Cut out a large part of the content

This commit is contained in:
Théophile Bastian 2019-10-08 12:00:43 +02:00
parent a22b55e086
commit 7cd581fbd0

View file

@ -66,17 +66,14 @@
{\lstinline[language=C, keepspaces=true, basicstyle=\ttfamily]}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\title[\sectionline] {Growing the DWARF tougher:\\
synthesis, validation and compilation}
\author[\slidecountline]{\textbf{Théophile Bastian} \\ \vspace{0.5em}
{{\footnotesize{}Based on work done with}\\
\textbf{Francesco Zappa Nardelli},
\textbf{Stephen Kell},
\textbf{Simon Ser}}}
\title[\sectionline] {Reliable and Fast DWARF-based Stack Unwinding}
\author[\slidecountline]{\textbf{Théophile Bastian},\\
\textbf{Stephen Kell}, \\
\textbf{Francesco Zappa Nardelli}}
\date{}
%\subject{}
%\logo{}
\institute{ENS Paris, INRIA}
\institute{ENS Paris, University of Kent, Inria}
\begin{document}
@ -87,7 +84,7 @@
\vspace{-2em}
\begin{center}
\begin{align*}
\text{Slides: } &\text{\url{https://tobast.fr/files/oracle18.pdf}} \\
\text{Slides: } &\text{\todo{add URL for this PDF}} \\
\end{align*}
\end{center}
\end{frame}
@ -264,31 +261,6 @@ $1 = 84
\end{center}
\end{frame}
\begin{frame}{Difficult to generate}
Major concern with DWARF\@: it is \alert{difficult to generate}
(correctly).
\begin{itemize}
\item{} \alert{Hard to generate}: each compiler pass must keep it in
sync
\item{} Most of it is \alert{seldom used} (\eg{} unwinding data of
dusty code), and thus \alert{seldom tested}
\end{itemize}
\vspace{1em}
Yields to
\begin{itemize}
\item{} unreliable DWARF\@: can cause headaches when debugging
\item{} or not generated at all (eg. OCaml until recently)
\end{itemize}
\vspace{1em}
\begin{center}
\Large\bf
$\leadsto$ Complex, buggy, untested
\end{center}
\end{frame}
\newcommand{\LinusMailOne}{
``Sorry, but last time was too f\dots painful. The whole (and
only) point of unwinders is to make debugging easy
@ -309,18 +281,29 @@ $1 = 84
\begin{frame}{A debugging hell: Linux kernel}
\LinusMailOne{}
\vspace{1em}
\LinusSource{}
\only<1-2>{
\vspace{1em}
\LinusSource{}
}
\pause{}\vspace{1em}
\begin{center}
\Large\bf
\alert{This is where we still are!}
\end{center}
\vspace{1em}
\only<2>{
\begin{center}
\Large\bf
\alert{This is where we still are!}
\end{center}
}
\only<3>{
\LinusMailTwo{}
\vspace{1em}
\LinusSource{}
}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Unwinding data validation}
\section{Unwinding data as an abstract state}
\newcommand{\tblrowval}[4]{#1 & #2 & \only<2->{#3} & \only<2->{#4} \\}
\newcommand{\blknote}[1]
@ -335,7 +318,7 @@ $1 = 84
\end{block}}
\newcommand{\tblhl}{\rowcolor{Tan}}
\begin{frame}{Validating an example}
\begin{frame}{Working on an example}
\newcommand{\firsttblrows}{
\tblrowval{\hspace{-2ex}<{\bf foo}>:}{}{\textbf{CFA}}{\textbf{ra}}
\rowonly<3>{\tblhl{}} \tblrowval{push}{\%r15}{rsp+8}{c-8}
@ -346,7 +329,6 @@ $1 = 84
\tblrowval{push}{\%rbp}{rsp+40}{c-8}
\tblrowval{push}{\%rbx}{rsp+48}{c-8}
\tblrowval{sub}{\$0x68,\%rsp}{rsp+56}{c-8}
\rowonly<9>{\tblhl{}} \tblrowval{cmp}{\$0x1,\%edi}{rsp+160}{c-8}%
}
\only<-8>{
@ -377,154 +359,10 @@ $1 = 84
\only<7>{The unwinding table can actually be seen as\\
an \alert{abstract interpretation} of the code\ldots}
\only<8>{\ldots and thus, for a given run, be
\alert{re-computed for verification}}
