m2-internship/report
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity

Rephrase and correct everything up to end of §1

Browse source
This commit is contained in:
Théophile Bastian 2018-08-18 21:12:05 +02:00
parent 4016b4f46c
commit f0809dbf1c
2 changed files with 203 additions and 195 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

109
report/fiche_synthese.tex
View file

@ -8,45 +8,45 @@
\subsection*{The general context}
The standard debugging data format for ELF binary files, DWARF, contains a lot
of information, which is generated mostly when passing \eg{} the switch
\lstbash{-g} to \prog{gcc}. This information, essentially provided for
debuggers, contains all that is needed to connect the generated assembly with
the original code, information that can be used by sanitizers (\eg{} the type
of each variable in the source language), etc.
The standard debugging data format for ELF binary files, DWARF, contains tables
that permit, for a given instruction pointer (IP), to understand how the
assembly instruction relates to the source code, where variables are currently
allocated in memory or if they are stored in a register, what are their type
and how to unwind the current stack frame. This inforation is generated when
passing \eg{} the switch \lstbash{-g} to \prog{gcc} or equivalents.
Even in stripped (non-debug) binaries, a small portion of DWARF data remains.
Among this essential data that is never stripped is the stack unwinding data,
which allows to unwind stack frames, restoring machine registers to the value
they had in the previous frame, for instance within the context of a debugger
or a profiler.
Even in stripped (non-debug) binaries, a small portion of DWARF data remains:
the stack unwinding data. This information is necessary to unwind stack
frames, restoring machine registers to the value they had in the previous
frame, for instance within the context of a debugger or a profiler.
This data is structured into tables, each row corresponding to an program
counter (PC) range for which it describes valid unwinding data, and each column
describing how to unwind a particular machine register (or virtual register
used for various purposes). These rules are mostly basic, consisting in offsets
from memory addresses stored in registers (such as \reg{rbp} or \reg{rsp}), but
in some cases, they can take the form of a stack-machine expression that can
access virtually all the process's memory and perform Turing-complete
This data is structured into tables, each row corresponding to an IP range for
which it describes valid unwinding data, and each column describing how to
unwind a particular machine register (or virtual register used for various
purposes). The vast majority of the rules actually used are basic --~see
Section~\ref{ssec:instr_cov}~\textendash, consisting in offsets from memory
addresses stored in registers (such as \reg{rbp} or \reg{rsp}). Yet, the
standard defines rules that take the form of a stack-machine expression that
can access virtually all the process's memory and perform Turing-complete
computation~\cite{oakley2011exploiting}.
\subsection*{The research problem}
As debugging data can easily take an unreasonable space if stored carelessly,
the DWARF standard pays a great attention to data compactness and compression,
and succeeds particularly well at it. But this, as always, is at the expense
of efficiency: accessing stack unwinding data for a particular program point
can be quite costly.
As debugging data can easily take an unreasonable space and grow larger than
the program itself if stored carelessly, the DWARF standard pays a great
attention to data compactness and compression, and succeeds particularly well
at it. But this, as always, is at the expense of efficiency: accessing stack
unwinding data for a particular program point is not a light operation --~in
the order of magnitude of $10\,\mu{}\text{s}$ on a modern computer.
This is often not a huge problem, as stack unwinding is mostly thought of as a
This is often not a huge problem, as stack unwinding is often thought of as a
debugging procedure: when something behaves unexpectedly, the programmer might
be interested in opening their debugger and exploring the stack. Yet, stack
unwinding might, in some cases, be performance-critical: for instance, profiler
programs needs to perform a whole lot of stack unwindings. Even worse,
exception handling relies on stack unwinding in order to find a suitable
catch-block! For such applications, it might be desirable to find a different
time/space trade-off, allowing a slightly space-heavier, but far more
time-efficient unwinding procedure.
time/space trade-off, storing a bit more for a faster unwinding.
This different trade-off is the question that I explored during this
internship: what good alternative trade-off is reachable when storing the stack
@ -69,29 +69,31 @@ This internship explored the possibility to compile DWARF's stack unwinding
data directly into native assembly on the x86\_64 architecture, in order to
provide fast access to the data at assembly level. This compilation process was
fully implemented and tested on complex, real-world examples. The integration
of compiled DWARF into existing, real-world projects have been made easy by
implementing an alternative version of the \textit{de facto} standard library
for this purpose, \prog{libunwind}.
of compiled DWARF into existing projects have been made easy by implementing an
alternative version of the \textit{de facto} standard library for this purpose,
\prog{libunwind}.
