\documentclass[conference]{IEEEtran} \IEEEoverridecommandlockouts{} % The preceding line is only needed to identify funding in the first footnote. % If that is unneeded, please comment it out. \usepackage{csquotes} \usepackage[style=ieee,backend=biber]{biblatex} \addbibresource{./bibliography.bib} \usepackage{amsmath,amssymb,amsfonts} \usepackage{algorithmic} \usepackage{booktabs} \usepackage{graphicx} \usepackage{textcomp} \usepackage{xcolor} \usepackage{caption} \usepackage{subcaption} % code listings \usepackage{minted} \usepackage{relsize} % acronyms \usepackage{acro} \acsetup{list-long-format=\capitalisewords} %additional packages %\usepackage[ngerman]{babel} \usepackage[utf8]{inputenc} \usepackage{hyperref} \usepackage{cleveref} \usepackage{url} %%fuer abkuerzungen begin \usepackage[acronym,hyperfirst = false]{glossaries} \glsdisablehyper{} %\usepackage[acronym,acronymlists={main, %abbreviationlist},shortcuts,toc,description,footnote]{glossaries} \newglossary[clg]{abbreviationlist}{cyi}{cyg}{List of Abbreviations} \newglossary[slg]{symbolslist}{syi}{syg}{Symbols} \renewcommand{\firstacronymfont}[1]{\emph{#1}} \renewcommand*{\glspostdescription}{} % Punkt am Ende jeder Beschreibung entfernen \renewcommand*{\acrnameformat}[2]{#2 (\acronymfont{#1})} % Langform der Akronyme \makeglossaries{} \date{\today} \input{glossary} %%fuer abkuerzungen end \include{acronyms} \begin{document} \title{Overview Over Attack Vectors and Countermeasures for Buffer Overflows} \author{\IEEEauthorblockN{Valentin Brandl} \IEEEauthorblockA{\textit{Faculity of Computer Science and Mathematics} \\ \textit{OTH Regensburg}\\ Regensburg, Germany \\ valentin.brandl@st.oth-regensburg.de\\ MatrNr. 3220018} } \maketitle \begin{abstract} This paper tries to explain the details behind buffer overflows, explore the problems stemming from those kinds of software vulnerabilities and discus possible countermeasures with focus on their effectiveness, performance impact and ease of use. It discusses compiler based (such as ASLR, NX, stack canaries) as well as type system based (e.g.\ dependent types) solutions to this prevalent type of software bugs based on their performance impact and the effort needed to introduce the mitigations into existing software projects. An analysis of the current state of the art informs the reader about what to expect when writing software today. The analysis shows that most techniques actually tackle the problem of exploiting buffer overflows for code execution but do nothing to prevent introducing them in the first place. \end{abstract} \begin{IEEEkeywords} Buffer Overflow, Software Security \end{IEEEkeywords} \section{Motivation}\label{ref:motivation} In the early days of programming, memory as managed manually to make the best use of slow hardware and low memory. This opened the door for many kinds of programming errors. Memory can be deallocated more than once (double-free), invalid pointers can be dereferenced (\mintinline{C}{NULL} pointer dereference; this is still a problem in many modern languages) or the program could read or write out of bounds of a buffer (information leaks, \acp{bof}). Languages that are affected by this are e.g.\ C, C++ and Fortran. While modern programming languages solve most if not all of these problems, critical parts of the worlds infrastructure are still implemented in these old languages, either because they allow the implementation of really performant programs, offer deterministic runtime behaviour (e.g.\ no pauses due to garbage collection), because they power legacy systems or for portability reasons. Scientists and software engineers have proposed lots of solutions to this problem over the years and this paper aims to compare and give an overview about those. Reading out of bounds can result in an information leak and is one of the less critical results of \ac{bof} in most cases, but there are exceptions, e.g.\ the Heartbleed bug~\cite{Heardbleed2014} in OpenSSL which allowed dumping secret keys from memory. Out of bounds writes are almost always critical and result in code execution vulnerabilities or at least application crashes. In 2018, 14\% (2368 out of 16556)~\cite{Cve2018} of all software vulnerabilities that have a CVE assigned, were overflow related. This shows that, even if this type of bug is very old and well known, it's still relevant today. \section{Background}\label{ref:background} \subsection{Technical Details} Code execution via \ac{bof} vulnerabilities almost always works by overwriting the return address in the current stack frame (known as \enquote{stack smashing})~\cite{Smashing2004}, so when the \mintinline{ASM}{RET} instruction