\alert{re-computed from scratch}}
\end{overlayarea}
}
}
\only<9->{
\begin{table}
\ttfamily\large
\begin{tabularx}{0.9\linewidth}{
l
b
>{\columncolor{SkyBlue}}s
>{\columncolor{SkyBlue}}s
}
\firsttblrows{}%
\end{tabularx}
\end{table}
\vspace{-0.8em}
\only<9>{\blklnote{If, within an execution,
\begin{itemize}
\item ra = \texttt{*(0xFFFF1098)}
\item \reg{rsp} = \texttt{0xFFFF1000}
\end{itemize}
We can \alert{evaluate both expressions} and \alert{compare}
}}
}
\end{frame}
\begin{frame}{Dynamic validation}
\textbf{Abstract state}
\begin{itemize}
\item \alert{Stack} of actual \alert{addresses} where
\alert{return addresses} are stored
\end{itemize}
\vspace{2em}\pause{}
\textbf{Abstract instruction semantics}
\begin{itemize}
\item[\alert{\texttt{call}}] push \alert{\reg{rsp}} on the stack
\item[\alert{\texttt{ret}}] pop from the stack
\end{itemize}
\vspace{2em}\pause{}
\textbf{Validation of each instruction}
\begin{itemize}
\item Evaluate the return address provided by DWARF
\item Compare it with the value at the top of the stack
\end{itemize}
\end{frame}
\begin{frame}{In practice: \texttt{eh\_frame\_check}}
Strategy implemented and working: \alert{\texttt{eh\_frame\_check}}
\begin{itemize}
\item \texttt{gdb} allows for Python instrumentation
\vspace{1em}\pause{}
\item Parse ELF and DWARF data (\texttt{pyelftools})
\item Run the binary inside gdb
\item Pause at each (assembly) step
\item Jointly evaluate DWARF data and the abstract stack
\item Report upon error
\end{itemize}
\vspace{1em}
Works, but… \alert{Python is slow}!
A few thousand of ASM instructions/second (good enough)
\end{frame}
\begin{frame}{A real bug!}
\begin{columns}[c]
\begin{column}{0.65\textwidth}
\lstinputlisting[language=C]{src/llvm_bug.c}
\end{column} \begin{column}{0.35\textwidth}
\textbf{\texttt{CSmith}\\+ \texttt{Creduce}\\+
\texttt{eh\_frame\_check}}
\vspace{2em}$\leadsto$ \alert{\bf LLVM (3.8) bug!}
\end{column}
\end{columns}
\end{frame}
\renewcommand{\tblrowval}[5]{#1 & #2 & #3 & #4 & #5 \\}
\begin{frame}{A real bug!}
\begin{columns}[c]
\column{0.7\textwidth}
\begin{align*}
\onslide<2->{\textbf{Abstract state} \qquad &
\left[\texttt{0xFFFF1000}\right]} \\
\onslide<3->{\reg{rsp} \qquad & %
~\,\texttt{%
\only<3-4>{0xFFFF1000}%
\only<5-8>{0xFFFF0FF8}%
\only<9->{0xFFFF1000}%
}
}
\end{align*}
\column{0.3\textwidth}
{\vspace{-4mm}\bf \fontsize{2cm}{5.5cm}\selectfont %
\only<4>{\cmark}%
\only<6>{\cmark}%
\only<8>{\cmark}%
\only<10->{\xmark}%
}
\end{columns}
\vspace{1em}
\begin{table}
\ttfamily\large
\begin{tabularx}{0.95\linewidth}{
l
l
b
>{\columncolor{SkyBlue}}s
>{\columncolor{SkyBlue}}s
}
\tblrowval{\hspace{-2ex}<{\bf foo}>:}{}{}{\textbf{CFA}}{\textbf{ra}}
%\rowonly<3>{\tblhl{}} \tblrowval{4004e0}{push}{\%rbx}{rsp+8}{c-8}
\rowonly<2-4>{\tblhl{}} \tblrowval{4004e0}{push}{\%rbx}{rsp+8}{c-8}
\rowonly<5-6>{\tblhl{}} \tblrowval{}{}{}{rsp+16}{c-8}
\tblrowval{}{[\ldots]}{}{}{}
\tblrowval{}{}{}{}{}
\rowonly<7-8>{\tblhl{}} \tblrowval{40061d}{pop}{\%rbx}{rsp+16}{c-8}
\rowonly<9->{\tblhl{}} \tblrowval{40061e}{retq}{}{rsp+16}{c-8}
\end{tabularx}
\end{table}
\begin{center}
\bf\Large %
\onslide<11>{$\leadsto$ LLVM bug \#13161}
\end{center}
\end{frame}
\begin{frame}{What for, in the end?}
\begin{itemize}[<+->]
\item We can \alert{find bugs} in compilers
\item We can \alert{validate DWARF tables}!
