6.858 Spring 2020 Lab 1: Buffer overflows

Handed out:	Monday, February 3, 2020
Part 1 due:	Friday, February 7, 2020 (5:00pm)
Part 2 due:	Friday, February 14, 2020 (5:00pm)
All parts due:	Friday, February 21, 2020 (5:00pm)

Introduction

You will do a sequence of labs in 6.858. These labs will give you practical experience with common attacks and counter-measures. To make the issues concrete, you will explore the attacks and counter-measures in the context of the zoobar web application in the following ways:

Lab 1 will introduce you to buffer overflow vulnerabilities, in the context of a web server called zookws. The zookws web server runs a simple python web application, zoobar, with which users transfer "zoobars" (credits) between each other. You will find buffer overflows in the zookws web server code, write exploits for the buffer overflows to inject code into the server over the network, and figure out how to bypass non-executable stack protection. Later labs look at other security aspects of the zoobar and zookws infrastructure.

Each lab requires you to learn a new programming language or some other piece of infrastructure. For example, in this lab you must become intimately familiar with certain aspects of the C language, x86 assembly language, gdb, etc. Detailed familiarity with many different pieces of infrastructure is needed to understand attacks and defenses in realistic situations: security weaknesses often show up in corner cases, and so you need to understand the details to craft exploits and design defenses for those corner cases. These two factors (new infrastructure and details) can make the labs time consuming. You should start early on the labs and work on them daily for some limited time (each lab has several exercises), instead of trying to do all exercises in a single shot just before the deadline. Take the time to understand the relevant details. If you get stuck, post a question on Piazza.

Several labs, including this lab, ask you to design exploits. These exploits are realistic enough that you might be able to use them for a real attack, but you should not do so. The point of the designing exploits is to teach you how to defend against them, not how to use them---attacking computer systems is illegal (see MIT network rules) and can get you into serious trouble. Don't do it.

NOTE: Since we re-use the same lab assignments across years, we ask that you please do not make your lab code publicly accessible (e.g., by checking your solutions into a public repository on GitHub). This helps keep the labs fair and interesting for students in future years.

Lab infrastructure

Exploiting buffer overflows requires precise control over the execution environment. A small change in the compiler, environment variables, or the way the program is executed can result in slightly different memory layout and code structure, thus requiring a different exploit. For this reason, this lab uses a virtual machine to run the vulnerable web server code.

To start working on this lab assignment, you'll need software that lets you run a virtual machine. For Linux users, we recommend running the course VM on KVM, which is built into the Linux kernel. KVM should be available through your distribution, and is preinstalled on Athena cluster computers; on Debian or Ubuntu, try apt-get install qemu-kvm. KVM requires hardware virtualization support in your CPU, and you must enable this support in your BIOS (which is often, but not always, the default). If you have another virtual machine monitor installed on your machine (e.g., VMware), that virtual machine monitor may grab the hardware virtualization support exclusively and prevent KVM from working.

Once you have virtual machine software installed on your machine, you should download the course VM image, and unpack it on your computer. This virtual machine contains an installation of Ubuntu 18.04 Linux.

To start the course VM using VMware, import 6.858-x86_64-v20.vmdk. Go to File > New, select "create a custom virtual machine", choose Linux > Debian 9.x 64-bit, choose Legacy BIOS, and use an existing virtual disk (and select the 6.858-x86_64-v20.vmdk file, choosing the "Take this disk away" option). Finally, click Finish to complete the setup.

To start the VM with kvm, run ./6.858-x86_64-v20.sh from a terminal (Ctrl+A x to force quit). If you get a permission denied error from this script, try adding yourself to the kvm group with sudo gpasswd -a `whoami` kvm, then log out and log back in.

You will use the student account in the VM for your work. The password for the student account is 6858. You can also get access to the root account in the VM using sudo; for example, you can install new software packages using sudo apt-get install pkgname.