is executed, an attacker controlled address is moved into the \ac{ip} register and the code pointed to by this address is executed~\cite{Detection2018}. Other ways include overwriting addresses in the \ac{plt} (the \ac{plt} contains addresses of dynamically linked library functions) of a binary so that, if a linked function is called, an attacker controlled function is called instead, or (in C++) overwriting the \ac{vmt}, which stores the pointers to an object's methods. A simple vulnerable C program might look like this: \begin{figure}[h!] \begin{minted}{c} void vuln(char *input) { char buf[50]; size_t len = strlen(input); for (size_t i = 0; i < len; i++) { buf[i] = input[i]; } } int main(int argc, char **argv) { vuln(argv[1]); return 0; } \end{minted} \caption{Vulnerable C program}\label{lst:vuln} \end{figure} A successful stack \ac{bof} exploit would place the payload in the memory by supplying it as an argument to the program (or by placing it in an environment variable, writing it to a file that the program reads, via network packet, \dots) and eventually overwrite the return address by providing an input with more than 50 bytes and therefore writing out of bounds. When executing the \mintinline{C}{return} instruction, and the \ac{ip} jumps into the payload, the attacker's code is executed. This works due to the way, how CPUs perform function calls: The stack frame of the current function lies between the \ac{bp} and \ac{sp} as shown in~\cref{fig:before}. When calling a function, the value of the \ac{bp} and \ac{ip} is pushed to the stack (\cref{fig:call}) and the CPU writes the address of the called function into the \ac{ip}. When the function returns, after restoring the old \ac{ip} from the stack, the execution continues from where the function call occurred earlier. If an overflow overwrites the old \ac{ip} (\cref{fig:exploit}), the attacker controls where execution continues. \begin{figure}[h!] \begin{subfigure}[b]{.3\textwidth} \includegraphics[width=\textwidth]{./dot/before.pdf} \caption{Stack layout before function call}\label{fig:before} \end{subfigure}\\ \begin{subfigure}[b]{.3\textwidth} \includegraphics[width=\textwidth]{./dot/call.pdf} \caption{Stack layout after function call}\label{fig:call} \end{subfigure}\\ \begin{subfigure}[b]{.3\textwidth} \includegraphics[width=\textwidth]{./dot/exploit.pdf} \caption{Stack layout after overflow}\label{fig:exploit} \end{subfigure} \caption{Stack layouts during an \ac{bof} exploit} \end{figure}% This is only one of several types and exploitation techniques. Others include \begin{itemize} \item Heap-based \ac{bof}: In this case there is no way of overwriting the return address but objects on the heap might contain function pointers (e.g.\ for dynamic dispatch) which can be overwritten to execute the attackers code, when called~\cite{Detection2018}. \item Integer overflow: Some calculation on fixed sized integers is used to allocate memory. The calculation leads to an integer overflow and only a small buffer is allocated~\cite{Detection2018}. Later the buffer is indexed with a big integer and performs a read or write outside the buffer. This kind of vulnerability can also lead to other problems because at least in C, signed integer overflow is undefined behaviour. \end{itemize} This paper does not explore other kinds of \ac{bof} in detail because the concept is always the same: Unchecked indexing into memory allows the attacker to overwrite some kind of return or call address, which allows hijacking of the execution flow. The most trivial kind of payloads is known as a \mintinline{ASM}{NOP} sled. Here the attacker appends as many \mintinline{ASM}{NOP} instructions before any shell-code (e.g.\ to invoke \mintinline{shell}{/bin/sh}) and points the overwritten \ac{ip} or function pointer somewhere inside the \mintinline{ASM}{NOP}s. The execution \enquote{slides} (hence the name) through the \mintinline{ASM}{NOP}s until it reaches the shell-code. Most of the mitigation techniques described in this paper protect against this kind of exploit but there are different and more complex ways of exploiting \acp{bof} that are not that easily mitigated. \section{Concept and Methods}\label{ref:concept} \subsection{Research Methods} This paper describes several techniques that have been proposed to mitigate the problems introduced by \acp{bof} and tries to answer the following questions: \begin{itemize} \item What is the performance impact? \item How effective is the technique? Did it actually prevent exploitation of \acp{bof}? \item How realistic is it for developers to use the technique in real-world code? Is an incremental introduction