Fuzzing is nowadays the attack technique used by a lot of pentesters and security researchers. Whether you’re looking for vulnerabilities in media files (pictures, videos, audio stuff) or just binary files, fuzzing is the right approach if you don’t want to do some static code analysis or debug the hell out of your targets.

When it comes to fuzzing there are a few tools to mention that have established during the last years. One of them is definitely AFL. AFL has a very powerful fuzzing engine and a lot of vulnerabilities have been identified by AFL. However, AFL’s simpleness of using files as parameters to your target binary, is also one big disadvantage. While in most cases you should be fine by reading your input from files, when dealing with sockets, the code has to be modified in order to read input from files instead of sockets. This step can be very complex and involves good C/C++ skills.

After reading “How I nearly almost saved the Internet” I thought it should be an easy task to adapt the code of my target in order to run it by AFL. This step was indeed very time-consuming and was done after endless debug sessions. Below I’ll try to give some overview which steps might be essential in fuzzing your target in an effective way. Modifying the source code for your needs definitely requires a deep understanding of the programms workflow. I hope this post will provide some useful tips how to speed up the fuzzing process - whatever your target will be.


Why would one want to fuzz conntrack-tools anyway? Before we continue: No 0day here! As sad as it sounds, I wasn’t able to find any crashes using AFL. But that is no guarantee for safe, secure code so there might be some vulnerabilities ready to be found.

As you probably know software security doest not only apply to web stuff but also to software that runs critical network infrastructure. Since the IT sec community seems to have a huge focus on web applications (it’s still the main entry point, right?), a few people (in relation to the first group mentioned) do have a look at network/system level stuff. In my case I wanted to audit and fuzz software that deals with connection tracking management. In my particular case I’ve used conntrackd to manage and synchronize the connection tracking information between 2 hosts.


On 2 virtual machines I’ve downloaded the conntrack-tools (also includes conntrackd), compiled the code and ran the daemons:

blockdiag code
  network nat-network {
	  address = ""

	  nat01 [address = ""];
	  nat02 [address = ""];


So nat01 and nat02 were supposed to exchange information about their connection tracking table. In a nutshell conntrackd:

Let’s have a look at some configuration file:

Sync {
        Mode FTFW{
    UDP Default {
        Port 3780
        Interface ens3 
        SndSocketBuffer 1249280
        RcvSocketBuffer 1249280
        Checksum on

This configuration says:

Basically this how synchronization works. Of course some details have been ommitted for the sake of simplicity. For the detailed configuration please refer to the documentation. In the end the daemons are started by specifying the config file:

$ ./src/conntrackd -C <config file>

If you experience some crazy error messages, make sure you have following kernel modules loaded: af_netlink

Network traffic

If you then sniff for packets you’ll see some magic packets:

$ tshark -nr nat-traffic.pcap -E separator="," -T fields -e udp.srcport -e udp.dstport -e udp.length -e data  | head -n 5

As you have noticed the UDP payloads above are very similar to each other. However, bigger payloads were also transmitted:



Now what valuable information can we extract from the payloads? Using scapy I’ve did some basic traffic analysis:

In [11]: p.payload.payload.show()

###[ UDP ]###
  sport     = 53085
  dport     = nnp
  len       = 76
  chksum    = 0x1aa6
###[ Raw ]###
     load      = '\x11\x00\x00DX\x15\xc1c\x00\x0c\x00\x00\n\x17h\x026\xef6\x8f\x00\x08\x00\x05\x00\x00\x01\x9e\x00\x05\x00\x02\x06\x00\x00\x00\x00\x08\x00\x03\x91\xf2\x01\xbb\x00\x05\x00\x04\x06\x00\x00\x00\x00\x08\x00\x06\x00\x00\x00\x1e\x00\x08\x00\x0c~d\xb2\xc3'

Apparently this is no ASCII data (who would have expected that anyway? :) Using gdb I wanted to know what’s inside the data and did some dynamic analysis. This way I was able to learn more about the data being passed over the wire and how the source code is structured.

But first I had to make the code debuggable:

$ CCFLAGS="-g3 -gdwarf2" ./configure

Afterwards you should be able to run it:

$ sudo ./src/conntrackd -C conntrackd.conf

Sniffing the traffic you should see sth like this:

$ sudo tshark -i ens3 -Y udp -E separator="," -T fields -e ip.src -e udp.srcport -e udp.dstport -e udp.length -e data,,48786,3780,16,1a10000858109293,,48786,3780,16,1a10000858109294,,48786,3780,16,1a10000858109295,,48786,3780,16,1a10000858109296,,48786,3780,16,1a10000858109297,,48786,3780,16,1a10000858109298,,48786,3780,16,1a10000858109299,,48786,3780,16,1a1000085810929a
^C8 packets captured

The daemon will try to send out some “keep-alive” messages to its counterpart. Now we send those packets manually and debug the interesting part in the code. For packet generation I’ll use scapy again (running as root since you’ll need to create RAW sockets):

In [1]: from scapy.all import *
WARNING: No route found for IPv6 destination :: (no default route?)

