How can I have DNS name resolving running while other protocols seem to be down?

Question

We are trying to implement a software based on Moxa UC-7112-LX embedded computer (uClinux OS). We use Cinteron MC52i GSM modem (regular GPRS service) and standart pppd to connect to the Internet.

Everything seems to be fine, right after the connection. Ping utility is working, Socket functions in my program work normally too. However after some time ppp connection brokes in a very peculiar way. These are the symptoms of that situation:

• When I call ping utility with some host name as parameter the system is able to resolve it's IP and starts sending ICMP packets but gets no response. I am trying different web resources names, so that the system cannot have their addresses cached or something. Whatever I choose, the system correctly resolves IP but can't get any ping responce.
• connect() and write() functions in my application give no error return but when it comes to read() the function returns with errno set to ECONNRESET (Connection reset by peer). The program uses standard socket functions (TCP protocol)
• the ppp link is shown as running ( ifconfig ppp0 )

So, the situation that I have is: the link is good enough to maintain DNS resolving service (UDP is working?) but NOT good enough to run TCP connection and receive ping echoes...

The situation does not appear all the time. Sometimes the system can work normally for days without any problem. Whenever the problem appears, simple reset solves everything.

I know that the system we use is quite exotic, and the situation described here may be connected with some buggy tcp stack or pppd implementation. Considering that the system is preconfigured by the manufacturer I don't have any options to rebuild/change the OS firmware.

Still I hope that someone have seen the similar situation on any linux-like system. Is there any way to test why DNS name resolving is working while the other network stuff does not? Is it possible to remove such connection state with some pppd settings?

Edit:

First of all, I'd like to address the possibility of local caching of the IP addresses. I don't have dig utility and I have no idea how to check which host gives the result to getaddrinfo() . Still I'm sure that the addresses are not cached cause I'm trying to ping totally random URLs. Also given the slow GPRS response time it is not necessary to have the time measuring utility to see that ping takes 1-2 seconds or more to resolve IP before starting sending out packets. Furthermore ncsd , BIND or any dns servers do not run locally on the machine. I understand that you may not see that as proof, but that's what I have given the utility set available on my system.

I'd like to give some additional information concerning the internet connection operation.

Normal connection state

The rc script at system load runs another script as background process:

sh /etc/connect &

The connect script is as follows:

#!/bin/sh
echo First connect attempt > /etc/ppp/conn.info
while true
do
date >> /etc/ppp/conn.info
pppd call mts
echo Reconnecting... >> /etc/ppp/conn.info
done

The reason that I've made a loop here is simple: the connection persists for several hours and after that it always breaks. Unfortunately my implementation of pppd does not support the logfile option (so I can't see why is it broken). persist does not seem to work either so I've come to the connect script above. The pppd options are:

/dev/ttyM0 115200 crtscts
connect 'chat -f /etc/ppp/peers/mts.chat'
noauth
user mts
password mts
noipdefault
usepeerdns
defaultroute

ifconfig ppp0 gives:

ppp0      Link encap:Point-Point Protocol  
          inet addr:172.22.22.109  P-t-P:192.168.254.254  Mask:255.255.255.255
          UP POINTOPOINT RUNNING NOARP MULTICAST  MTU:1500  Metric:1
          RX packets:34 errors:0 dropped:0 overruns:0 frame:0
          TX packets:36 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:3 
          RX bytes:3130 (3.0 KiB)  TX bytes:2250 (2.1 KiB) 

And thats where it starts getting strange. Whenever I connect I'm getting different inet addr but Ptp is always the same: 192.168.254.254. This is the same address that appears in default gateway entry, as given by netstat -rn :

Kernel IP routing table
Destination     Gateway         Genmask         Flags   MSS Window  irtt Iface
192.168.254.254 0.0.0.0         255.255.255.255 UH        0 0          0 ppp0
192.168.4.0     0.0.0.0         255.255.255.0   U         0 0          0 eth1
192.168.15.0    0.0.0.0         255.255.255.0   U         0 0          0 eth0
192.168.0.0     192.168.15.1    255.255.0.0     UG        0 0          0 eth0
0.0.0.0         192.168.254.254 0.0.0.0         UG        0 0          0 ppp0 

route -Cevn is unavailable on my system, route gives the same info as above.