possible? \end{itemize} The paper focuses on solutions for the C language, since it is still the second most used language as of December 2019~\cite{Tiobe2019}. Some of the described techniques are language agnostic but this is not a focus of this paper. In the end, there is a discussion about the current state. For the literature research, the paper~\citetitle{Detection2018} served as a base. From there on, the author performed a snowball system search with combinations of the keywords \enquote{buffer}, \enquote{overflow}, \enquote{detection}, \enquote{prevention} and \enquote{dependent typing} using \url{https://scholar.google.com/}. Evaluation and prioritization of results is done using the following criteria: \begin{itemize} \item Type of publication in the following order: \begin{enumerate} \item conference paper \item unreleased paper \item books \item online sources \end{enumerate} \item Number of citations \item Publisher \item Author's reputation and institute \item Overall quality (first by checking structure and abstract, then by the actual content) \end{itemize} \subsection{\Ac{rbc}} The easiest and maybe single most effective method to prevent \acp{bof} is to check, if a write or read operation is out of bounds. This requires storing the size of a buffer together with the pointer to the buffer (so called fat pointers) and check for each read or write in the buffer, if it is in bounds at runtime. Almost any language that comes with a managed runtime, uses \ac{rbc}. For this technique to be effective in general, writes to raw pointers must be disallowed. Otherwise the security checks can be circumvented. \Ac{rbc} introduces a runtime overhead for every indexed read or write operation. This is a problem if a program runs on limited hardware or might impact real-time properties. Introducing \ac{rbc} into an existing codebase is not easy. Using fat pointers in a few functions does not prevent other parts of the code to use raw pointers into the same buffer. So for this to be effective, the whole codebase needs to be changed to disallow raw pointers, which, depending on the size, might not be feasible. Still, if done correctly and consequently, there will be no \ac{bof} vulnerabilities. \Ac{dos} might is still possible depending on how invalid indexing is handled, because the program might terminate gracefully when a out of bounds index is used. \subsection{Prevent/Detect Overwriting Return Address} Since stack based \ac{bof} exploits work by overwriting the return address in the current stack frame, preventing or at least detecting this, can be quite effective without much overhead at runtime. \citeauthor{Rad2001} describe a technique that stores a redundant copy of the return address in a secure memory area that is guarded by read-only memory, so it cannot be overwritten by overflows. When returning, the copy of the return address is compared to the one in the current stack frame and only if it matches, the \mintinline{ASM}{RET} instruction is actually executed~\cite{Rad2001}. While this is effective against stack based \acp{bof}, in the described form, it does not protect against \ac{vmt} or \ac{plt} overwrites. An extension could be made to also protect the \ac{plt} and \ac{vmt} but custom constructs using function pointers would remain vulnerable. Since this technique is a compiler extension, no modification of the codebase is required to enable it, and while it does not prevent all kinds of \ac{bof}, it mitigates all stack based \acp{bof} with only minimal overhead when calling and returning from a function. An older technique from 1998 proposes to put a canary word (named after the canaries that were used in mines to detect low oxygen levels) between the data of a stack frame and the return address~\cite{Stackguard1998}\cite{AtkDef2016}. When returning, a check is performed, to confirm, the canary is intact, if it is not, a \ac{bof} occurred. This technique is implemented by major compilers~\cite{Gcc2003} but can be defeated, if there is an information leak that leaks the canary to the attacker. The attacker is then able to construct a payload, that keeps the canary intact. This mitigation has a minimal performance impact~\cite{Gcc2003} and offers a good level of protection. It is a compiler extension so there is no need for modification of the code base. \subsection{Type System Solutions} \citeauthor{Dep2007} propose an extension to the C type system that extends it with dependent types. These types have an associated value, e.g.