In [2]: import binascii

In [3]: p=Ether()/IP(dst="")/UDP(dport=3780)

In [4]: payload = binascii.unhexlify("1a10000858109293")

In [5]: p/Raw(load=payload)
Out[5]: <Ether  type=IPv4 |<IP  frag=0 proto=udp dst= |<UDP  dport=nnp |<Raw  load='\x1a\x10\x00\x08X\x10\x92\x93' |>>>>

In [7]: sendp(p/Raw(load=payload))
Sent 1 packets.

Data flow

After heavy debugging sessions I was able to understand the data flow inside the code. Below I’ll try to explain how conntrackd basically works:

  1. The configuration file gets parsed and several options are set
  2. In src/sync-mode.c you have init_sync where a lot of vodoo happens but the most important part is:
	for (i=0; i<STATE_SYNC(channel)->channel_num; i++) {
		int fd = channel_get_fd(STATE_SYNC(channel)->channel[i]);
		fcntl(fd, F_SETFL, O_NONBLOCK);

		switch(channel_type(STATE_SYNC(channel)->channel[i])) {
			register_fd(fd, channel_accept_cb,
			register_fd(fd, channel_handler,

Depending on the sync mode, then for every channel a channel handler is being registered. This will be called every time an “event” occurs, that means a packet arrives:

Since I chose to test for UDP packets, let’s have a look at channel_handler:

static void channel_handler(void *data)
	struct channel *c = data;
	int k;

	for (k=0; k<CONFIG(event_iterations_limit); k++) {
		if (channel_handler_routine(c) == -1) {

This will then call channel_handler_routine:

static int channel_handler_routine(struct channel *m)
	ssize_t numbytes;
	ssize_t remain, pending = cur - __net;
	char *ptr = __net;

	numbytes = channel_recv(m, cur, sizeof(__net) - pending);
	if (numbytes <= 0)
		return -1;

	remain = numbytes;
	if (pending) {
		remain += pending;
		cur = __net;

	while (remain > 0) {
		struct nethdr *net = (struct nethdr *) ptr;
		int len;


	return 0;

Obvisouly channel_recv is called:

int channel_recv(struct channel *c, char *buf, int size)
	return c->ops->recv(c->data, buf, size);

Depending on which channel (variable c) is currently active channel_recv will then call the channel’s recv function. The channel for UDP is stored in a structure called channel_udp:

struct channel_ops channel_udp = {
	.headersiz	= 28, /* IP header (20 bytes) + UDP header 8 (bytes) */
	.open		= channel_udp_open,
	.close		= channel_udp_close,
	.send		= channel_udp_send,
	.recv		= channel_udp_recv,
	.get_fd		= channel_udp_get_fd,
	.isset		= channel_udp_isset,
	.accept_isset	= channel_udp_accept_isset,
	.stats		= channel_udp_stats,
	.stats_extended = channel_udp_stats_extended,

As you can notice channel_udp.recv is a function pointer to channel_udp_recv:

static int
channel_udp_recv(void *channel, char *buf, int size)
	struct udp_channel *m = channel;
	return udp_recv(m->server, buf, size);

So now we’re getting closer to the function that is actually responsible for reading data from the wire. In udp_recv all the magic happens:

ssize_t udp_recv(struct udp_sock *m, void *data, int size)
	ssize_t ret;
	socklen_t sin_size = sizeof(struct sockaddr_in);

        ret = recvfrom(m->fd,
		       (struct sockaddr *)&m->addr,
	if (ret == -1) {
		if (errno != EAGAIN)
		return ret;

	m->stats.bytes += ret;

	return ret;

Nothing special about: Just call recvfrom, read the data into data and update some statistics.

The plan

If we want to do some fuzzing we’ll have to modify udp_recv to read from some file instead of sockets. Re-writing the function was actually a quite easy task:

#ifdef FUZZ
ssize_t my_udp_recv(struct udp_sock *m, void *data, int size, char *filename) {
    ssize_t ret;
    long f_size;
    FILE *f;
	socklen_t sin_size = sizeof(struct sockaddr_in);

    // Read file
    if (!(f = fopen(CONFIG(fuzz_file), "rb"))) {
        fprintf(stderr, "[FUZZ] Couldn't open file\n");
    // Get file length
    fseek(f, 0L, SEEK_END);
    f_size = ftell(f);

    ret = fread(data, f_size, 1, f);
    if (ret == -1) {
		if (errno != EAGAIN)
		return ret;
    m->stats.bytes += f_size;

	return f_size;

Easy, right? Now my_udp_recv will read the file specified by filename and store the contents inside data. However, this simple modification implied several changes across the whole base.

You can see the whole changes here.