Multiple approaches have been tried, in order to determine which compilation
process leads to the best time/space trade-off.
Unexpectedly, the part that proved hardest of the project was finding a
benchmarking protocol that was both relevant and reliable. Unwinding one single
frame is way too fast to provide a reliable benchmarking on a few samples
(around $10\,\mu s$ per frame). Having a lot of samples is not easy, since one
must avoid unwinding the same frame over and over again, which would only
benchmark the caching mechanism. The other problem is to distribute evenly the
unwinding measures across the various program positions, including directly
into the loaded libraries (\eg{} the \prog{libc}).
Unexpectedly, the part that proved hardest of the project was finding and
implementing a benchmarking protocol that was both relevant and reliable.
Unwinding one single frame is way too fast to provide a reliable benchmarking
on a few samples (around $10\,\mu s$ per frame). Having enough samples for this
purpose --~at least a few thousands~-- is not easy, since one must avoid
unwinding the same frame over and over again, which would only benchmark the
caching mechanism. The other problem is to distribute evenly the unwinding
measures across the various IPs, including directly into the loaded libraries
(\eg{} the \prog{libc}).
The solution eventually chosen was to modify \prog{perf}, the standard
profiling program for Linux, in order to gather statistics and benchmarks of
its unwindings. Modifying \prog{perf} was an additional challenge that turned
out to be harder than expected, since the source code is pretty opaque to
someone who doesn't know the project well. This, in particular, required to
produce an alternative version of \prog{libunwind} interfaced with the compiled
debugging data.
someone who doesn't know the project well, and the optimisations make some
parts counter-intuitive. This, in particular, required to produce an
alternative version of \prog{libunwind} interfaced with the compiled debugging
data.
% What is your solution to the question described in the last paragraph?
%
@ -108,15 +110,16 @@ debugging data.
The goal was to obtain a compiled version of unwinding data that was faster
than DWARF, reasonably heavier and reliable. The benchmarks mentioned have
yielded convincing results: on the experimental setup created (detailed later
in this report), the compiled version is around 26 times faster than the DWARF
version, while it remains only around 2.5 times bigger than the original data.
yielded convincing results: on the experimental setup created (detailed on
Section~\ref{sec:benchmarking} below), the compiled version is around 26 times
faster than the DWARF version, while it remains only around 2.5 times bigger
than the original data.
The implementation is not yet release-ready, as it does not support 100\ \% of
the DWARF5 specification~\cite{dwarf5std} --~see Section~\ref{ssec:ehelfs}
below. Yet, it supports the vast majority --~around $99.9$\ \%~-- of the cases
seen in the wild, and is decently robust compared to \prog{libunwind}, the
reference implementation. Indeed, corner cases occur often, and on a 27000
below. Yet, it supports the vast majority --~more than $99.9$\ \%~-- of the
cases seen in the wild, and is decently robust compared to \prog{libunwind},
the reference implementation. Indeed, corner cases occur often, and on a 27000
samples test, 885 failures were observed for \prog{libunwind}, against 1099 for
the compiled DWARF version (see Section~\ref{ssec:timeperf}).
@ -130,13 +133,13 @@ virtually any operating system.
\subsection*{Summary and future work}
In most cases of everyday's life, a slow stack unwinding is not a problem, or
even an annoyance. Yet, having a 26 times speed-up on stack unwinding-heavy
tasks, such as profiling, can be really useful to profile large programs,
particularly if one wants to profile many times in order to analyze the impact
of multiple changes. It can also be useful for exception-heavy programs. Thus,
it might be interesting to implement a more stable version, and try to
interface it cleanly with mainstream tools, such as \prog{perf}.
In most cases of everyday's life, a slow stack unwinding is not a problem, left
apart an annoyance. Yet, having a 26 times speed-up on stack unwinding-heavy
tasks can be really useful to \eg{} profile large programs, particularly if one
wants to profile many times in order to analyze the impact of multiple changes.
It can also be useful for exception-heavy programs. Thus, it might be
interesting to implement a more stable version, and try to interface it cleanly
with mainstream tools, such as \prog{perf}.