But I'm never able to ping the 192.168.254.254, not even when everything is working as intended: tcp connection, ping, DNS etc. Here is the result of traceroute:

traceroute to kernel.org (149.20.4.69), 30 hops max, 40 byte packets
 1  172.16.4.210 (172.16.4.210)  528.765 ms  545.269 ms  616.67 ms
 2  172.16.4.226 (172.16.4.226)  563.034 ms  526.176 ms  537.07 ms
 3  10.250.85.161 (10.250.85.161)  572.805 ms  564.073 ms  556.766 ms
 4  172.31.250.9 (172.31.250.9)  556.513 ms  563.383 ms  580.724 ms
 5  172.31.250.10 (172.31.250.10)  518.15 ms  526.403 ms  537.574 ms
 6  pub2.kernel.org (149.20.4.69)  538.058 ms  514.222 ms  538.575 ms
 7  pub2.kernel.org (149.20.4.69)  537.531 ms  538.52 ms  537.556 ms
 8  pub2.kernel.org (149.20.4.69)  568.695 ms  523.099 ms  570.983 ms
 9  pub2.kernel.org (149.20.4.69)  526.511 ms  534.583 ms  537.994 ms
##### traceroute loops here - why??  #######

So, I can assume that 172.16.4.210 is peer's address. Such address is pingable in any case (see below). I have no idea why the structure of traceroute output is like this (packets come from internal network of ISP right to the destination, 'loop' at the destination address - it just should not be like this).

Also I would like to note that I can ping DNS server but traceroute does not go all the way up to it.

You may notice that there are eth0 and eth1 devices. They are irrelevant to the case. eth1 is not connected and eth0 is connected to lan without internet access.

Bad connection state

So, some time passes and the situation under question appears. I can't ping anything but DNS server (and peer, the address for which I get from traceroute result for the DNS) and cant communicate with remote host via tcp. DNS resolving is working

The network utilites give the same output as in normal state. I have the same unpingable peer (192.168.254.254 from ifconfig result), the routing table is the same:

# ifconfig ppp0
ppp0      Link encap:Point-Point Protocol
          inet addr:172.22.22.109  P-t-P:192.168.254.254  Mask:255.255.255.255
          UP POINTOPOINT RUNNING NOARP MULTICAST  MTU:1500  Metric:1
          RX packets:297 errors:0 dropped:0 overruns:0 frame:0
          TX packets:424 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:3
          RX bytes:33706 (32.9 KiB)  TX bytes:27451 (26.8 KiB)

# route
Kernel IP routing table
Destination     Gateway         Genmask         Flags Metric Ref    Use Iface
192.168.254.254 *               255.255.255.255 UH    0      0        0 ppp0
192.168.4.0     *               255.255.255.0   U     0      0        0 eth1
192.168.15.0    *               255.255.255.0   U     0      0        0 eth0
192.168.0.0     192.168.15.1    255.255.0.0     UG    0      0        0 eth0
default         192.168.254.254 0.0.0.0         UG    0      0        0 ppp0

Note that the original ppp connection (one which I used to provide the output from normal state) persisted. My /etc/connect script did not loop (there was no new record in a makeshift log the script makes).

Here goes the ping to DNS server:

# cat /etc/resolv.conf
#search moxa.com
nameserver 213.87.0.1
nameserver 213.87.1.1
# ping 213.87.0.1
PING 213.87.0.1 (213.87.0.1): 56 data bytes
64 bytes from 213.87.0.1: icmp_seq=0 ttl=59 time=559.8 ms
64 bytes from 213.87.0.1: icmp_seq=1 ttl=59 time=509.9 ms
64 bytes from 213.87.0.1: icmp_seq=2 ttl=59 time=559.8 ms

And traceroute:

# traceroute 213.87.0.1
traceroute to 213.87.0.1 (213.87.0.1), 30 hops max, 40 byte packets
 1  172.16.4.210 (172.16.4.210)  542.449 ms  572.858 ms  595.681 ms
 2  172.16.4.214 (172.16.4.214)  590.392 ms  565.887 ms  676.919 ms
 3  * * *
 4  217.8.237.62 (217.8.237.62)  603.1 ms  569.078 ms  553.723 ms
 5  * * *
 6  * * *
 ## and so on ###

*** lines may look like trouble but im getting the same traceroute for that DNS in normal situation

ping to 172.16.4.210 works fine as well.

Now to TCP. I've started a simple echo server on my PC and tried to connect via telnet to it (the actual ip address is not shown):

# telnet XXX.XXX.XXX.XXX 9060
Trying XXX.XXX.XXX.XXX(25635)...
Connected to XXX.XXX.XXX.XXX.
Escape character is '^]'.
aaabbbccc
Connection closed by foreign host.

So thats what happened here. Successfull connect() just like in my custom application is followed by Connection closed... when telnet called read() . The actual server did not receive any incoming connection. Why did 'connect()' return normally (it could not get the handshake response from the host!) is beyond my scope of knowledge.

Sure enough same telnet test works fine in normal state.

Note:

I did not publish this on serverfault cause of the embedded nature of my system. serverfault as far as I understand deals with more conventional systems (like x86s running 'normal' linux). I just hope that stackoverflow has more embedded experts who know such systems as my Moxa.

Answer 1

Q : How can I have DNS name resolving running while other protocols seem to be down?