You can either log into the virtual machine using its console, or use ssh to log into the virtual machine over the (virtual) network. The latter also lets you easily copy files into and out of the virtual machine with scp or rsync. How you access the virtual machine over the network depends on how you're running it. If you're using VMWare, you'll first have to find the virtual machine's IP address. To do so, log in on the console, run ip addr show dev eth0, and note the IP address listed beside inet. With kvm, you can use localhost as the IP address for ssh and HTTP. You can now log in with ssh by running the following command from your host machine: ssh -p 2222 student@IPADDRESS. To avoid having to type the password each time, you may want to set up an SSH Key.

Getting started

The course Git repository is available at https://web.mit.edu/6858/2020/lab.git. To get the lab code, log into the VM using the student account and clone the source code for lab 1 as follows:

Before you proceed with this lab assignment, make sure you can compile the zookws web server:

The component of zookws that receives HTTP requests is zookd. It is written in C and serves static files and executes dynamic scripts. For this lab you don't have to understand the dynamic scripts; they are written in Python and the exploits in this lab apply only to C code. The HTTP-related code is in http.c. Here is a tutorial about the HTTP protocol.

In order to run the web server in a predictable fashion---so that its stack and memory layout is the same every time---you will use the clean-env.sh script. This is the same way in which we will run the web server during grading, so make sure all of your exploits work on this configuration!

The reference binaries of zookd are provided in bin.tar.gz, which we will use for grading. Make sure your exploits work on those binaries. The make check command will always use both clean-env.sh and bin.tar.gz to check your submission.

Now, make sure you can run the zookws web server and access the zoobar web application from a browser running on your machine, as follows:

The ./clean-env.sh commands starts zookd on port 8080. To open the zoobar application, open your browser and point it at the URL http://IPADDRESS:8080/, where IPADDRESS is the VM's IP address we found above. If something doesn't seem to be working, try to figure out what went wrong, or contact the course staff, before proceeding further.

Part 1: Finding buffer overflows

In the first part of this lab assignment, you will find buffer overflows in the provided web server. To do this lab, you will need to understand the basics of buffer overflows. To help you get started with this, you should read Smashing the Stack in the 21st Century, which goes through the details of how buffer overflows work, and how they can be exploited.

Exercise 1. Study the web server's C code (in zookd.c and http.c), and find one example of code that allows an attacker to overwrite the return address of a function. Hint: look for buffers allocated on the stack. Write down a description of the vulnerability in the file answers.txt. For your vulnerability, describe the buffer which may overflow, how you would structure the input to the web server (i.e., the HTTP request) to overflow the buffer and overwrite the return address, and the call stack that will trigger the buffer overflow (i.e., the chain of function calls starting from process_client).

It is worth taking your time on this exercise and familiarizing yourself with the code, because your next job is to exploit the vulnerability you identified. In fact, you may want to go back and forth between this exercise and Exercises 2 and 3, as you work out the details and document them. That is, if you find a buffer overflow that you think can be exploited, you can use Exercises 2 and 3 to figure out if it indeed can be exploited. It will be helpful to draw a stack diagram like the figures in Smashing the Stack in the 21st Century.

Now, you will start developing exploits to take advantage of the buffer overflows you have found above. We have provided template Python code for an exploit in /home/student/lab/exploit-template.py, which issues an HTTP request. The exploit template takes two arguments, the server name and port number, so you might run it as follows to issue a request to zookws running on localhost:

You are free to use this template, or write your own exploit code from scratch. Note, however, that if you choose to write your own exploit, the exploit must run correctly inside the provided virtual machine.

You may find gdb useful in building your exploits (though it is not required for you to do so). As zookd forks off many processes (one for each client), it can be difficult to debug the correct one. The easiest way to do this is to run the web server ahead of time with clean-env.sh and then attaching gdb to an already-running process with the -p flag. You can find the PID of a process by using pgrep; for example, to attach to zookd-exstack, start the server and, in another shell, run

Keep in mind that a process being debugged by gdb will not get killed even if you terminate the parent zookd process using ^C. If you are having trouble restarting the web server, check for leftover processes from the previous run, or be sure to exit gdb before restarting zookd. You can also save yourself some typing by using b instead of break, and c instead of continue.