Another question worth exploring might be whether it is possible to shrink even
more the original DWARF unwinding data, which would be stored in a format not

289
report/report.tex
View file

@ -1,7 +1,7 @@
\title{DWARF debugging data, compilation and optimization}
\author{Théophile Bastian\\
Under supervision of Francesco Zappa-Nardelli\\
Under supervision of Francesco Zappa Nardelli\\
{\textsc{parkas}, \'Ecole Normale Supérieure de Paris}}
\date{March -- August 2018\\August 20, 2018}
@ -54,21 +54,17 @@ Under supervision of Francesco Zappa-Nardelli\\
\subsection*{Source code}\label{ssec:source_code}
The source code of all the implementations made during this internship is
available at \url{https://git.tobast.fr/m2-internship/}. See
All the source code produced during this internship is available openly. See
Section~\ref{ssec:code_avail} for details.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Stack unwinding data presentation}
The compilation process presented in this section is implemented in
\prog{dwarf-assembly}.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Stack frames and x86\_64 calling conventions}
On most platforms, programs make use of a \emph{call stack} to store
On every common platform, programs make use of a \emph{call stack} to store
information about the nested function calls at the current execution point, and
keep track of their nesting. This call stack is conventionally a contiguous
memory space mapped close to the top of the addressing space. Each function
@ -80,15 +76,15 @@ restored before returning, the function's return address and local variables.
On the x86\_64 platform, with which this report is mostly concerned, the
calling convention that is followed is defined in the System V
ABI~\cite{systemVabi} for the Unix-like operating systems --~among which Linux.
Under this calling convention, the first six arguments of a function are passed
in the registers \reg{rdi}, \reg{rsi}, \reg{rdx}, \reg{rcx}, \reg{r8},
\reg{r9}, while additional arguments are pushed onto the stack. It also defines
which registers may be overwritten by the callee, and which parameters must be
restored before returning. This restoration, most of the time, is done by
pushing the register value onto the stack in the function prelude, and
restoring it just before returning. Those preserved registers are \reg{rbx},
\reg{rsp}, \reg{rbp}, \reg{r12}, \reg{r13}, \reg{r14}, \reg{r15}.
ABI~\cite{systemVabi} for the Unix-like operating systems --~among which Linux
and MacOS\@. Under this calling convention, the first six arguments of a
function are passed in the registers \reg{rdi}, \reg{rsi}, \reg{rdx},
\reg{rcx}, \reg{r8}, \reg{r9}, while additional arguments are pushed onto the
stack. It also defines which registers may be overwritten by the callee, and
which registers must be restored before returning. This restoration, for most
compilers, is done by pushing the register value onto the stack in the function
prelude, and restoring it just before returning. Those preserved registers are
\reg{rbx}, \reg{rsp}, \reg{rbp}, \reg{r12}, \reg{r13}, \reg{r14}, \reg{r15}.
\begin{wrapfigure}{r}{0.4\textwidth}
\centering
@ -104,29 +100,32 @@ use \reg{rbp} (``base pointer'') to save this value of \reg{rip}, by writing
the old value of \reg{rbp} just below the return address on the stack, then
copying \reg{rsp} to \reg{rbp}. This makes it easy to find the return address
from anywhere within the function, and also allows for easy addressing of local
variables. Yet, using \reg{rbp} to save \reg{rip} is not always done, since it
somehow ``wastes'' a register. This decision is, on x86\_64 System V, up to the
compiler.
variables. To some extents, it also allows for hot debugging, such as saving a
useful core dump upon segfault. Yet, using \reg{rbp} to save \reg{rip} is not
always done, since it somehow ``wastes'' a register. This decision is, on
x86\_64 System V, up to the compiler.
Often, a function will start by subtracting some value to \reg{rsp}, allocating
some space in the stack frame for its local variables. Then, it will push on
some space in the stack frame for its local variables. Then, it will push on
the stack the values of the callee-saved registers that are overwritten later,
effectively saving them. Before returning, it will pop the values of the saved
registers back to their original registers and restore \reg{rsp} to its former
value.
\subsection{Stack unwinding}
\subsection{Stack unwinding}\label{ssec:stack_unwinding}
For various reasons, it might be interesting, at some point of the execution of
a program, to glance at its program stack and be able to extract informations
from it. For instance, when running a debugger such as \prog{gdb}, a frequent
usage is to obtain a \emph{backtrace}, that is, the list of all nested function
calls at this point. This actually reads the stack to find the different stack
frames, and decode them to identify the function names, parameter values, etc.
calls at the current IP\@. This actually reads the stack to find the different
stack frames, and decode them to identify the function names, parameter values,
etc.
