Julia Evans

How is a binary executable organized? Let's explore it!

I used to think that executables were totally impenetrable. I’d compile a C program, and then that was it! I had a Magical Binary Executable that I could no longer read.

It is not so! Executable file formats are regular file formats that you can understand. I’ll explain some simple tools to start! We’ll be working on Linux, with ELF binaries. (binaries are kind of the definition of platform-specific, so this is all platform-specific.) We’ll be using C, but you could just as easily look at output from any compiled language.

Let’s write a simple C program, hello.c:

#include <stdio.h>

int main() {

Then we compile it (gcc -o hello hello.c), and we have a binary called hello. This originally seems impenetrable (how do we even binary?!), but let’s see how we can investigate it! We’re going to learn what symbols, sections, and segments are. At a high level:

  • symbols are like function names, and are used to answer “If I call printf and it’s defined somewhere else, how do I find it?”
  • symbols are organized into sections – code lives in one section (.text), and data in another (.data, .rodata)
  • sections are organized into segments

Throughout we’ll use a tool called readelf to look at these.

So, let’s dive into our binary!

Step 1: open it in a text editor!

This is most naive possible way to view a binary. If run cat hello, I get something like this:

@@@@@@��88@@@@�� ((`(`�
1```H��k����H���5 H�[]�fff.�H�=p
UH��t�H��]�H`��]�UH����@�����]�����������H�l$�L�d$�H�- L�%

There’s text here, though! This was not a total failure. In particular it says “Penguin!” and “ELF”. ELF is the name of the binary format. So that’s something! Then there are a bunch of unprintable symbols, which isn’t a huge surprise because this is a binary.

Step 2: use readelf to see the symbol table

Throughout we’re going to use a tool called readelf to explore our binary. Let’s start by running readelf --symbols on it. (another popular tool to do this is nm)

$ readelf --symbols hello 
   Num:    Value          Size Type    Bind   Vis      Ndx Name
    48: 0000000000000000     0 FUNC    GLOBAL DEFAULT  UND puts@@GLIBC_2.2.5
    59: 0000000000400410     0 FUNC    GLOBAL DEFAULT   13 _start
    61: 00000000004004f4    16 FUNC    GLOBAL DEFAULT   13 main

(full output here)

Here we see three symbols: main is the address of my main() function. puts looks like a reference to the printf function I called in it (which I guess the compiler changed to puts as an optimization?). _start is pretty important.

When the program starts running, you might think it starts at main. It doesn’t! It actually goes to _start. This does a bunch of Very Important Things that I don’t understand very well, including calling main. So I won’t explain them.

So, what’s a symbol?


When you write a program, you might write a function called hello. When you compile the program, the binary for that function is labelled with a symbol called hello. If I call a function (like printf) from a library, we need a way to look up the code for that function! The process of looking up functions from libraries is called linking. It can happen either just after we compile the program (“static linking”) or when we run the program (“dynamic linking”).

So symbols are what allow linking to work! Let’s find the symbol for printf! It’ll be in libc, where all the C standard library functions are.

If I run nm on my copy of libc, it tells me “no symbols”. But the internet tells me I can use objdump -tT instead! This works! objdump -tT /lib/x86_64-linux-gnu/libc-2.15.so gives me this output.

If you look at it, you’ll see sprintf, strlen, fork, exec, and everything you might expect libc to have. From here we can start to imagine how dynamic linking works – we see that hello calls puts, and then we can look up the location of puts in libc’s symbol table.

Step 3: use objdump to see the binary, and learn about sections!

Opening our binary in a text editor was a bad way to open it. objdump is a better way. Here’s an excerpt:

$ objdump -s hello
Contents of section .text:
 400410 31ed4989 d15e4889 e24883e4 f0505449  1.I..^H..H...PTI
 400420 c7c0a005 400048c7 c1100540 0048c7c7  ....@.H....@.H..
 400430 f4044000 e8c7ffff fff49090 4883ec08  ..@.........H...
Contents of section .interp:
 400238 2f6c6962 36342f6c 642d6c69 6e75782d  /lib64/ld-linux-
 400248 7838362d 36342e73 6f2e3200           x86-64.so.2.    
Contents of section .rodata:
 4005f8 01000200 50656e67 75696e21 00        ....Penguin!.   

You can see that it shows us all the bytes in the file as hex on the left, and a translation into ASCII on the right.

The are a whole bunch of sections here (see this gist for the whole thing). This shows you all the bytes in your binary! Some sections we care about:

  • .text is the program’s actual code (the assembly). _start and main are both part of the .text section.
  • .rodata is where some read-only data is stored (in this case, our string “Penguin!”)
  • .interp is the filename of the dynamic linker!