When a process being debugged by gdb forks, by default gdb continues to debug the parent process and does not attach to the child. Since zookd forks a child process to service each request, you may find it helpful to have gdb attach to the child on fork, using the command set follow-fork-mode child. We have added that command to /home/student/lab/.gdbinit, which will take effect if you start gdb in that directory.

For this and subsequent exercises, you may need to encode your attack payload in different ways, depending on which vulnerability you are exploiting. In some cases, you may need to make sure that your attack payload is URL-encoded; that is, use + instead of space and %2b instead of +. Here is a URL encoding reference and a handy conversion tool. You can also use quoting functions in the python urllib module to URL-encode bytes (in particular, urllib.parse.quote_from_bytes, followed by .encode('ascii') to get the bytes from the string). In other cases, you may need to include binary values into your payload. The Python struct module can help you do that. For example, struct.pack("<Q", x) will produce an 8-byte (64-bit) binary encoding of the integer x.

Exercise 2. Write an exploit that uses a buffer overflow to crash the web server (or one of the processes it creates). You do not need to inject code at this point. Verify that your exploit crashes the server by checking the last few lines of dmesg | tail, using gdb, or observing that the web server crashes (i.e., it will print Child process 9999 terminated incorrectly, receiving signal 11)

Provide the code for the exploit in a file called exploit-2.py.

The vulnerability you found in Exercise 1 may be too hard to exploit. Feel free to find and exploit a different vulnerability.

Submit your answers to the first part of this lab assignment by running make submit-a. Alternatively, run make prepare-submit-a and upload the resulting lab1a-handin.tar.gz file to the submission web site.

Part 2: Code injection

In this part, you will use your buffer overflow exploits to inject code into the web server. The goal of the injected code will be to unlink (remove) a sensitive file on the server, namely /home/student/grades.txt. Use zookd-exstack, since it has an executable stack that makes it easier to inject code. The zookws web server should be started as follows.

You can build the exploit in two steps. First, write the shell code that unlinks the sensitive file, namely /home/student/grades.txt. Second, embed the compiled shell code in an HTTP request that triggers the buffer overflow in the web server.

When writing shell code, it is often easier to use assembly language rather than higher-level languages, such as C. This is because the exploit usually needs fine control over the stack layout, register values and code size. The C compiler will generate additional function preludes and perform various optimizations, which makes the compiled binary code unpredictable.

We have provided shell code for you to use in /home/student/lab/shellcode.S, along with Makefile rules that produce /home/student/lab/shellcode.bin, a compiled version of the shell code, when you run make. The provided shell code is intended to exploit setuid-root binaries, and thus it runs a shell. You will need to modify this shell code to instead unlink /home/student/grades.txt.

To help you develop your shell code for this exercise, we have provided a program called run-shellcode that will run your binary shell code, as if you correctly jumped to its starting point. For example, running it on the provided shell code will cause the program to execve("/bin/sh"), thereby giving you another shell prompt:

Exercise 3 (warm-up). Modify shellcode.S to unlink /home/student/grades.txt. Your assembly code can either invoke the SYS_unlink system call, or call the unlink() library function.

You may find strace useful when trying to figure out what system calls your shellcode is making. Much like with gdb, you attach strace to a running program:

It will then print all of the system calls that program makes. If your shell code isn't working, try looking for the system call you think your shell code should be executing (i.e., unlink), and see whether it has the right arguments.

Next, we construct a malicious HTTP request that injects the compiled byte code to the web server, and hijack the server's control flow to run the injected code. When developing an exploit, you will have to think about what values are on the stack, so that you can modify them accordingly.

