by dhcpd, and packets processed.

This seems to me like pure configuration overkill. It would require that
people compile and load netgraph to run dhclient, and that the netgraph
graph be created and populated correctly.

> What is benefit to get rid of bpf? What is problem with it?

IPV6_NEXTHOP is intended to support this for IPv6. It doesn't right
now, it returns EAFNOSUPPORT.

Please see my attached game plan.

BMS
-------------- next part --------------
The story with FreeBSD and 'ip unnumbered' in IPv4 land

 - Multicast group membership
   - Dealt with in ip_multicast_if()
     (can specify ifIndex in 0/8 as per RFC1724 hack)

 - IP forwarding FIB
   - Routes can be specified with the rt_gate field set to an
     sockaddr_dl (AF_LINK) filled out only with an interface name
     and no RTF_GATEWAY or RTF_LLINFO flag set for an unnumbered
     serial interface.
   - For host mode, sending packets to destinations whose prefix
     and next-hop match such a route in the FIB are OK, but the
     source address may need to be explicitly specified (it may
     default to INADDR_ANY otherwise).

 - Unicast sends
   - Currently no way to explicitly send a udp datagram to such
     an interface without specifying a destination with a route
     pointing to that interface.

  - Running BGP over an unnumbered interface is possible.

    For an active session open, there has to be a route to one of
    the addresses configured on the remote peer which traverses
    the unnumbered interface, and this address must be specified
    during connect().

    For a passive session open, it would probably require that
    net.inet.ip.check_interface is 0, so that incoming sessions
    can be accepted on the unnumbered interface for passive open.

    The IP address for the passive open (seen during accept() will
    of course have to correspond to one of the other addresses
    configured on the local host.

 - For a protocol which needs to send broadcasts via the existing
   unicast path and which cannot use multicast, there is no way
   of doing this in the current iteration of the code.

TODO:
  - Add support for explicitly specifying the interface used during
    ip_output(). This would require SO_DONTROUTE and IP_SENDIF to
    be set. SO_BINDTODEVICE could be emulated.

    IP_SENDIF specified as a control message would be purely
    temporary and only necessary for an unconnected datagram
    socket send.

    The inpcb already has an interface index field for IPv6. This
    should be unravelled for IPv4 too.

    IP_SENDIF passed to setsockopt() would be like SO_BINDTODEVICE
    only instead of specifying an interface name, the interface
    index is specified and cached in the inpcb.

    SO_BINDTODEVICE would perform a lookup of the interface index
    by name, just like Linux, then perform the same action as
    IP_SENDIF when used as a socket option. This means calling
    ifunit() to resolve the name to an ifnet, then just cache
    ifnet->if_index in the inpcb.

    SO_BINDTODEVICE can either be implemented purely in the
    Linuxulator, or within src/sys/net/ itself.

    The actual hack in ip_output() would need to check if the
    interface had gone away, and if it had, drop the datagram -
    this stuff only takes effect if SO_DONTROUTE is specified.
    This is easily done...

    (SO_DONTROUTE of course implies IP_ROUTETOIF - they are
     defined to be identical).
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 167 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-net/attachments/20041028/d9a5c5e8/attachment.bin