This operation is far from trivial. Often, a stack frame will only make sense
with correct machine registers values, which can be restored from the previous
stack frame, imposing to \emph{walk} the stack, reading the entries one after
when the correct values are stored in the machine registers. These values,
however, are to be restored from the previous stack frame, where they are
stored. This imposes to \emph{walk} the stack, reading the entries one after
the other, instead of peeking at some frame directly. Moreover, the size of one
stack frame is often not that easy to determine when looking at some
instruction other than \texttt{return}, making it hard to extract single frames
@ -138,28 +137,29 @@ frame, and thus be able to decode the next frame recursively, is called
Let us consider a stack with x86\_64 calling conventions, such as shown in
Figure~\ref{fig:call_stack}. Assuming the compiler decided here \emph{not} to
use \reg{rbp}, and assuming the function \eg{} allocates a buffer of 8
use \reg{rbp}, and assuming the function allocates \eg{} a buffer of 8
integers, the area allocated for local variables should be at least $32$ bytes
long (for 4-bytes integers), and \reg{rsp} will be pointing below this area.
Left apart analyzing the assembly code produced, there is no way to find where
the return address is stored, relatively to \reg{rsp}, at some arbitrary point
of the function. Even when \reg{rbp} is used, there is no easy way to guess
where each callee-saved register is stored in the stack frame, and worse, which
callee-saved registers were saved, since it is optional to save a register
that the function never touches.
where each callee-saved register is stored in the stack frame, since the
compiler is free to do as it wishes. Even worse, it is not trivial to know
callee-saved registers were at all, since if the function does not alter a
register, it does not have to save it.
With this example, it seems pretty clear that it is often necessary to have
additional data to perform stack unwinding. This data is often stored among the
debugging informations of a program, and one common format of debugging data is
DWARF\@.
With this example, it seems pretty clear tha some additional data is necessary
to perform stack unwinding reliably, without only performing a guesswork. This
data is stored along with the debugging informations of a program, and one
common format of debugging data is DWARF\@.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Unwinding usage and frequency}
Stack unwinding is a more common operation that one might think at first. The
most commonly thought use-case is simply to get a stack trace of a program, and
provide a debugger with the information it needs: for instance, when inspecting
a stack trace in \prog{gdb}, it is quite common to jump to a previous frame:
use case mostly thought of is simply to get a stack trace of a program, and
provide a debugger with the information it needs. For instance, when inspecting
a stack trace in \prog{gdb}, a common operation is to jump to a previous frame:
\lstinputlisting{src/segfault/gdb_session}
@ -176,39 +176,43 @@ Linux --~see Section~\ref{ssec:perf} --, is used to measure and analyze in
which functions a program spends its time, identify bottlenecks and find out
which parts are critical to optimize. To do so, modern profilers pause the
traced program at regular, short intervals, inspect their stack, and determine
which function is currently being run. They also often perform a stack
unwinding to determine the call path to this function, to determine which
function indirectly takes time: \eg, a function \lstc{fct_a} can call both
\lstc{fct_b} and \lstc{fct_c}, which take a lot of time; spend practically no
time directly in \lstc{fct_a}, but spend a lot of time in calls to the other
two functions that were made from \lstc{fct_a}.
which function is currently being run. They also perform a stack unwinding to
figure out the call path to this function, in order to determine which function
indirectly takes time: for instance, a function \lstc{fct_a} can call both
\lstc{fct_b} and \lstc{fct_c}, which both take a lot of time; spend practically
no time directly in \lstc{fct_a}, but spend a lot of time in calls to the other
two functions that were made from \lstc{fct_a}. Knowing that after all,
\lstc{fct_a} is the culprit can be useful to a programmer.
Exception handling also requires a stack unwinding mechanism in most languages.
Indeed, an exception is completely different from a \lstinline{return}: while the
latter returns to the previous function, the former can be caught by virtually
any function in the call path, at any point of the function. It is thus
necessary to be able to unwind frames, one by one, until a suitable
\lstc{catch} block is found. The C++ language, for one, includes a
Indeed, an exception is completely different from a \lstinline{return}: while
the latter returns to the previous function, at a well-defined IP, the former
can be caught by virtually any function in the call path, at any point of the
function. It is thus necessary to be able to unwind frames, one by one, until a
suitable \lstc{catch} block is found. The C++ language, for one, includes a
stack-unwinding library similar to \prog{libunwind} in its runtime.