A : Your local DNS resolver ( bind is another possibility besides ncsd ) might be caching the first response. dig will tell you where you are getting the response from:

[mpenning@Bucksnort ~]$ dig cisco.com

; <<>> DiG 9.6-ESV-R4 <<>> +all cisco.com
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 22106
;; flags: qr rd ra; QUERY: 1, ANSWER: 1, AUTHORITY: 2, ADDITIONAL: 0

;; QUESTION SECTION:
;cisco.com.         IN  A

;; ANSWER SECTION:
cisco.com.      86367   IN  A   198.133.219.25

;; AUTHORITY SECTION:
cisco.com.      86367   IN  NS  ns2.cisco.com.
cisco.com.      86367   IN  NS  ns1.cisco.com.

;; Query time: 1 msec       <----------------------- 1msec is usually cached
;; SERVER: 127.0.0.1#53(127.0.0.1)  <--------------- Answered by localhost
;; WHEN: Wed Dec  7 04:41:21 2011
;; MSG SIZE  rcvd: 79

[mpenning@Bucksnort ~]$

If you are getting a very quick (low milliseconds) answer from 127.0.0.1 , then it's very likely that you're getting a locally cached answer from a prior query of the same DNS name (and it's quite common for people to use caching DNS resolvers on a ppp connection to reduce connection time, as well as achieving a small load reduction on the ppp link).

If you suspect a cached answer, do a dig on some other DNS name to see whether it can resolve too.

• If random DNS names continue resolution and you still cannot make a TCP connection to a certain host, this is worthy of noting when you edit the question after this investigation.
• If random DNS names don't resolve, then this is indicative of something like the loss of your default route, or the ppp connection going down.

Other diagnostic information

If you find yourself in either of the last situations I described, you need to do some IP and ppp-level debugs before this can be isolated further. As someone mentioned, tcpdump is quite valuable at this point, but it sounds like you don't have it available.

I assume you are not making a TCP connection to the same IP address of your DNS server. There are many possibilities at this point... If you can still resolve random DNS names, but TCP connections are failing, it is possible that the problem you are seeing is on the other side of the ppp connection, that the kernel routing cache (which holds a little TCP state information like MSS ) is getting messed up, you have too much packet loss for tcp , or any number of things.

Let's assume your topology is like this:

       10.1.1.2/30     10.1.1.1/30
       [ppp0]          [pppX]
uCLinux----------------------AccessServer---->[To the reset of the network]

When you initiate your ppp connection, take note of your IP address and the address of your default gateway:

ip link show ppp0      # display the link status of your ppp0 intf (is it up?)
ip addr show ppp0      # display the IP address of your ppp0 interface
ip route show          # display your routing table
route -Cevn           # display the kernel's routing cache

Similar results can be found if you don't have the iproute2 package as part of your distro ( iproute2 provides the ip utility):

ifconfig ppp0               # display link status and addresses on ppp0
netstat -rn                 # display routing table
route -Cevn                 # display kernel routing table

For those with the iproute2 utilities (which is almost everybody these days), ifconfig has been deprecated and replaced by the ip commands; however, if you have an older 2.2 or 2.4-based system you may still need to use ifconfig .

Troubleshooting steps :

1. When you start having the problem, first check whether you can ping the address of pppX on your access server.

   • If you can not ping the ip address of pppX on the other side, then it is highly unlikely your DNS is getting resolved by anything other than a cached response on your uCLinux machine.
   • If you can ping pppX , then try to ping the ip address of your TCP peer and the IP address of the DNS (if it is not on localhost ). Unless there is a firewall involved, you must be able to ping it successfully for any of this to work.

2. If you can ping the ip address of pppX but you cannot ping your TCP peer's ip address, check your routing table to see whether your default route is still pointing out ppp0

3. If your default route points through ppp0 , check whether you can still ping the ip address of the default route.

4. If you can ping your default route and you can ping the remote host that you're trying to connect to, check the kernel's routing cache for the IP address of the remote TCP host.... look for anything odd or suspicious

5. If you can ping the remote TCP host (and you need to do about 200 pings to be sure... tcp is sensitive to significant packet loss & GPRS is notoriously lossy), try making a successful telnet <remote_host> <remote_port> . If both are successful, then it's time to start looking inside your software for clues.

If you still can't untangle what is happening, please include the output of the aforementioned commands when you come back... as well as how you're starting the ppp connection.

Answer 2

Pings should never be part of an end-user application(see note), and no program should rely on ping to function. At best ping might tell us that a part of the TCP/IP stack was running on the remote. See my argument here .

What the OP describes as a problem doesn't seem to be a problem. All network connections fail, the resolver may or may not use the network, and ping isn't really helpful. I would guess that the OP can check that the modem is connected or not, and if it isn't connect again.

edit: Pseudo code

do until success
try
  connect "foobar.com"
  try
    write data
    read response
  catch
    not success
  endtry

catch error
  'modem down - reconnect
  not success
end try
loop

Note: the exception would be if you are writing a network monitoring application for a networking person.