depot/third_party/tvl/users/tazjin/presentations/bootstrapping-2018/presentation.tex
Default email c4fb0432ae Project import generated by Copybara.
GitOrigin-RevId: 3fc1143a04da49a92c3663813c6a0c1e8ccd477f
2020-09-29 23:42:59 -04:00

251 lines
6.4 KiB
TeX

\documentclass[12pt]{beamer}
\usetheme{metropolis}
\newenvironment{code}{\ttfamily}{\par}
\title{Where does \textit{your} compiler come from?}
\date{2018-03-13}
\author{Vincent Ambo}
\institute{Norwegian Unix User Group}
\begin{document}
\maketitle
%% Slide 1:
\section{Introduction}
%% Slide 2:
\begin{frame}{Chicken and egg}
Self-hosted compilers are often built using themselves, for example:
\begin{itemize}
\item C-family compilers bootstrap themselves \& each other
\item (Some!) Common Lisp compilers can bootstrap each other
\item \texttt{rustc} bootstraps itself with a previous version
\item ... same for many other languages!
\end{itemize}
\end{frame}
\begin{frame}{Chicken, egg and ... lizard?}
It's not just compilers: Languages have runtimes, too.
\begin{itemize}
\item JVM is implemented in C++
\item Erlang-VM is C
\item Haskell runtime is C
\end{itemize}
... we can't ever get away from C, can we?
\end{frame}
%% Slide 3:
\begin{frame}{Trusting Trust}
\begin{center}
\huge{Could this be exploited?}
\end{center}
\end{frame}
%% Slide 4:
\begin{frame}{Short interlude: A quine}
\begin{center}
\begin{code}
((lambda (x) (list x (list 'quote x)))
\newline\vspace*{6mm} '(lambda (x) (list x (list 'quote x))))
\end{code}
\end{center}
\end{frame}
%% Slide 5:
\begin{frame}{Short interlude: Quine Relay}
\begin{center}
\includegraphics[
keepaspectratio=true,
height=\textheight
]{quine-relay.png}
\end{center}
\end{frame}
%% Slide 6:
\begin{frame}{Trusting Trust}
An attack described by Ken Thompson in 1983:
\begin{enumerate}
\item Modify a compiler to detect when it's compiling itself.
\item Let the modification insert \textit{itself} into the new compiler.
\item Add arbitrary attack code to the modification.
\item \textit{Optional!} Remove the attack from the source after compilation.
\end{enumerate}
\end{frame}
%% Slide 7:
\begin{frame}{Damage potential?}
\begin{center}
\large{Let your imagination run wild!}
\end{center}
\end{frame}
%% Slide 8:
\section{Countermeasures}
%% Slide 9:
\begin{frame}{Diverse Double-Compiling}
Assume we have:
\begin{itemize}
\item Target language compilers $A$ and $T$
\item The source code of $A$: $ S_{A} $
\end{itemize}
\end{frame}
%% Slide 10:
\begin{frame}{Diverse Double-Compiling}
Apply the first stage (functional equivalence):
\begin{itemize}
\item $ X = A(S_{A})$
\item $ Y = T(S_{A})$
\end{itemize}
Apply the second stage (bit-for-bit equivalence):
\begin{itemize}
\item $ V = X(S_{A})$
\item $ W = Y(S_{A})$
\end{itemize}
Now we have a new problem: Reproducibility!
\end{frame}
%% Slide 11:
\begin{frame}{Reproducibility}
Bit-for-bit equivalent output is hard, for example:
\begin{itemize}
\item Timestamps in output artifacts
\item Non-deterministic linking order in concurrent builds
\item Non-deterministic VM \& memory states in outputs
\item Randomness in builds (sic!)
\end{itemize}
\end{frame}
\begin{frame}{Reproducibility}
\begin{center}
Without reproducibility, we can never trust that any shipped
binary matches the source code!
\end{center}
\end{frame}
%% Slide 12:
\section{(Partial) State of the Union}
\begin{frame}{The Desired State}
\begin{center}
\begin{enumerate}
\item Full-source bootstrap!
\item All packages reproducible!
\end{enumerate}
\end{center}
\end{frame}
%% Slide 13:
\begin{frame}{Bootstrapping Debian}
\begin{itemize}
\item Sparse information on the Debian-wiki
\item Bootstrapping discussions mostly resolve around new architectures
\item GCC is compiled by depending on previous versions of GCC
\end{itemize}
\end{frame}
\begin{frame}{Reproducing Debian}
Debian has a very active effort for reproducible builds:
\begin{itemize}
\item Organised information about reproducibility status
\item Over 90\% reproducibility in Debian package base!
\end{itemize}
\end{frame}
\begin{frame}{Short interlude: Nix}
\begin{center}
\includegraphics[
keepaspectratio=true,
height=0.7\textheight
]{nixos-logo.png}
\end{center}
\end{frame}
\begin{frame}{Short interlude: Nix}
\begin{center}
\includegraphics[
keepaspectratio=true,
height=0.90\textheight
]{drake-meme.png}
\end{center}
\end{frame}
\begin{frame}{Short interlude: Nix}
\begin{center}
\includegraphics[
keepaspectratio=true,
height=0.7\textheight
]{nixos-logo.png}
\end{center}
\end{frame}
\begin{frame}{Bootstrapping NixOS}
Nix evaluation can not recurse forever: The bootstrap can not
simply depend on a previous GCC.
Workaround: \texttt{bootstrap-tools} tarball from a previous
binary cache is fetched and used.
An unfortunate magic binary blob ...
\end{frame}
\begin{frame}{Reproducing NixOS}
Not all reproducibility patches have been ported from Debian.
However: Builds are fully repeatable via the Nix fundamentals!
\end{frame}
\section{Future Developments}
\begin{frame}{Bootstrappable: stage0}
Hand-rolled ``Cthulhu's Path to Madness'' hex-programs:
\begin{itemize}
\item No non-auditable binary blobs
\item Aims for understandability by 70\% of programmers
\item End goal is a full-source bootstrap of GCC
\end{itemize}
\end{frame}
\begin{frame}{Bootstrappable: MES}
Bootstrapping the ``Maxwell Equations of Software'':
\begin{itemize}
\item Minimal C-compiler written in Scheme
\item Minimal Scheme-interpreter (currently in C, but intended to
be rewritten in stage0 macros)
\item End goal is full-source bootstrap of the entire GuixSD
\end{itemize}
\end{frame}
\begin{frame}{Other platforms}
\begin{itemize}
\item Nix for Darwin is actively maintained
\item F-Droid Android repository works towards fully reproducible
builds of (open) Android software
\item Mobile devices (phones, tablets, etc.) are a lost cause at
the moment
\end{itemize}
\end{frame}
\begin{frame}{Thanks!}
Resources:
\begin{itemize}
\item bootstrappable.org
\item reproducible-builds.org
\end{itemize}
@tazjin | mail@tazj.in
\end{frame}
\end{document}