Technically, exception handling could be implemented without any stack
unwinding, by using \lstc{setjmp}/\lstc{longjmp} mechanics~\cite{niditoexn}.
However, this is not possible to implement it straight away in C++ (and some
other languages), because the stack needs to be properly unwound in order to
trigger the destructors of stack-allocated objects. Furthermore, this is often
undesirable: \lstc{setjmp} has a quite big overhead, which is introduced
whenever a \lstc{try} block is encountered. Instead, it is often preferred to
have strictly no overhead when no exception happens, at the cost of a greater
overhead when an exception is actually fired --~after all, they are supposed to
be \emph{exceptional}. For more details on C++ exception handling,
see~\cite{koening1990exception} (especially Section~16.5). Possible
implementation mechanisms are also presented in~\cite{dinechin2000exn}.
unwinding, by using \lstc{setjmp} and \lstc{longjmp}
mechanics~\cite{niditoexn}. However, it is not possible to implement this
straight away in C++ (among others), because the stack needs to be
properly unwound in order to trigger the destructors of stack-allocated
objects. Furthermore, this is often undesirable: \lstc{setjmp} introduces an
overhead, which is hit whenever a \lstc{try} block is encountered. Instead, it
is often preferred to have strictly no overhead when no exception happens, at
the cost of a greater overhead when an exception is actually fired --~after
all, they are supposed to be \emph{exceptional}. For more details on C++
exception handling, see~\cite{koening1990exception} (especially Section~16.5).
Possible implementation mechanisms are also presented
in~\cite{dinechin2000exn}.
In both of these two previous cases, performance \emph{can} be a problem. In
the latter, a slow unwinding directly impacts the overall program performance,
particularly if a lot of exceptions are thrown and caught far away in their
call path. In the former, profiling \emph{is} performance-heavy and often quite
slow when analyzing large programs anyway.
call path. As for the former, profiling \emph{is} performance-heavy and slow:
for a session analyzing the \prog{tor-browser} for two and a half minutes,
\prog{perf} spends $100\,\mu \text{s}$ analyzing each of the $325679$ samples,
that is, $300\,\text{ms}$ per second of program run with default settings.
One of the causes that inspired this internship were also Stephen Kell's
\prog{libcrunch}~\cite{kell2016libcrunch}, which makes a heavy use of stack
@ -229,43 +233,43 @@ The DWARF data commonly includes type information about the variables in the
original programming language, correspondence of assembly instructions with a
line in the original source file, \ldots
The format also specifies a way to represent unwinding data, as described in
the previous paragraph, in an ELF section originally called
\lstc{.debug_frame}, most often found as \ehframe.
Section~\ref{ssec:stack_unwinding} above, in an ELF section originally called
\lstc{.debug_frame}, but most often found as \ehframe.
For any binary, debugging information can easily get quite large if no
attention is payed to keeping it as compact as possible. In this matter, DWARF
does an excellent job, and everything is stored in a very compact way. This,
however, as we will see, makes it both difficult to parse correctly and quite
slow to interpret.
For any binary, debugging information can easily take up space and grow bigger
than the program itself if no attention is paid at keeping it as compact as
possible when designing the file format. On this matter, DWARF does an
excellent job, and everything is stored in a very compact way. This, however,
as we will see, makes it both difficult to parse correctly and relatively slow
to interpret.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{DWARF unwinding data}
The unwinding data, which we will call from now on the \ehframe, contains, for
each possible instruction pointer (that is, an instruction address within the
program), a set of ``registers'' that can be unwound, and a rule describing how
to do so.
each possible IP, a set of ``registers'' that can be unwound, and a rule
describing how to do so.
The DWARF language is completely agnostic of the platform and ABI, and in
particular, is completely agnostic of a particular platform's registers. Thus,
when talking about DWARF, a register is merely a numerical identifier that is
often, but not necessarily, mapped to a real machine register by the ABI\@.
as far as DWARF is concerned, a register is merely a numerical identifier that
is often, but not necessarily, mapped to a real machine register by the ABI\@.
In practice, this data takes the form of a collection of tables, one table per
Frame Description Entry (FDE). A FDE, in turn, is a DWARF entry describing such
a table, that has a range of IPs on which it has authority. Most often, but not
necessarily, it corresponds to a single function in the original source code.