\item \ldots{}well, only along \alert{one execution path}\ldots
\item but mostly we are close to a working \alert{algorithm} to
\alert{synthesize unwinding data from binaries}!
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
@ -575,11 +413,11 @@ $1 = 84
\begin{frame}{\cfa{} expression}
Two possibilities:
\begin{itemize}
\item Either \reg{rbp} is used as base pointer \pause{}
\item Or we must track \cfa{} wrt. \reg{rsp}
\item Either we track \cfa{} wrt. \reg{rsp}
\begin{itemize}
\item And update it after each instruction if needed
\end{itemize}
\item and update it after each instruction if needed
\end{itemize} \pause{}
\item Or \reg{rbp} is used as base pointer: easy
\end{itemize}
\end{frame}
@ -600,7 +438,7 @@ $1 = 84
\begin{itemize}
\item \alert{Upon split} (eg. \texttt{X})\alert{:} nothing special,
propagate end state of X to children nodes A and B
propagate end state of X to child nodes A and B
\item \alert{Upon join} (eg. \texttt{while\_end})\alert{:} check
consistency of both input states
\begin{itemize}
@ -611,44 +449,6 @@ $1 = 84
\end{frame}
\renewcommand{\tblrowval}[4]{#1 & #2 & #3 & #4 \\}
\begin{frame}{Trust the compiler to avoid tricky unwinding}
\begin{columns}[c]
\column{0.55\textwidth}
\lstinputlisting[language=C,firstline=3,lastline=7]
{src/use_rbp/use_rbp.c}
\pause{}
\begin{itemize}
\item At each loop cycle, \alert{\texttt{y} is larger} and
\alert{allocated on the stack}
\item Thus, \reg{rsp} is \alert{farther from \cfa} at each cycle: no
constant rule $\cfa = \reg{rsp} + k$.
\item A complex DWARF expression is possible, but
\alert{the compiler won't}.
\end{itemize}
\pause{}
\column{0.45\textwidth}
\lstbash{\$ gcc -O0 -g -c src.c -fomit-frame-pointer}
\vspace{1em}
\begin{center}
\tt
\begin{tabular}{l l l l}
\tblrowval{LOC} {CFA} {rbp} {ra}
\tblrowval{000} {rsp+8} {u} {c-8}
\tblrowval{001} {rsp+16} {c-16} {c-8}
\tblrowval{004} {\alert{rbp}+16} {c-16} {c-8}
\tblrowval{010} {\alert{rbp}+16} {c-16} {c-8}
\tblrowval{0ce} {rsp+8} {c-16} {c-8}
\end{tabular}
\end{center}
\end{columns}
\end{frame}
\begin{frame}{}
\vfill
\centering
@ -699,10 +499,6 @@ $1 = 84
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Compilation example: original C, DWARF}
\lstinputlisting[language=C]{src/fib7/fib7.cfde}
\end{frame}
\begin{frame}[shrink]{Compilation example: generated C}
\lstinputlisting[language=C]{src/fib7/fib7.eh_elf_basic.c}
\end{frame}
@ -710,28 +506,7 @@ $1 = 84
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Compilation choices}
\textbf{In order to keep the compiler \alert{simple} and \alert{easily
testable}, the whole DWARF5 instruction set is not supported.}
\begin{itemize}
\item Focus on \alert{x86\_64}
\item Focus on unwinding return address \\
\vspace{0.3ex}
$\leadsto$ \textit{Allows building a backtrace}
\begin{itemize}
\item \alert{suitable for perf, not for gdb}
\item Only supports \alert{unwinding registers}: \reg{rip}, \reg{rsp},
\reg{rbp}, \reg{rbx}
\item Supports the \alert{wide majority} ($> 99.9\%$) of instructions
used
\item Among \alert{4000} randomly sampled filed, only \alert{24}
containing unsupported instructions
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{Interface: libunwind}
\begin{frame}{Mostly plug-and-play: libunwind interface}
\begin{itemize}
\item \alert{libunwind}: \textit{de facto} standard library for
unwinding
@ -752,79 +527,11 @@ $1 = 84
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Size optimisation: outlining}
\begin{itemize}
\item Most of the rows boil down to \alert{a few common rows}.