\ a pointer type can have the buffer size associated to it~\cite{Dep2007}. This prevents indexing into a buffer with out-of-bounds values. This extension is a superset of C so compilation of any valid C code is possible using the extension and incremental improvement of the codebase is possible. If the type extension is advanced enough, the additional information might form the base for a formal verification. In some cases, inference of the type extensions is possible~\cite{Dep2007}. This technique prevents all kinds of overflows, if used, but requires changes to the codebase and is only effective where these changes are applied. Since it is a compile-time solution, it affects the compile-time but has no negative effect on the runtime. \subsection{Address Space Layout Randomization} \Ac{aslr} aims to prevent exploitation of \acp{bof} by placing code at random locations in memory~\cite{AtkDef2016}. That way, it is not trivial to set the return address to point to the payload in memory. This is effective against every kind of \ac{bof} vulnerability but it is still possible to exploit \ac{bof} vulnerabilities in combination with information leaks or other techniques like heap spraying. Also on 32 bit systems, the address space is small enough to try a brute-force attempt until the payload in memory is hit~\cite{Effectiveness2014}. This is another technique that works without modification of the code base. Also there is no runtime overhead because nothing changed except the location of the program. \subsection{w\^{}x Memory} w\^{}x (also known as \ac{nx} or \ac{dep}) makes memory either writable or executable~\cite{AtkDef2016}. That way, an attacker cannot place arbitrary payloads in memory. There are still techniques to exploit this by reusing existing executable code. The ret-to-libc exploiting technique uses existing calls to the libc with attacker controlled parameters, e.g.\ if the program uses the \mintinline{shell}{system} command, the attacker can plant \mintinline{shell}{/bin/sh} as parameter on the stack, followed by the address of \mintinline{shell}{system} and get a shell on the system. \Ac{rop} (a superset of ret-to-libc exploits) uses so called \ac{rop} gadgets, combinations of memory modifying instructions followed by the \mintinline{ASM}{RET} instruction to build instruction chains, that execute the desired shell-code. This is achieved by placing the desired return addresses in the right order on the stack and reuses the existing code to circumvent the w\^{}x protection. These combinations of memory modification followed by \mintinline{ASM}{RET} instructions, known as \ac{rop} chains, are Turing complete~\cite{Rop2007}, so in theory it is possible to construct any imaginable payload, as long as the exploited program contains enough gadgets and the overflowing buffer has enough space. \section{Discussion}\label{ref:discussion} \subsection{Effectiveness} \subsubsection{\ac{aslr}} \Ac{aslr} has proven effective and sees wide use in production. Most major operating systems implement this technique~\cite{FBSDaslr}. Some even use kernel \ac{aslr}~\cite{Linuxaslr}. Since this mechanism is active at runtime, it does not require any changes in the code itself, the program only has to be compiled as a \ac{pie}. On 32-bit CPUs, only 16-bit of the address are randomized. These 16-bit can be brute forced in a few minutes or seconds~\cite{AslrEffective2004}. There is no runtime overhead since the only change is the position of the program in memory. This technique can and should be used on modern systems because there is no additional work required, except maybe recompilation. \subsubsection{w\^{}x} The rise of code reuse exploits like \ac{rop} and ret-to-libc, shows the ineffectiveness of w\^{}x protection. It makes vulnerabilities harder to exploit by preventing the most naive types of payloads but it doesn't actually prevent exploits from happening. \Ac{nx} does not prevent any exploits but makes it harder for an attacker that does not know the system, the program is running on (e.g.\ a network service). It has no runtime overhead and is a compile-time option so it does not hurt to enable \ac{nx}. \subsubsection{Runtime Bounds Checks} Checking for overflows at runtime is very effective but can have a huge performance impact so it is not feasible in every case. It also comes with other footguns. There might be integer overflows when calculating the bounds which might introduce other problems. \subsection{State of the Art} Operating systems started to compile C code to \acp{pie} by default~\cite{ArchPie2017} and \ac{aslr} is enabled, too. Same goes for \ac{nx} and stack canaries~\cite{ArchPie2017}. The combination of these mitigations makes it hard to write general exploits for modern operating systems. To check the current state, the