When you're constructing an exploit, you will often need to know the addresses of specific stack locations, or specific functions, in a particular program. One way to do this is to add printf() statements to the function in question. For example, you can use printf("Pointer: %p\n", &x); to print the address of variable x or function x. However, this approach requires some care: you need to make sure that your added statements are not themselves changing the stack layout or code layout. We (and make check) will be grading the lab without any printf statements you may have added.

A more fool-proof approach to determine addresses is to use gdb. For example, suppose you want to know the stack address of the pn[] array in the http_serve function in zookd-exstack, and the address of its saved return pointer. You can obtain them using gdb by first starting the web server (remember clean-evn!), and then attaching gdb to it:

Be sure to run gdb from the ~/lab directory, so that it picks up the set follow-fork-mode child command from ~/lab/.gdbinit. Now you can issue an HTTP request to the web server, so that it triggers the breakpoint, and so that you can examine the stack of http_serve.

From this, you can tell that, at least for this invocation of http_serve, the pn[] buffer on the stack lives at address 0x7fffffffd4a0, and the saved value of %rip (the return address in other words) is at 0x7fffffffdcc8. If you want to see register contents, you can also use info registers.

Exercise 3. Starting from one of your exploits from Exercise 2, construct an exploit that hijacks the control flow of the web server and unlinks /home/student/grades.txt. Save this exploit in a file called exploit-3.py.

Verify that your exploit works; you will need to re-create /home/student/grades.txt after each successful exploit run.

Suggestion: first focus on obtaining control of the program counter. Sketch out the stack layout that you expect the program to have at the point when you overflow the buffer, and use gdb to verify that your overflow data ends up where you expect it to. Step through the execution of the function to the return instruction to make sure you can control what address the program returns to. The next, stepi, and x commands in gdb should prove helpful.

Once you can reliably hijack the control flow of the program, find a suitable address that will contain the code you want to execute, and focus on placing the correct code at that address---e.g. a derivative of the provided shell code.

The test either prints "PASS" or "FAIL". We will grade your exploits in this way. Do not change the Makefile.

The standard C compiler used on Linux, gcc, implements a version of stack canaries (called SSP). You can explore whether GCC's version of stack canaries would or would not prevent a given vulnerability by using the SSP-enabled versions of zookd: zookd-withssp.

Submit your answers to the first two parts of this lab assignment by running make submit-b. Alternatively, run make prepare-submit-b and upload the resulting lab1b-handin.tar.gz file to the submission web site.

Part 3: Return-to-libc attacks

Many modern operating systems mark the stack non-executable in an attempt to make it more difficult to exploit buffer overflows. In this part, you will explore how this protection mechanism can be circumvented. Run the web server configured with binaries that have a non-executable stack, as follows.

The key observation to exploiting buffer overflows with a non-executable stack is that you still control the program counter, after a ret instruction jumps to an address that you placed on the stack. Even though you cannot jump to the address of the overflowed buffer (it will not be executable), there's usually enough code in the vulnerable server's address space to perform the operation you want.

Thus, to bypass a non-executable stack, you need to first find the code you want to execute. This is often a function in the standard library, called libc, such as execve, system, or unlink. Then, you need to arrange for the stack and registers to be in a state consistent with calling that function with the desired arguments. Finally, you need to arrange for the ret instruction to jump to the function you found in the first step. This attack is often called a return-to-libc attack.

One challenge with return-to-libc attacks is that you need to pass arguments to the libc function that you want to invoke. The x86-64 calling conventions make this a challenge because the first 6 arguments are passed in registers. For example, the first argument must be in register %rdi (see man 2 syscall, which documents the calling convention). So, you need an instruction that loads the first argument into %rdi. In Exercise 3, you could have put that instruction in the buffer that your exploit overflows. But, in this part of the lab, the stack is marked non-executable, so executing the instruction would crash the server, but wouldn't execute the instruction.

