Routing problem with udp, and a multihomed host in 2.4.20

Routing problem with udp, and a multihomed host in 2.4.20

[Neil Brown]

I have three subnets.  A, B and C.

I have a Linux 2.4.20 server, named bartok, with three interfaces, one
on each subnet (so it can listen to broadcast requests everywhere).

This server has a default route pointing at A.

I have a router that routes between all these subnets, and others.

I have a client that sits on subnet B.

When my client tries to talk to bartok, it might try any of 3 IP
addresses (Due to round-robining in the DNS).

All IP addresses work find when establishing a TCP connection
e.g. telnet or ssh.
They don't for UDP. 
  e.g.   rpcinfo -u bartok mountd

If I use the address on B, it works fine (as you would expect - same
subnet).
If I use the address on A it works fine (that is the default route
interface).

But if I use the address on C, it doesn't.

What happens is:

    - request goes from client to router
    - request goes from router to bartok interface on C
    - bartok issues an ARP for client on C interface which is WRONG
    - nobody replies to the ARP because client is on B, not C.

If I turn on proxy-arp on the router I can get the reply back, but I
would rather not do that.

So why does a reply to a UDP request arriving on subnet C from some
other subnet try to ARP out on subnet C instead of being routed
normally, while replies to TCP requests get routed properly and work
fine?

Is this a bug, or is there some configuration I can change?

I have double checked the subnet masks and broadcast addresses and
they work fine.
We have rp_filter set to 0 on all interfaces.
forwarding is set to 0, but setting it to 1 makes no difference.

Thanks,
NeilBrown

Re: Routing problem with udp, and a multihomed host in 2.4.20

[David S. Miller]

On Wed, 2003-02-12 at 15:18, Neil Brown wrote:
> Is this a bug, or is there some configuration I can change?

Specify the correct 'src' parameter in your 'ip' route
command invocations.

Re: Routing problem with udp, and a multihomed host in 2.4.20

[Neil Brown]

On  February 12, davem@redhat.com wrote:
> On Wed, 2003-02-12 at 15:18, Neil Brown wrote:
> > Is this a bug, or is there some configuration I can change?
> 
> Specify the correct 'src' parameter in your 'ip' route
> command invocations.

Thanks... but I think I need a bit more help.

bartok # ./ip route show
129.94.232.0/24 via 129.94.172.66 dev eth1 
129.94.242.0/24 dev eth0  proto kernel  scope link  src 129.94.242.45 
129.94.241.0/24 via 129.94.174.2 dev eth1 
129.94.172.0/22 dev eth1  proto kernel  scope link  src 129.94.172.12 
129.94.208.0/22 dev eth2  proto kernel  scope link  src 129.94.208.2 
default via 129.94.242.1 dev eth0 
bartok # ip addr show
1: lo: <LOOPBACK,UP> mtu 16436 qdisc noqueue 
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet 127.0.0.1/8 scope host lo
2: eth0: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:10:4b:1c:a3:a4 brd ff:ff:ff:ff:ff:ff
    inet 129.94.242.45/24 brd 129.94.242.255 scope global eth0
3: eth1: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:a0:c9:8f:7f:3c brd ff:ff:ff:ff:ff:ff
    inet 129.94.172.12/22 brd 129.94.242.255 scope global eth1
4: eth2: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:90:27:37:bb:d5 brd ff:ff:ff:ff:ff:ff
    inet 129.94.208.2/22 brd 129.94.242.255 scope global eth2


The client in question is 129.94.211.194
It sends a UDP request to 129.94.172.12
The reply, addressed to 129.94.211.194 causes ARP requests on bartok's
eth1

Surely the route that applies is the 4th one, which seems to have a
correct 'src' parameter.

BTW, The routes were all created with 'route', not 'ip', incase it
matters.

NeilBrown

Re: Routing problem with udp, and a multihomed host in 2.4.20

[David S. Miller]

   From: Neil Brown <neilb@cse.unsw.edu.au>
   Date: Thu, 13 Feb 2003 20:28:34 +1100

   On  February 12, davem@redhat.com wrote:
   > On Wed, 2003-02-12 at 15:18, Neil Brown wrote:
   > > Is this a bug, or is there some configuration I can change?
   > 
   > Specify the correct 'src' parameter in your 'ip' route
   > command invocations.
   
   Thanks... but I think I need a bit more help.
   
Sorry, I forgot to add that you need to enable the
arp_filter sysctl as well to make this work properly.

It should work once you do this.