\begin{itemize}
\item[$\leadsto$] \textbf{\emph{outline} them!}
\end{itemize}
\pause{}
\item On libc, $20\,827$ rows $\rightarrow$ $302$ outlined ($1.5\,\%$)
\item Turn the big switch into a binary search \alert{if/else tree}
\end{itemize}
\pause{}
\bigskip{}
\begin{center}
$\leadsto$ only \textbf{2.5 times bigger than DWARF}
\end{center}
\end{frame}
\begin{frame}{Example with outlining}
\lstinputlisting[language=C]{src/fib7/fib7.eh_elf_outline.c}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Benchmarking}
\begin{frame}{Benchmarking requirements}
\begin{enumerate}
\item Thousands of samples (single unwind: $10\,\mu{}s$)
\item Interesting enough program to unwind: nested functions, complex
FDEs
\item Mitigate caching: don't always unwind from the \emph{same} point
\item Yet be fair: don't always unwind from totally different places
\item Distribute evenly: if possible, also from within libraries
\end{enumerate}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{perf instrumentation}
\textbf{\alert{perf} is a state-of-the-art polling profiler for Linux.}
\begin{itemize}
\item{} used to get readings of the time spent in each function
\item{} works by regularly stopping the program, unwinding its stack,
then aggregating the gathered data
\end{itemize}
\pause{}\bigskip{}
\textbf{Instrumenting perf matches all the requirements!}
\begin{itemize}
\item{} \alert{Plug \ehelfs{} into perf}: use \ehelfs{} instead of
DWARF to unwind the stack
\item{} Implement \alert{unwinding performance counters} inside perf
\bigskip{}
\item{} Use perf on \alert{hackbench}, a kernel stress-test program
\begin{itemize}
\item Small program
\item Lots of calls
\item Relies on libc, libpthread
\end{itemize}
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Results}
\begin{frame}{Time performance}
\todo{UPDATE}
\small
\centering
\begin{tabular}{l r r r r r}
@ -857,6 +564,7 @@ $1 = 84
\end{frame}
\begin{frame}{Space performance}
\todo{UPDATE}
\begin{center}
\begin{tabular}{r r r r r r}
\toprule
@ -884,40 +592,23 @@ $1 = 84
\section*{Conclusion}
\setcounter{section}{0}
\begin{frame}{}
\LinusMailOne{}
\begin{frame}{A fragment of our article}
The original article \textbf{Reliable and Fast DWARF-based Stack Unwinding}
contains
\only<1>{\LinusSource{}}
\only<2->{\vspace{1em}}
\pause{}
\vspace{1em}
\LinusMailTwo{}
\begin{itemize}
\item{} DWARF unwinding tables validation;
\item{} DWARF unwinding tables synthesis;
\item{} DWARF-based unwinding speedup.
\end{itemize}
\LinusSource{}
\pause{}
\vspace{1em}
\begin{center}
\bf
Give us a few months: we will make Linus reconsider \texttt{;)}
Come and chat if interested! \texttt{:)}
\end{center}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}
\vspace{5mm}
\includegraphics[width=\linewidth]{img/keep_breathing}
\vspace{-1cm}
\begin{center}
\large
\begin{align*}
\textbf{Slides: } &\text{\url{https://tobast.fr/files/oracle18.pdf}} \\
\end{align*}
\end{center}
\end{frame}
\end{document}