The major difference between sections and segments is that sections are used at link time (by ld) and segments are used at execution time. objdump shows us the contents of the sections, which is nice, but doesn’t give us as much metadata about the sections as I’d like. Let’s try readelf instead:

$ readelf --sections hello
Section Headers:
  [Nr] Name              Type             Address           Offset
       Size              EntSize          Flags  Link  Info  Align
  [13] .text             PROGBITS         0000000000400410  00000410
       00000000000001d8  0000000000000000  AX       0     0     16
  [15] .rodata           PROGBITS         00000000004005f8  000005f8
       000000000000000b  0000000000000000   A       0     0     4
  [24] .data             PROGBITS         0000000000601010  00001010
       0000000000000010  0000000000000000  WA       0     0     8
  [25] .bss              NOBITS           0000000000601020  00001020
       0000000000000010  0000000000000000  WA       0     0     8
  [26] .comment          PROGBITS         0000000000000000  00001020
       000000000000002a  0000000000000001  MS       0     0     1
Key to Flags:
  W (write), A (alloc), X (execute), M (merge), S (strings), l (large)
  I (info), L (link order), G (group), T (TLS), E (exclude), x (unknown)
  O (extra OS processing required) o (OS specific), p (processor specific)

(full output)

Neat! We can see .text is executable and read-only, .rodata (“read only data”) is read-only, and .data is read-write.

Step 4: Look at some assembly!

We mentioned briefly that .text contains assembly code. We can actually look at what it is really easily. If we were magicians, we would already be able to read and understand this:

Contents of section .text:
 400410 31ed4989 d15e4889 e24883e4 f0505449  1.I..^H..H...PTI
 400420 c7c0a005 400048c7 c1100540 0048c7c7  ....@.H....@.H..
 400430 f4044000 e8c7ffff fff49090 4883ec08  ..@.........H...

It starts with 31ed4989. Those are bytes that our CPU interprets as code! And runs! However we are not magicians (I don’t know what 31 ed means!) and so we will use a disassembler instead.

$ objdump -d ./hello
Disassembly of section .text:

0000000000400410 <_start>:
  400410:       31 ed                   xor    %ebp,%ebp
  400412:       49 89 d1                mov    %rdx,%r9
  400415:       5e                      pop    %rsi
  400416:       48 89 e2                mov    %rsp,%rdx
  400419:       48 83 e4 f0             and    $0xfffffffffffffff0,%rsp

full output here

So we see that 31 ed is xoring two things. Neat! That’s all the assembly we’ll do for now.

Step 5: Segments!

Finally, a program is organized into segments or program headers. Let’s look at the segments for our program using readelf --segments hello.

Program Headers:
  [... removed ...]
  INTERP         0x0000000000000238 0x0000000000400238 0x0000000000400238
                 0x000000000000001c 0x000000000000001c  R      1
      [Requesting program interpreter: /lib64/ld-linux-x86-64.so.2]
  LOAD           0x0000000000000000 0x0000000000400000 0x0000000000400000
                 0x00000000000006d4 0x00000000000006d4  R E    200000
  LOAD           0x0000000000000e28 0x0000000000600e28 0x0000000000600e28
                 0x00000000000001f8 0x0000000000000208  RW     200000
  [... removed ...]

 Section to Segment mapping:
  Segment Sections...
   01     .interp 
   02 .interp .note.ABI-tag .note.gnu.build-id .gnu.hash .dynsym
       .dynstr .gnu.version .gnu.version_r .rela.dyn .rela.plt .init .plt
       .text .fini .rodata .eh_frame_hdr .eh_frame
   03     .ctors .dtors .jcr .dynamic .got .got.plt .data .bss 
   04     .dynamic 
   05     .note.ABI-tag .note.gnu.build-id 
   06     .eh_frame_hdr 
   08     .ctors .dtors .jcr .dynamic .got 

Segments are used to determine how to separate different parts of the program into memory. The first LOAD segment is marked R E (read / execute) and the second is RW (read/write). .text is in the first segment (we want to read it but never write to it), and .data, .bss are in the second (we need to write to them, but not execute them).

Not magic!

Executables aren’t magic. ELF is a file format like any other! You can use readelf, nm, and objdump to inspect your Linux binaries. Try it out! Have fun.

Other resources:

Thanks very much to the amazing Allison Kaptur and Dan Luu for reading a draft of this.

What happens if you write a TCP stack in Python? You can be a kernel hacker!