During Hacker School, I wanted to understand networking better, and I decided to write a miniature TCP stack as part of that. I was much more comfortable with Python than C and I’d recently discovered the scapy networking library which made sending packets really easy.
So I started writing teeceepee!
The basic idea was
- open a raw network socket that lets me send TCP packets
- send a HTTP request to
- get and parse a response
I didn’t care much about proper error handling or anything; I just wanted to get one webpage and declare victory :)
Step 1: the TCP handshake
I started out by doing a TCP handshake with Google! (this won’t necessarily run correctly, but illustrates the principles). I’ve commented each line.
The way a TCP handshake works is:
- me: SYN
- google: SYNACK!
- me: ACK!!!
Pretty simple, right? Let’s put it in code.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Wait, sequence numbers?
What’s all this about sequence numbers? The whole point of TCP is to make sure you can resend packets if some of them go missing. Sequence numbers are a way to check if you’ve missed packets. So let’s say that Google sends me 4 packets, size 110, 120, 200, and 500 bytes. Let’s pretend the initial sequence number is 0. Then those packets will have sequence numbers 0, 110, 230, and 430.
So if I suddenly got a 100-byte packet with a sequence number of 2000, that would mean I missed a packet! The next sequence number should be 930!
How can Google know that I missed the packet? Every time I receive a packet from Google, I need to send an ACK (“I got the packet with sequence number 230, thanks!”). If the Google server notices I haven’t ACKed a packet, then it can resend it!
The TCP protocol is extremely complicated and has all kinds of rate limiting logic in it, but we’re not going to talk about any of that. This is all you’ll need to know about TCP for this post!
For a more in-depth explanation, including how SYN packets affect sequence numbers, I found Understanding TCP sequence numbers very clear.
Step 2: OH NO I already have a TCP stack
So I ran the code above, and I had a problem. IT DIDN’T WORK.
But in a kind of funny way! I just didn’t get any responses. I looked in Wireshark (a wonderful tool for spying on your packets) and it looked like this:
1 2 3
Wait, what? I never sent a
RST means STOP THE
CONNECTION IT’S OVER. That is not in my code at all!
This is when I remembered that I already have a TCP stack on my computer, in my kernel. So what was actually happening was:
1 2 3 4
The new exchange was like this:
1 2 3 4 5 6
And it worked! Okay, awesome, we can now send packets AND GET RESPONSES without my kernel interfering! AWESOME.
Step 3: get a webpage!
There is an intervening step here where I fix tons of irritating bugs preventing Google from sending me the HTML for http://google.com. I eventually fixed all of these, and emerge victorious!
I needed to
- put together a packet containing a HTTP GET request
- make sure I can listen for lots of packets in response, not just one
- spend a lot of time fixing bugs with sequence numbers
- try to close the connection properly
Step 4: realize Python is slow
Once I had everything working, I used Wireshark again to look at what packets were being sent back and forth. It looked something like this:
1 2 3 4 5 6 7
The sequences of packets from Google and ACKs from me looked something like: P P P A P P P P P A P P A P P P P A. Google was sending me packets way faster than my program could keep up and send ACKs. Then, hilariously, Google’s server would assume that there were network problems causing me to not ACK its packets.
And it would eventually reset the connection because it would decide there were connection problems.
But the connection was fine! My program was totally responding! It was just that my Python program was way too slow to respond to packets in the millisecond times it expected.
(edit: this diagnosis seems to be incorrect :) you can read some discussion about what may be actually going on here)
If you’re actually writing a production TCP stack, don’t use Python. (surprise!)
I was really happy that it actually worked, though! The ARP spoofing
was extremely finicky, but I wrote a version of
curl using it which
worked about 25% of the time. You can see all the absurd code at
I think this was actually way more fun and instructive than trying to write a TCP stack in an appropriate language like C :)