The solution to this problem is to find a piece of code in the server that loads an address into %rdi. Such a piece of code is referred to as a "borrowed code chunk", or more generally as a rop gadget, because it is a tool for return-oriented programming (rop). Luckily, zookd.c accidentally has a useful gadget: see the function accidentally.

Exercise 4. Starting from your exploit in Exercise 2 and 3, construct an exploit that unlinks /home/student/grades.txt when run on the binaries that have a non-executable stack. Name this new exploit exploit-4.py.

In this attack you are going to take control of the server over the network without injecting any code into the server. You should use a return-to-libc attack where you redirect control flow to code that already existed before your attack. The outline of the attack is to perform a buffer overflow that:

causes the argument to the chosen libc function to be on stack
then causes accidentally to run so that argument ends up in %rdi
and then causes accidentally to return to the chosen libc function

It will be helpful to draw a stack diagram like the figures in Smashing the Stack in the 21st Century at (1) the point that the buffer overflows and (2) at the point that accidentally runs.

Challenge! (optional) The accidentally function is a bit artificial. For extra credit, figure out how to perform the return-to-libc attack without relying on that function (delete it and find another way to make your exploit work). Provide your attack in exploit-challenge.py. Also, briefly explain the attack and provide ROP gadgets you use in answers.txt.

You will need to find another chunk of code to reuse that gives you control over %rdi. You can read through the disassembly (e.g. using objdump) to look for useful ROP gadgets.

Because of the nature of x86/x86-64, you can use another technique to find sequences of instructions that don't even appear in the disassembly! Instructions are variable-length (from 1 to 15 bytes), and by causing a misaligned parse (by jumping into the middle of an intended instruction), you can cause a sequence of machine code to be misinterpreted. For example, the instruction sequence pop %r15; ret corresponds to the machine code 41 5F C3. But instead of executing from the start of this instruction stream, if you jump 1 byte in, the machine code 5F C3 corresponds to the assembly pop %rdi; ret.

Automated tools such as ROPgadget.py can assist you in searching for ROP gadgets, even finding gadgets that arise from misaligned parses.

Part 4: Fixing buffer overflows and other bugs

Now that you have figured out how to exploit buffer overflows, you will try to find other kinds of vulnerabilities in the same code. As with many real-world applications, the "security" of our web server is not well-defined. Thus, you will need to use your imagination to think of a plausible threat model and policy for the web server.

Exercise 5. Look through the source code and try to find more vulnerabilities that can allow an attacker to compromise the security of the web server. Describe the attacks you have found in answers.txt, along with an explanation of the limitations of the attack, what an attacker can accomplish, why it works, and how you might go about fixing or preventing it. You should ignore bugs in zoobar's code. They will be addressed in future labs.

One approach for finding vulnerabilities is to trace the flow of inputs controlled by the attacker through the server code. At each point that the attacker's input is used, consider all the possible values the attacker might have provided at that point, and what the attacker can achieve in that manner.

You should find at least two vulnerabilities for this exercise.

Finally, you will fix the vulnerabilities that you have exploited in this lab assignment.

Exercise 6. For each buffer overflow vulnerability you have exploited in Exercises 2, 3, and 4, fix the web server's code to prevent the vulnerability in the first place. Do not rely on compile-time or runtime mechanisms such as stack canaries, removing -fno-stack-protector, baggy bounds checking, etc.

Make sure that your code actually stops your exploits from working. Use make check-fixed to run your exploits against your modified source code (as opposed to the staff reference binaries from bin.tar.gz). These checks should report FAIL (i.e., exploit no longer works). If they report PASS, this means the exploit still works, and you did not correctly fix the vulnerability.

Note that your submission should not make changes to the Makefile and other grading scripts. We will use our unmodified version during grading.

You should also make sure your code still passes all tests using make check, which uses the unmodified lab binaries.

You are done! Submit your answers to the lab assignment by running make submit. Alternatively, run make prepare-submit and upload the resulting lab1-handin.tar.gz file to the submission web site.