My submissions for r2wars 2020
If I learnt one thing, it's that ARM is the future
r2wars is a CoreWar-like game thar runs within the radare2 ESIL virtual machine. In short, you have two programs running in a shared memory space (1kb), with the goal of killing the other and surviving as long as possible. r2wars was conducted as a part of r2con2020.
My first submission was an incredibly simple “bomber”. All it does is write code to a location, jump there, and continue executing the same thing over and over.
mov eax, 0xfeebfeeb; just some bad jumps mov ebx, eax mov ecx, eax mov edx, eax mov ebp, eax mov edi, eax mov esp, 0x3fc mov esi, 0x3fd mov [esi], 0xe6ff60 jmp esi
Specifically, it writes
0xe6ff60, which is
pushal jmp esi
effectively looping over and over.
pushal is a very interesting x86
instruction, that pushes all the registers and decrements the stack
esp by how many ever bytes were pushed. Nifty, especially if
you’re looking for high throughput (to bomb the address space). Here, it
starts bombing from
0x000 (and below, because there’s no
bounds checking in place), and ends up killing itself, since writing
outside of the arena (
0x400) is illegal.
Ultimately, this bot placed 7th out of 9 contestants — an underwhelming outcome. I had to fix this.
I sat for a second and recollected the different reasons for my bot
getting killed, and the one that occurred the most was my bot
insta-dying to bad instructions being written from
0x400 — i.e. from
near where I’m positioned. Nearly all competing bots write from bottom
pushal decrements the stack pointer. So the obvious
solution was to reposition my initial payload way above, at
of course, it goes without saying that this assumes everyone’s using
pushal (they are).
mov eax, 0xffffffff mov ecx, eax mov edx, eax mov ebx, eax mov ebp, eax mov esi, eax check: mov edi, 0x000 cmp [edi], 0 jne planb mov esp, 0x400 inc edi mov [edi], 0xe7ff6060; pushal, jmp edi jmp edi planb: mov edi, 0x3fb mov [edi], 0xe7ff6060 mov esp, 0x3fa jmp edi
I also added a (pretty redundant) check to see if the stuff at
0, since the entire arena is initially
0x0. My reasoning, albeit
flawed, was that if it wasn’t 0, then it was unsafe to go there. In
hindsight, it would’ve been safer, since it’s already been written
over by somebody. In any case,
planb never got executed because of
what I’d mentioned earlier — everyone writes from
0x000, for that matter. So I’m relatively safer than
I was in day 1.
These changes paid off, though. I placed 4th on day 2, out of 13 contestants! This screenshot was taken on my phone as I was eating dinner.
All wasn’t well though — I still lost 4 matches, for the reasons below:
- I’d get snuffed out before my bomb wave from
0x400would reach the opponent.
- I’d end up bombing myself without hitting anyone on the way up.
I needed to add some checks to prevent killing myself in the process of bombing.
mov eax, 0xffffffff mov ecx, eax mov edx, eax mov ebx, eax mov ebp, eax mov esi, eax mov edi, 0x000 mov esp, 0x400 mov [edi], 0x20fc8360 mov [edi+4], 0xff600374 mov [edi+8], 0x0400bce7 mov [edi+12], 0xe7ff0000 jmp edi
If you noticed, the initial payload I’m writing to the address at
is a bit more complex this time — let’s break it down.
0x20fc8360 0xff600374 0x0400bce7 0xe7ff0000
This translates to:
60 pushal 83 FC 20 cmp esp, 0x20 74 03 je 9 60 pushal FF E7 jmp edi BC 04 00 00 00 mov esp, 0x400; <- 0x9 FF E7 jmp edi
I check if the stack pointer is
0x20 (decrements from
0x400 due to
pushal); if yes, reset to
0x400, else continue looping. This
prevented me from writing myself over, and also resets bombing from
0x400 — better chance of hitting someone we missed in our first wave.
Sounds good? That’s what I thought too. Day 3 had a bunch of new bot
submissions (and some updated submissions), and a lot of them checked
0x000 for existence of a bot, effectively recking me. I placed 8th out
of 14 contestants, with 7 wins and 6 losses. Tough day.
day 4: the finals
I spent a lot of time refactoring my bot. I tried all kinds of things,
even reworked it to be entirely mobile using the
trick, but I just wasn’t satisfied. In the end, I decided to address the
problem in the simplest way possible. You’re checking
I’ll reposition my initial payload to
And this surprisingly tiny change landed me in 4th place out of 15
contestants, which was way better than I’d anticipated! The top spots
were all claimed by ARM, and naturally so — they had a potential
throughput of 64 bytes per cycle thanks to
stmia, compared to x86’s 32
bytes. Pretty neat!
links and references
- Anisse’s r2wars 2019 post
- Emile’s intro to r2wars
- How not to suck at r2wars
- r2wars: Shall we play a game?
- Shell Storm’s online (dis)assembler
- r2wars game engine
- Anisse’s bot workspace
- My bot dev workspace
- r2con YouTube
This was my first ever r2wars, and it was an incredible experience. Who woulda thunk staring at colored boxes on the screen would be so much fun?! So much so that my parents walked over to see what all the fuss was about. Shoutout to Abel and pancake for taking the time out to work on this, and especially Abel for dealing with all the last minute updates and bot submissions!
All things said, mine was still the best x86 bot — so that’s a win. ;)