Each column of the table is a register (\eg{} \reg{rsp}), with two additional
special registers, CFA (Canonical Frame Address) and RA (Return Address),
containing respectively the base pointer of the current stack frame and the
return address of the current function. For instance, on a x86\_64
Each column of the table is a register (\eg{} \reg{rsp}), along with two
additional special registers, CFA (Canonical Frame Address) and RA (Return
Address), containing respectively the base pointer of the current stack frame
and the return address of the current function. For instance, on a x86\_64
architecture, RA would contain the unwound value of \reg{rip}, the instruction
pointer. Each row has a certain validity interval, on which it describes
accurate unwinding data. This range starts at the instruction pointer it is
associated with, and ends at the start IP of the next table row (or the end IP
of the current FDE if it was the last row). In particular, there can be no ``IP
hole'' within a FDE --~unlike FDEs themselves, which can leave holes between
them.
associated with, and ends at the start IP of the next table row --~or the end
IP of the current FDE if it was the last row. In particular, there can be no
``IP hole'' within a FDE --~unlike FDEs themselves, which can leave holes
between them.
\begin{figure}[h]
\begin{minipage}{0.45\textwidth}
@ -312,7 +316,7 @@ them.
\caption{Stack frame schema}\label{table:ex1_stack_schema}
\end{table}
For instance, the C source code in Listing~\ref{lst:ex1_c} above, when compiled
For instance, the C source code in Listing~\ref{lst:ex1_c}, when compiled
with \lstbash{gcc -O1 -fomit-frame-pointer -fno-stack-protector}, yields the
assembly code in Listing~\ref{lst:ex1_asm}. The memory layout of the stack
frame is presented in Table~\ref{table:ex1_stack_schema}, to help understanding
@ -380,9 +384,9 @@ Figure~\ref{fig:fde_line_density} was generated on a random sample of around
The most commonly used library to perform stack unwinding, in the Linux
ecosystem, is \prog{libunwind}~\cite{libunwind}. While it is very robust and
quite efficient, most of its optimization comes from fine-tuned code and good
caching mechanisms. While parsing DWARF, \prog{libunwind} is forced to parse
the relevant FDE from its start, until it finds the row it was seeking.
decently efficient, most of its optimization comes from fine-tuned code and
good caching mechanisms. When parsing DWARF, \prog{libunwind} is forced to
parse the relevant FDE from its start, until it finds the row it was seeking.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
@ -392,9 +396,9 @@ the relevant FDE from its start, until it finds the row it was seeking.
We will now define semantics covering most of the operations used for FDEs
described in the DWARF standard~\cite{dwarf5std}, such as seen in
Listing~\ref{lst:ex1_dwraw}, with the exception of DWARF expressions. These are
not exhaustively treated because they are quite rich and would take a lot of
time and space to formalize, and in the meantime are only seldom used (see the
DWARF statistics regarding this).
not exhaustively treated because they form a rich language and would take a lot
of time and space to formalize, and in the meantime are only seldom used (see
the DWARF statistics regarding this).
These semantics are defined with respect to the well-formalized C language, and
are passing through an intermediary language. The DWARF language can read the
@ -650,7 +654,7 @@ earlier. The translation from $\intermedlang$ to C is defined as follows:
if(ip >= $loc$) {
for(int reg=0; reg < NB_REGS; ++reg)
new_ctx[reg] = $\semR{row[reg]}$;
goto end_ifs; // Avoid if/else if problems
goto end_ifs; // Avoid using `else if` (easier for generation)
}
\end{lstlisting}
\end{itemize}
@ -688,9 +692,8 @@ licenses.
The rough idea of the compilation is to produce, out of the \ehframe{} section
of a binary, C code that resembles the code shown in the DWARF semantics from
Section~\ref{sec:semantics} above. This C code is then compiled by GCC in
\lstbash{-O2} mode, since it already provides a good level of optimization and
compiling in \lstbash{-O3} takes way too much time. This saves us the trouble
of optimizing the generated C code whenever GCC does that by itself.
\lstbash{-O2} mode. This saves us the trouble of optimizing the generated C
code whenever GCC does that by itself.
The generated code consists in a single monolithic function, \lstc{_eh_elf},
taking as arguments an instruction pointer and a memory context (\ie{} the
@ -698,15 +701,15 @@ value of the various machine registers) as defined in
Listing~\ref{lst:unw_ctx}. The function will then return a fresh memory
context, containing the values the registers hold after unwinding this frame.
