ELF is an extremely flexible format for representing binary code in a system. By following the ELF standard you can represent a kernel binary just as easily as a normal executable or a system library. The same tools can be used to inspect and operate on all ELF files and developers who understand the ELF file format can translate their skills to most modern UNIX systems.

ELF extends on COFF and gives the header sufficient flexibility to define an arbitrary number of sections, each with its own properties. This facilitates easier dynamic linking and debugging.

ELF Overview
Figure 3.1 ELF Overview

3.1 ELF File Header

Overall, the file has a file header which describes the file in general and then has pointers to each of the individual sections that make up the file. Example 3.1.1, The ELF Header shows the description as given in the API documentation for ELF32 (the 32-bit form of ELF). This is the layout of the C structure which defines a ELF header.

typedef struct {
        unsigned char e_ident[EI_NIDENT];
        Elf32_Half    e_type;
        Elf32_Half    e_machine;
        Elf32_Word    e_version;
        Elf32_Addr    e_entry;
        Elf32_Off     e_phoff;
        Elf32_Off     e_shoff;
        Elf32_Word    e_flags;
        Elf32_Half    e_ehsize;
        Elf32_Half    e_phentsize;
        Elf32_Half    e_phnum;
        Elf32_Half    e_shentsize;
        Elf32_Half    e_shnum;
        Elf32_Half    e_shstrndx;
} Elf32_Ehdr;
Example 3.1.1 The ELF Header
$ readelf --header /bin/ls

ELF Header:
  Magic:   7f 45 4c 46 01 02 01 00 00 00 00 00 00 00 00 00 
  Class:                             ELF32
  Data:                              2's complement, big endian
  Version:                           1 (current)
  OS/ABI:                            UNIX - System V
  ABI Version:                       0
  Type:                              EXEC (Executable file)
  Machine:                           PowerPC
  Version:                           0x1
  Entry point address:               0x10002640
  Start of program headers:          52 (bytes into file)
  Start of section headers:          87460 (bytes into file)
  Flags:                             0x0
  Size of this header:               52 (bytes)
  Size of program headers:           32 (bytes)
  Number of program headers:         8
  Size of section headers:           40 (bytes)
  Number of section headers:         29
  Section header string table index: 28

Example 3.1.2 The ELF Header, as shown by readelf

Example 3.1.2, The ELF Header, as shown by readelf shows a human readable form as present by the readelf program, which is part of GNU binutils.

The e_ident array is the first thing at the start of any ELF file, and always starts with a few "magic" bytes. The first byte is 0x7F and then the next three bytes are "ELF". You can inspect an ELF binary to see this for yourself with something like the hexdump command.

ianw@mingus:~$ hexdump -C /bin/ls | more
00000000  7f 45 4c 46 01 02 01 00  00 00 00 00 00 00 00 00  |.ELF............|

... (rest of the program follows) ...
Example 3.1.3 Inspecting the ELF magic number

Note the 0x7F to start, then the ASCII encoded "ELF" string. Have a look at the standard and see what the rest of the array defines and what the values are in a binary.

Next we have some flags for the type of machine this binary is created for. The first thing we can see is that ELF defines different type sized versions, one for 32 bit and one for 64 bit versions; here we inspect the 32 bit version. The difference is mostly that on 64 bit machines addresses obviously required to be held in 64 bit variables. We can see that the binary has been created for a big endian machine that uses 2's complement to represent negative numbers. Skipping down a bit we can see the Machine tells us this is a PowerPC binary.

The apparently innocuous entry point address seems straight forward enough; this is the address in memory that the program code starts at. Beginning C programmers are told that main() is the first program called in your program. Using the entry point address we can actually verify that it isn't.

$ cat test.c
#include <stdio.h>

int main(void)
        printf("main is : %p\n", &main);
        return 0;

$ gcc -Wall -o test test.c

$ ./test
main is : 0x10000430

$ readelf --headers ./test | grep 'Entry point'
  Entry point address:               0x100002b0

$ objdump --disassemble ./test | grep 100002b0
100002b0 <_start>:
100002b0:       7c 29 0b 78     mr      r9,r1

Example 3.1.4 Investigating the entry point

In Example 3.1.4, Investigating the entry point we can see that the entry point is actually a function called _start. Our program didn't define this at all, and the leading underscore suggests that it is in a separate namespace. We examine how a program starts up in detail in Section 8.2, Starting the program.