author investigates, which mitigations are enabled by default in the latest release (9.2) of the \ac{gcc} and the latest commit of the LLVM-project (\mintinline[breaklines]{shell}{181ab91efc9}) by building both compilers using the default configuration. The experiments are performed on a 64-bit Debian 9.11 system running on version 4.19.0 of the Linux kernel. The following commands compile the source codes: \begin{figure}[h!] \begin{subfigure}[b]{.3\textwidth} \begin{minted}{shell} mkdir objdir \ && cd objdir \ && ../configure \ --build=x86_64-linux-gnu \ --host=x86_64-linux-gnu \ --target=x86_64-linux-gnu \ --disable-multilib \ && make -j8 \end{minted} \caption{\ac{gcc} compilation script}\label{lst:gcc} \end{subfigure} \\ \begin{subfigure}[b]{.3\textwidth} \begin{minted}{shell} mkdir build \ && cd build \ && cmake -DLLVM_ENABLE_PROJECTS=clang \ -DCMAKE_BUILD_TYPE=Release \ -G "Unix Makefiles" ../llvm \ && make -j8 \end{minted} \caption{clang compilation script}\label{lst:clang} \end{subfigure} \end{figure} The \mintinline{shell}{build}, \mintinline{shell}{host} and \mintinline{shell}{target} parameters in~\cref{lst:gcc} describe the target platform for the compiler and \mintinline{shell}{disable-multilib} disables 32-bit support, which is not needed for this experiment. The \mintinline{sh}{-j8} flag only tells make to use all 8 available cores for compilation. \mintinline{shell}{CMAKE_BUILD_TYPE=Release} creates a release build of the clang compiler (see~\cref{lst:clang}). The fresh builds of \ac{gcc} and clang compile the code from~\cref{lst:vuln} to check which mitigations are enabled by default. After using \mintinline[breaklines]{shell}{gcc -o vuln.gcc vuln.c} and \mintinline[breaklines]{shell}{clang -o vuln.clang vuln.c} to compile the source code, the \mintinline{shell}{checksec.sh} tool~\cite{Checksec2019} shows which mitigations are active in the new binary: \begin{table}[h!] \begin{center} \begin{tabular}{lrr} \toprule Mitigation & Active in \ac{gcc}? & Active in clang? \\ \toprule Stack Canary & No & No \\ \midrule \ac{nx} & Yes & Yes \\ \midrule \ac{pie} & No & No \\ \bottomrule \end{tabular} \caption{Enabled mitigations in a default \ac{gcc} and clang build}\label{tab:mitigations} \end{center} \end{table} Surprisingly enough, two of the most popular C compilers enable only one of the described compile-time mitigations by default (see~\cref{tab:mitigations}). Maintainer of operating system packages of the compiler might choose a more secure configuration for the compiler as shown in~\cite{ArchPie2017} but still, compiler vendors might want to choose better defaults, too. So far, all discussed mitigations don't change anything about the existence of \acp{bof} but just try to prevent the exploitation for code execution. The vulnerable programs terminate if the stack canary is overwritten, a call into \ac{nx} memory occurs or execution continues inside garbage data due to \ac{aslr}. The underlying problem persists, only the worst results are mitigated. \Ac{dos} is still a problem in safety critical systems (e.g.\ cars, planes, medical devices) or in any area with real-time requirements. Language extensions to fix the problem of \acp{bof} as described in~\cite{Dep2007} require lots of discipline to use them everywhere. They are only useful if the whole codebase uses the new features. Introducing them in an existing codebase is quite unrealistic since it requires lots of modifications. On the other hand, this actually prevents \acp{bof} from happening and not just from being exploited, so it looks like an interesting concept for safety critical software. \section{Conclusion}\label{ref:conclusion} While there are many techniques, that protect against different types of \acp{bof}, none of them is effective in every situation but in combination they offer good protection against code execution attacks. Maybe the time has come, where usage of memory unsafe languages has to be stopped where it is not inevitable. There are many modern programming languages, that aim for the same problem space as C, C++ or Fortran but without the issues coming from these languages. If it is feasible to use a garbage collector, languages like Go, Java or even scripting languages like Python might work just fine. If real-time properties are required, Rust could be the way to go, without any language runtime and with deterministic memory management. For any other problem, almost any other memory safe language is better than using unsafe C. \printbibliography{} % \bibliographystyle{IEEEtran} % \bibliography{bibliography} % \printacronyms{} \end{document} % vim: set filetype=tex ts=2 sw=2 tw=80 et spell :