The body of the function itself is mostly a huge switch, taking advantage of
the non-standard --~yet widely implemented in C compilers~-- syntax for range
switches, in which each \lstinline{case} can refer to a range. All the FDEs are
merged together into this switch, each row of a FDE being a switch case.
Separating the various FDEs in the C code --~other than with comments~-- is,
unlike what is done in DWARF, pointless, since accessing a ``row'' has a linear
cost, and the C code is not meant to be read, except maybe for debugging
purposes. The switch cases bodies then fill a context with unwound values, then
return it.
The body of the function itself consists in a single monolithic switch, taking
advantage of the non-standard --~yet widely implemented in C compilers~--
syntax for range switches, in which each \lstinline{case} can refer to a range.
All the FDEs are merged together into this switch, each row of a FDE being a
switch case. Separating the various FDEs in the C code --~other than with
comments~-- is, unlike what is done in DWARF, pointless, since accessing a
``row'' has a linear cost, and the C code is not meant to be read, except maybe
for debugging purposes. The switch cases bodies then fill a context with
unwound values, then return it.
A setting of the compiler also optionally enables another parameter to the
\lstc{_eh_elf} function, \lstc{deref}, which is a function pointer. This
@ -720,13 +723,13 @@ Unlike in the \ehframe, and unlike what should be done in a release,
real-world-proof version of the \ehelfs, the choice was made to keep this
implementation simple, and only handle the few registers that were needed to
simply unwind the stack. Thus, the only registers handled in \ehelfs{} are
\reg{rip}, \reg{rbp}, \reg{rsp} and \reg{rbx}, the latter being used quite
often in \prog{libc} to hold the CFA address. This is enough to unwind the
stack reliably, and thus enough for profiling, but is not sufficient to analyze
every stack frame as \prog{gdb} would do after a \lstbash{frame n} command.
Yet, if one was to enhance the code to handle every register, it would not be
much harder and would probably be only a few hours of code refactoring and
rewriting.
\reg{rip}, \reg{rbp}, \reg{rsp} and \reg{rbx}, the latter being used a few
times in \prog{libc} to hold the CFA address in common functions. This is
enough to unwind the stack reliably, and thus enough for profiling, but is not
sufficient to analyze every stack frame as \prog{gdb} would do after a
\lstbash{frame n} command. Yet, if one was to enhance the code to handle every
register, it would not be much harder and would probably be only a few hours of
code refactoring and rewriting.
\lstinputlisting[language=C, caption={Unwinding context}, label={lst:unw_ctx}]
{src/dwarf_assembly_context/unwind_context.c}
@ -743,16 +746,15 @@ by including an error flag by lack of $\bot$ value.
This generated data is stored in separate shared object files, which we call
\ehelfs. It would have been possible to alter the original ELF file to embed
this data as a new section, but getting it to be executed just as any
portion of the \lstc{.text} section would probably have been painful, and
keeping it separated during the experimental phase is quite convenient. It is
possible to have multiple versions of \ehelfs{} files in parallel, with various
options turned on or off, and it doesn't require to alter the base system by
editing \eg{} \texttt{/usr/lib/libc-*.so}. Instead, when the \ehelf{} data is
required, those files can simply be \lstc{dlopen}'d. It is also possible to
imagine, in a future environment production, packaging \ehelfs{} files
separately, so that people interested in heavy computation can have the choice
to install them.
this data as a new section, but getting it to be executed just as any portion
of the \lstc{.text} section would probably have been painful, and keeping it
separated during the experimental phase is convenient. It is possible to have
multiple versions of \ehelfs{} files in parallel, with various options turned
on or off, and it doesn't require to alter the base system by editing \eg{}
\texttt{/usr/lib/libc-*.so}. Instead, when the \ehelf{} data is required, those
files can simply be \lstc{dlopen}'d. It is also possible to imagine, in a
future environment production, packaging \ehelfs{} files separately, so that
people interested in heavy computation can have the choice to install them.
This, in particular, means that each ELF file has its unwinding data in a
separate \ehelf{} file --~just like with DWARF, where each ELF retains its own
@ -781,10 +783,8 @@ possible to produce a compiled version very close to the one described in
Section~\ref{sec:semantics}. Although the unwinding speed cannot yet be
actually benchmarked, it is already possible to write in a few hundred lines of
C code a simple stack walker printing the functions traversed. It already works
without any problem on the easily tested cases, since corner cases are mostly
found in standard and highly optimized libraries, and it is not that easy to get
the program to stop and print a stack trace from within a system library
without using a debugger.
well on the standard cases that are easily tested, and can be used to unwind
the stack of simple programs.