After that the header contains pointers to where in the file other important parts of the ELF file start, like a table of contents.

3.2 Symbols and Relocations

The ELF specification provides for symbol tables which are simply mappings of strings (symbols) to locations in the file. Symbols are required for linking; for example assigning a value to a variable foo declared as extern int foo; would require the linker to find the address of foo, which would involve looking up "foo" in the symbol table and finding the address.

Closely related to symbols are relocations. A relocation is simply a blank space left to be patched up later. In the previous example, until the address of foo is known it can not be used. However, on a 32-bit system, we know the address of foo must be a 4-byte value, so any time the compiler needs to use that address (to say, assign a value) it can simply leave 4-bytes of blank space and keep a relocation that essentially says to the linker "place the real value of "foo" into the 4 bytes at this address". As mentioned, this requires the symbol "foo" to be resolved. Section 2.1, Relocations contains further information on relocations.

3.3 Sections and Segments

The ELF format specifies two "views" of an ELF file — that which is used for linking and that which is used for execution. This affords significant flexibility for systems designers.

We talk about sections in object code waiting to be linked into an executable. One or more sections map to a segment in the executable.

3.3.1 Segments

As we have done before, it is sometimes easier to look at the higher level of abstraction (segments) before inspecting the lower layers.

As we mentioned the ELF file has an header that describes the overall layout of the file. The ELF header actually points to another group of headers called the program headers. These headers describe to the operating system anything that might be required for it to load the binary into memory and execute it. Segments are described by program headers, but so are some other things required to get the executable running.

typedef struct {
          Elf32_Word p_type;
          Elf32_Off  p_offset;
          Elf32_Addr p_vaddr;
          Elf32_Addr p_paddr;
          Elf32_Word p_filesz;
          Elf32_Word p_memsz;
          Elf32_Word p_flags;
          Elf32_Word p_align;
Example The Program Header

The definition of the program header is seen in Example, The Program Header. You might have noticed from the ELF header definition above how there were fields e_phoff, e_phnum and e_phentsize; these are simply the offset in the file where the program headers start, how many program headers there are and how big each program header is. With these three bits of information you can easily find and read the program headers.

Program headers more than just segments. The p_type field defines just what the program header is defining. For example, if this field is PT_INTERP the header is defined as meaning a string pointer to an interpreter for the binary file. We discussed compiled versus interpreted languages previously and made the distinction that a compiler builds a binary which can be run in a stand alone fashion. Why should it need an interpreter? As always, the true picture is a little more complicated. There are several reasons why a modern system wants flexibility when loading executable files, and to do this some information can only be adequately acquired at the actual time the program is set up to run. We see this in future chapters where we look into dynamic linking. Consequently some minor changes might need to be made to the binary to allow it to work properly at runtime. Thus the usual interpreter of a binary file is the dynamic loader, so called because it takes the final steps to complete loading of the executable and prepare the binary image for running.

Segments are described with a value of PT_LOAD in the p_type field. Each segment is then described by the other fields in the program header. The p_offset field tells you how far into the file on disk the data for the segment is. The p_vaddr field tells you what address that data is to live at in virtual memory (p_addr describes the physical address, which is only really useful for small embedded systems that do not implement virtual memory). The two flags p_filesz and p_memsz work to tell you how big the segment is on disk and how big it should be in memory. If the memory size is greater than the disk size, then the overlap should be filled with zeros. In this way you can save considerable space in your binaries by not having to waste space for empty global variables. Finally p_flags indicates the permissions on the segment. Execute, read and write permissions can be specified in any combination; for example code segments should be marked as read and execute only, data sections as read and write with no execute.

There are a few other segment types defined in the program headers, they are described more fully in the standards specification.

3.3.2 Sections

As we have mentioned, sections make up segments. Sections are a way to organise the binary into logical areas to communicate information between the compiler and the linker. In some special binaries, such as the Linux kernel, sections are used in more specific ways (see Section 6.2, Custom sections).