Afterwards I was able to compile the code with this new feature:

$ git clone https://github.com/dorneanu/conntrack-fuzzing
$ cd conntrack-fuzzing
$ CFLAGS=-DFUZZ LDFLAGS=-ldl ./configure
$ ./src/conntrackd --help 
Connection tracking userspace daemon v1.4.4
Usage: ./src/conntrackd [commands] [options]

Daemon mode commands:
  -X [fuzz-file], specify fuzz file
  -d [options]          Run in daemon mode


Now you can specify a file (-X) which will be used as parameter for my_udp_recv. The “daemon” will now just read the contents of the file, analyze the input and then simply exit.

Generating input values

Now that I was able to specify files as data input, I had to generate some input examples for the screening process of AFL. By sniffing enought traffic between 2 productive hosts running conntrackd I got a lots of payloads I could then use for AFL. All I had to do was to store those values as binary files:

In [1]: payload = "1000003c5815c1c9000c00000a171f0cc3b265fa000800050000019800050002110000000008000305be0035000800060000001e0008000c7e64b2c3"

In [2]: import binascii

In [3]: with open("/tmp/input1", "wb") as f:
   ...:     f.write(binascii.unhexlify(payload))
   ...:     f.close()

In [4]: !hexdump -C /tmp/input1
00000000  10 00 00 3c 58 15 c1 c9  00 0c 00 00 0a 17 1f 0c  |...<X...........|
00000010  c3 b2 65 fa 00 08 00 05  00 00 01 98 00 05 00 02  |..e.............|
00000020  11 00 00 00 00 08 00 03  05 be 00 35 00 08 00 06  |...........5....|
00000030  00 00 00 1e 00 08 00 0c  7e 64 b2 c3              |........~d..|

For my test cases I’ve chosen ca. 4-5 different payloads and stored them as binary files. Now I was ready to run AFL.

Prepare for AFL

Inside the project I’ve then created the directory structure for AFL:

$ mkdir -p fuzzing/{ftfw,notrack}/input

For every sync mode (ftwfw and notrack) I’ve then copied the input values into input. As a next step I had to compile the source using afl-gcc:

$ CC=afl-gcc AFL_USE_ASAN=1 CFLAGS=-DFUZZ LDFLAGS=-ldl ./configure
$ make -j 4

Running AFL

AFL itself had to be ran as root since conntrackd requires privileged permissions:

$ echo core >/proc/sys/kernel/core_pattern
$ afl-fuzz -i fuzzing/ftfw/input/ -o fuzzing/ftfw/output ./src/conntrackd -C conntrackd.conf -X @@
[+] You have 2 CPU cores and 1 runnable tasks (utilization: 50%).
[+] Try parallel jobs - see /usr/local/share/doc/afl/parallel_fuzzing.txt.
[*] Checking core_pattern...
[*] Setting up output directories...
[*] Scanning 'fuzzing/ftfw/input/'...
[+] No auto-generated dictionary tokens to reuse.
[*] Creating hard links for all input files...
[*] Validating target binary...
[*] Attempting dry run with 'id:000000,orig:r1'...
[*] Spinning up the fork server...
[+] All right - fork server is up.
    len = 136, map size = 597, exec speed = 3185 us
[*] Attempting dry run with 'id:000001,orig:r2'...
    len = 204, map size = 622, exec speed = 2569 us
[*] Attempting dry run with 'id:000002,orig:r3'...
    len = 68, map size = 619, exec speed = 2545 us
[*] Attempting dry run with 'id:000003,orig:r4'...
    len = 476, map size = 597, exec speed = 2867 us
[!] WARNING: No new instrumentation output, test case may be useless.
[+] All test cases processed.

[!] WARNING: Some test cases look useless. Consider using a smaller set.
[+] Here are some useful stats:

    Test case count : 2 favored, 0 variable, 4 total
       Bitmap range : 597 to 622 bits (average: 608.75 bits)
        Exec timing : 2545 to 3185 us (average: 2792 us)

[*] No -t option specified, so I'll use exec timeout of 20 ms.
[+] All set and ready to roll!


You can also monitor the current input using:

$ watch -n1 'cat fuzzing/ftfw/output/.cur_input | hexdump -C'


While the preparation steps were indeed time-consuming I think I have definitely improved my gdb skills looking at the code and trying to understand how things work. Additionally I was able to trace the data flow within the code - from the sockets till to the functions where data got analyzed. The fuzzing process itself was quite straight-forward and mainly managed by AFL itself. Unfortunately (depends on the perspective you’re looking from) I wasn’t able to find any crashes inside conntrack-tools which is actually good. By crashing the daemon some network connectivity would have been heavily affected. But again: The fact that I wasn’t able to find any issues, is no guarantee for secure software. Maybe other will have a look at the code and do some static code analysis.

For me personally this was a hell of fun and I’m already looking forward to my next fuzzing project. Stay tuned and feel free to drop your comments for additional questions/suggestions regarding the steps explained here.