1

I use libtins (It uses Pcap) to capture link layer packets and forward to a network namespace where the actual application runs on.

Client(Browser) -> Server -> Pcap -> Pcap Send -> br0 (Bridge) -> Namespace -> Application 

Now, I see that the packets are forwarded and visible on Tcpdump snapshot but I don't see they are received by the application it-self, no signs of outgoing packets either.

It appear that the loopback interface in namespace haven't gotten packets either so looks like Kernel doesn't route the packets to loopback for some reason.

I believe the application by default listen to loopback interface, and based on my assumption it's where localhost runs on.

I've tried to bind the application into the veth0 network interface's IP in namespace but it didn't work either.

I am able ping any website inside network namespace.

Regarding Checksum, TCP Dump shows the packets has correct checksum

Here how I set up my network namespace,

sysctl -w net.ipv4.ip_forward=1 &&
sysctl -w net.ipv6.conf.all.forwarding=1 &&
ip netns add namespace1 &&
ip link add veth0 type veth peer name veth1 &&
ip link set veth0 netns namespace1 &&
ip netns exec namespace1 ip addr add 192.168.1.11/24 dev veth0 &&
ip link add name br0 type bridge &&
ip link set br0 up &&
ip link set veth1 up &&
ip netns exec namespace1 ip link set veth0 up &&
ip netns exec namespace1 ip link set lo up &&
ip link set veth1 master br0 &&
ip addr add 192.168.1.10/24 brd + dev br0 &&
ip -all netns exec ip route add default via 192.168.1.10 &&
iptables -t nat -A POSTROUTING -s 192.168.1.0/24 -j MASQUERADE

Here an example of a packet captured by TCP Dump on network namespace,

16:50:11.742116 IP (tos 0x2a,ECT(0), ttl 115, id 8487, offset 0, flags [DF], proto TCP (6), length 52)
    MYHOMEIP.51202 > SERVERIP.https: Flags [SEW], cksum 0x59a4 (correct), seq 332112346, win 64240, options [mss 1460,nop,wscale 8,nop,nop,sackOK], length 0

Here is output of ss -lntp on namespace,

State           Recv-Q          Send-Q                     Local Address:Port                     Peer Address:Port          Process
LISTEN          0               511                                    *:443                                 *:*              users:(("node",pid=30798,fd=20))
LISTEN          0               511                                    *:7000                                *:*              users:(("node",pid=30798,fd=19))

Looking at the output, I see Send Q is filled with 511 and I'm not sure what that means.

The answer below doesn't add anything new and it didn't fix my problem.

I would love to award a bounty but can't find the button.

Apparently, A similar question has been asked on this site years ago but no one succeeded providing working solution.

  • Hi, a general tip, tcpdump see the packet before the firewall, make sure it’s allowed in your firewall traffic, as it would explain why tcpdump see the oacket while your apps not. – yagmoth555 Jul 14 at 2:09
  • @yagmoth555 I can curl and view the site directly from default namespace. Just when I forward packets, it start different behavior as I described above. – sqlbie Jul 14 at 3:59
1

Talk to veth

Your exact commands work on my system.

Here's what I've checked:

man nc
ip netns exec namespace1 nc -l 192.168.1.11 8123

Now on the remote client:

ip ro add 192.168.1.11/32 via SERVERIP
echo MyTestData | nc -v 192.168.1.11 8123

Talk to loopback

You are saying something about loopback though. Obviously won't work for 127/8, so I've prepared routing:

ip netns exec namespace1 ip a add 192.168.2.2 dev lo
ip ro add 192.168.2.2/32 via 192.168.1.11

Now on the remote client:

ip ro add 192.168.2.2/32 via SERVERIP
echo MyTestData | nc -v  192.168.2.2 8123

That being said, one more caveat. You seem to be coming from netadmin background. The server is not a router and so using loopback for anything else than a 127/8 "traffic to itself" is a very uncommon practice on application servers. Clients are expected to talk to the eth/veth.

But I cannot tell all the clients to add a route

Looking at your previous question (now deleted) about how kernel is supposed to handle routing in case client talks to "public" IP I'd like to remind about a thing called DNAT. (It's something else that MASQUERADE. You often use DNAT together with MASQUERADE.)

So instead of ip ro add at a client you can do DNAT without touching the client. I don't want to come off offensive, I'm stating the obvious, just in case.

| improve this answer | |
  • Client is Chrome Browser not the Default Namespace. I'm doing load balancing prototype and I forward packets to this isolated namespace which the NodeJS application runs on and listening to a port, but it appear that Kernel doesn't route the packets to loopback. Node listen to localhost. – sqlbie Jul 11 at 23:23
  • I can run tcpdump on that namespace and see packets but Node don't see them. I believe the reason is that Kernel doesn't route the packets to localhost which Node listens on. – sqlbie Jul 11 at 23:26
  • I meant by loopback is localhost which node runs on, when I try tcpdump on namespace B's loopback interface, I don't see any packets. – sqlbie Jul 11 at 23:28
  • @sqlbie Gimme your ss -lntp for that nodejs because I really don't know what does it mean to listen on a localhost. – kubanczyk Jul 12 at 19:42
  • @sqlbie Also, what's wrong with using nc like I did. It will rule out all the doubts - the nodejs app can be broken, right? – kubanczyk Jul 12 at 19:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.