Re: Routing problem with udp, and a multihomed host in 2.4.20

[Neil Brown]

On Thursday February 13, davem@redhat.com wrote:
>    From: Neil Brown <neilb@cse.unsw.edu.au>
>    Date: Thu, 13 Feb 2003 20:28:34 +1100
> 
>    On  February 12, davem@redhat.com wrote:
>    > On Wed, 2003-02-12 at 15:18, Neil Brown wrote:
>    > > Is this a bug, or is there some configuration I can change?
>    > 
>    > Specify the correct 'src' parameter in your 'ip' route
>    > command invocations.
>    
>    Thanks... but I think I need a bit more help.
>    
> Sorry, I forgot to add that you need to enable the
> arp_filter sysctl as well to make this work properly.
> 
> It should work once you do this.

Nope...

Maybe I'm not explaining myself well enough.  
So I expermented a bit more and did some "strace"ing, and read some
man pages....

It turns out that the problem occurs when send_msg is used to send a
UDP packet, and the control information contains
              struct in_pktinfo {
                  unsigned int   ipi_ifindex;  /* Interface index */
                  struct in_addr ipi_spec_dst; /* Local address */
                  struct in_addr ipi_addr;     /* Header Destination address */
              };
specifying the address and interface of the message that we are
replying to.
I'll include all the numbers below for completenes, but the brief
description goes:
 Three subnets, A,B,C  all connected by a router.
 Client X on subnet B - default route to router.
 
 Server Y: three interfaces:
     eth0 on A -  default route to router on A
     eth1 on B  ( and so directly connected to client X)
     eth2 on C

 Packet from X to Y:C  (i.e. address of eth2 on Y) goes through router
 to Y.
 Y responds with sendmsg specifying that the incoming packet was on
 eth2 and was addressed to Y:C.

 What *should* (IMO) happen is the response should have Y:C as the
 source address, and that packet should be routed with a preference
 to eth2.  As eth2 in not on B, and there are no known routes to B via
 eth2, the reply should be routed normally: i.e. directly to eth1.

 What *does* happen is that the reply is sent on eth2 as though the
 client X were local to eth2.  i.e. an ARP request is sent to find the
 MAC address, and then the packets is sent to this MAC address.

 It might be reasonable that my *should* case would require
 ip_forwarding begin turned on, but I have ip_forwarding turned on and
 it doesn't help.
 
 In any case the *does* case is wrong because it sends a packet on an
 interface to a neighbour that in known not to be directly attached to
 that interface.

Does that make my situation clearer?

Thanks,

NeilBrown
-------------------------
The numbers: 

On a multi homed host with the following interfaces:

bartok # ./ip address show
1: lo: <LOOPBACK,UP> mtu 16436 qdisc noqueue 
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet 127.0.0.1/8 scope host lo
2: eth0: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:10:4b:1c:a3:a4 brd ff:ff:ff:ff:ff:ff
    inet 129.94.242.45/24 brd 129.94.242.255 scope global eth0
3: eth1: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:a0:c9:8f:7f:3c brd ff:ff:ff:ff:ff:ff
    inet 129.94.172.12/22 brd 129.94.242.255 scope global eth1
4: eth2: <BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast qlen 100
    link/ether 00:90:27:37:bb:d5 brd ff:ff:ff:ff:ff:ff
    inet 129.94.208.2/22 brd 129.94.242.255 scope global eth2


and the following routes:

bartok # ./ip route show
129.94.232.0/24 via 129.94.172.66 dev eth1 
129.94.242.0/24 dev eth0  proto kernel  scope link  src 129.94.242.45 
129.94.241.0/24 via 129.94.174.2 dev eth1 
129.94.172.0/22 dev eth1  proto kernel  scope link  src 129.94.172.12 
129.94.208.0/22 dev eth2  proto kernel  scope link  src 129.94.208.2 
default via 129.94.242.1 dev eth0 


1/ A TCP SYN/ACK with 
	source 129.94.172.12 dest 129.94.211.194 
  that is in response to a TCP SYN with
	source 129.94.211.194 dest 129.94.172.12
  that arrived on eth1 will be sent directly to
    129.94.211.194 on eth2

  This is what you would expect.

2/ A UDP packet with 
	source 129.94.172.12 dest 129.94.211.194 
   that is sent (sendto) on a newly created and bound
   SOCK_DGRAM socket will be sent directly to
    129.94.211.194 on eth2

   This is also what you would expect.

3/ A UDP packet sent on a newly created unbound
   socket (bound to 0.0.0.0) to 129.94.211.194
   will have

	source 129.94.208.2 dest 129.94.211.194
   and will be sent directly on eth2

   Again as you would expect.