The major drawback of this approach, without any particular care taken, is the
space waste. The space taken by those tentative \ehelfs{} is analyzed in
@ -835,14 +835,16 @@ made in order to shrink the \ehelfs.
\medskip
The major optimization that most reduced the output size was to use an if/else
tree implementing a binary search on the program counter relevant intervals,
instead of a huge switch. In the process, we also \emph{outline} a lot of code,
that is, find out identical ``switch cases'' bodies --~which are not switch
cases anymore, but if bodies~--, move them outside of the if/else tree,
identify them by a label, and jump to them using a \lstc{goto}, which
de-duplicates a lot of code and contributes greatly to the shrinking. In the
process, we noticed that the vast majority of FDE rows are actually taken among
very few ``common'' FDE rows.
tree implementing a binary search on the instruction pointer relevant
intervals, instead of a single monolithic switch. In the process, we also
\emph{outline} code whenever possible, that is, find out identical ``switch
cases'' bodies --~which are not switch cases anymore, but if bodies~--, move
them outside of the if/else tree, identify them by a label, and jump to them
using a \lstc{goto}, which de-duplicates a lot of code and contributes greatly
to the shrinking. In the process, we noticed that the vast majority of FDE rows
are actually taken among very few ``common'' FDE rows. For instance, in the
\prog{libc}, out of a total of $20827$ rows, only $302$ ($1.5\,\%$) remain
after the outlining.
This makes this optimization really efficient, as seen later in
Section~\ref{ssec:results_size}, but also makes it an interesting question
@ -861,10 +863,10 @@ DWARF data could be efficiently compressed in this way.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Benchmarking}
\section{Benchmarking}\label{sec:benchmarking}
Benchmarking turned out to be, quite surprisingly, the hardest part of the
project. It ended up requiring a lot of investigation to find a working
project. It ended up requiring a good deal of investigation to find a working
protocol, and afterwards, a good deal of code reading and coding to get the
solution working.
@ -884,13 +886,15 @@ are made from different locations is somehow cheating, since it makes useless
distribution. All in all, the benchmarking method must have a ``natural''
distribution of unwindings.
Another requirement is to also distribute quite evenly the unwinding points
across the program: we would like to benchmark stack unwindings crossing some
standard library functions, starting from inside them, etc.
Another requirement is to also distribute evenly enough the unwinding points
across the program to mimic real-world unwinding: we would like to benchmark
stack unwindings crossing some standard library functions, starting from inside
them, etc.
Finally, the unwound program must be interesting enough to enter and exit a lot
of functions, nest function calls, have FDEs that are not as simple as in
Listing~\ref{lst:ex1_dw}, etc.
Finally, the unwound program must be interesting enough to enter and exit
functions often, building a good stack of nested function calls (at least 5
frequently), have FDEs that are not as simple as in Listing~\ref{lst:ex1_dw},
etc.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
@ -915,7 +919,8 @@ program, but instead inject code in it.
\subsection{Benchmarking with \prog{perf}}\label{ssec:bench_perf}
In the context of this internship, the main advantage of \prog{perf} is that it
does a lot of stack unwinding. It also meets all the requirements from
unwinds the stack on a regular, controllable basis, easily unwinding thousands
of time in a few seconds. It also meets all the requirements from
Section~\ref{ssec:bench_req} above: since it stops at regular intervals and
unwinds, the unwindings are evenly distributed \wrt{} the frequency of
execution of the code, which is a natural enough setup for the benchmarks to be
@ -966,10 +971,10 @@ the repositories \prog{perf-eh\_elf} and \prog{libunwind-eh\_elf}.
The first approach tried to benchmark was trying to create some specific C code
that would meet the requirements from Section~\ref{ssec:bench_req}, while
calling itself a benchmarking procedure from time to time. This was abandoned
quite quickly, because generating C code interesting enough to be unwound
turned out hard, and the generated FDEs invariably ended out uninteresting. It
would also never have met the requirement of unwinding from fairly distributed
calling itself a benchmarking procedure from time to time. This was quickly
abandoned, because generating C code interesting enough to be unwound turned
out hard, and the generated FDEs invariably ended out uninteresting. It would
also never have met the requirement of unwinding from fairly distributed
locations anyway.
Another attempt was made using CSmith~\cite{csmith}, a random C code generator