However:

4/ A UDP packet send on an unbound socket (bound to a port but not an
   IP address) to 129.94.211.194, via a sendmsg request with
   in_pktinfo specifing that the incoming packet was recieved on eth1
   and had
	source 129.94.211.194 dest 129.94.172.12
   will have
	source 129.94.172.12 dest 129.94.211.194

   and will be sent directly to 129.94.211.194 ON ETH1

   By 'sent directly' I mean if the arp table has an entry for 
   129.94.211.194 on eth1, it will be sent to that MAC address, and if
   it doesn't an ARP request will be broadcast on eth1 to find an
   appropriate MAC address.

   This is *wrong*.

   I am happy that the source address is 129.94.172.12 in this case
   while in case 3 it is 129.94.208.2.  I am not happy that it
   directly sends to eth1.

Re: Routing problem with udp, and a multihomed host in 2.4.20

[Herbert Xu]

Neil Brown <neilb@cse.unsw.edu.au> wrote:
> 
> It turns out that the problem occurs when send_msg is used to send a
> UDP packet, and the control information contains
>              struct in_pktinfo {
>                  unsigned int   ipi_ifindex;  /* Interface index */
>                  struct in_addr ipi_spec_dst; /* Local address */
>                  struct in_addr ipi_addr;     /* Header Destination address */
>              };
> specifying the address and interface of the message that we are
> replying to.

So your application is forcing the packet to go out on a specific
interface bypassing the routing table...
-- 
Debian GNU/Linux 3.0 is out! ( http://www.debian.org/ )
Email:  Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: Routing problem with udp, and a multihomed host in 2.4.20

[Neil Brown]

On Friday February 14, neilb@cse.unsw.edu.au wrote:
> 
> It turns out that the problem occurs when send_msg is used to send a
> UDP packet, and the control information contains
>               struct in_pktinfo {
>                   unsigned int   ipi_ifindex;  /* Interface index */
>                   struct in_addr ipi_spec_dst; /* Local address */
>                   struct in_addr ipi_addr;     /* Header Destination address */
>               };
> specifying the address and interface of the message that we are
> replying to.

Well, I took the plunge and hunted in and around the networking
code....

in net/ipv4/route.c, in ip_route_output_slow  there is an if clause:
   if (fib_lookup(&key, &res)) {

In my situation, fib_lookup will fail because key contains a 
non-zero "oif" (output interface), but that interface does not have a
valid route to key.dst.

What should be done in that case?  Well in my situation, the oif is
just a hint, not a requirement (it is, after all, the interface that
the request arrived on.  It is not necessarily the interface that the
reply has to go out on).  So possibly it should clear oif and try
fib_lookup again.  But it doesn't.
This is a branch for "if oif is non-zero", and it is largely a
comment:

		if (oldkey->oif) {
			/* Apparently, routing tables are wrong. Assume,
			   that the destination is on link.

			   WHY? DW.
			   Because we are allowed to send to iface
			   even if it has NO routes and NO assigned
			   addresses. When oif is specified, routing
			   tables are looked up with only one purpose:
			   to catch if destination is gatewayed, rather than
			   direct. Moreover, if MSG_DONTROUTE is set,
			   we send packet, ignoring both routing tables
			   and ifaddr state. --ANK


			   We could make it even if oif is unknown,
			   likely IPv6, but we do not.
			 */

			if (key.src == 0)
				key.src = inet_select_addr(dev_out, 0,
							   RT_SCOPE_LINK);
			res.type = RTN_UNICAST;
			goto make_route;

This comment seems to be considering a case where oif has been set as
an explicit request for the message to go on that interface, which is
not the case for "sendmsg" with an IP_PKTINFO attachment.

Maybe when ip_cmsg_send interprets the IP_PKTINFO and sets ipc->oif,
it should set some flag to say "hint".  And then fib_lookup, or
possibly ip_route_output_slow, could test for that hint and re-try
with no oif if the first try fails..

But even that is not a complete solution.  We would also need to
modify ip_route_output_key which scans a table of recently computed
routes to avoid having to do fib_lookup for every packet.  This table
would need to know about oif's that are hints, and ones that are
requirements.

At about this point it got all a bit too complicated, so I haven't
bothered to try to patch the kernel to make it work right.

Rather, I noticed that if the interface specified does have a route to
the destination, that route will be taken rather than attempting to
directly send to the dest.

So I have added default routes to all of my interfaces, not just the
prefered one, and my symptoms have gone away (and I don't need
proxy-arp on the router any more).

This will satisfy my needs for now, but I feel that something should
be done to fix this problem the "right" way.  I'm just not sure what.

NeilBrown