We've seen how segments ultimately come down to a blob of data in a file on disk with some descriptions about where it should be loaded and what permissions it has. Sections have a similar header to segments, as shown in Example, Sections .

typedef struct {
          Elf32_Word sh_name;
          Elf32_Word sh_type;
          Elf32_Word sh_flags;
          Elf32_Addr sh_addr;
          Elf32_Off  sh_offset;
          Elf32_Word sh_size;
          Elf32_Word sh_link;
          Elf32_Word sh_info;
          Elf32_Word sh_addralign;
          Elf32_Word sh_entsize;
Example Sections

Sections have a few more types defined for the sh_type field; for example a section of type SH_PROGBITS is defined as a section that hold binary data for use by the program. Other flags say if this section is a symbol table (used by the linker or debugger for example) or maybe something for the dynamic loader. There are also more attributes, such as the allocate attribute which flags that this section will need memory allocated for it.

Below we will examine the program listed in Example, Sections .

#include <stdio.h>

int big_big_array[10*1024*1024];

char *a_string = "Hello, World!";

int a_var_with_value = 0x100;

int main(void)
	big_big_array[0] = 100;
	printf("%s\n", a_string);
	a_var_with_value += 20;
Example Sections

Example, Sections readelf output shows the output of readelf with some parts stripped clarity. Using this output we can analyse each part of our simple program and see where it ends up in the final output binary.

$ readelf --all ./sections
ELF Header:
  Size of section headers:           40 (bytes)
  Number of section headers:         37
  Section header string table index: 34

Section Headers:
  [Nr] Name              Type            Addr     Off    Size   ES Flg Lk Inf Al
  [ 0]                   NULL            00000000 000000 000000 00      0   0  0
  [ 1] .interp           PROGBITS        10000114 000114 00000d 00   A  0   0  1
  [ 2] .note.ABI-tag     NOTE            10000124 000124 000020 00   A  0   0  4
  [ 3] .hash             HASH            10000144 000144 00002c 04   A  4   0  4
  [ 4] .dynsym           DYNSYM          10000170 000170 000060 10   A  5   1  4
  [ 5] .dynstr           STRTAB          100001d0 0001d0 00005e 00   A  0   0  1
  [ 6] .gnu.version      VERSYM          1000022e 00022e 00000c 02   A  4   0  2
  [ 7] .gnu.version_r    VERNEED         1000023c 00023c 000020 00   A  5   1  4
  [ 8] .rela.dyn         RELA            1000025c 00025c 00000c 0c   A  4   0  4
  [ 9] .rela.plt         RELA            10000268 000268 000018 0c   A  4  25  4
  [10] .init             PROGBITS        10000280 000280 000028 00  AX  0   0  4
  [11] .text             PROGBITS        100002b0 0002b0 000560 00  AX  0   0 16
  [12] .fini             PROGBITS        10000810 000810 000020 00  AX  0   0  4
  [13] .rodata           PROGBITS        10000830 000830 000024 00   A  0   0  4
  [14] .sdata2           PROGBITS        10000854 000854 000000 00   A  0   0  4
  [15] .eh_frame         PROGBITS        10000854 000854 000004 00   A  0   0  4
  [16] .ctors            PROGBITS        10010858 000858 000008 00  WA  0   0  4
  [17] .dtors            PROGBITS        10010860 000860 000008 00  WA  0   0  4
  [18] .jcr              PROGBITS        10010868 000868 000004 00  WA  0   0  4
  [19] .got2             PROGBITS        1001086c 00086c 000010 00  WA  0   0  1
  [20] .dynamic          DYNAMIC         1001087c 00087c 0000c8 08  WA  5   0  4
  [21] .data             PROGBITS        10010944 000944 000008 00  WA  0   0  4
  [22] .got              PROGBITS        1001094c 00094c 000014 04 WAX  0   0  4
  [23] .sdata            PROGBITS        10010960 000960 000008 00  WA  0   0  4
  [24] .sbss             NOBITS          10010968 000968 000000 00  WA  0   0  1
  [25] .plt              NOBITS          10010968 000968 000060 00 WAX  0   0  4
  [26] .bss              NOBITS          100109c8 000968 2800004 00  WA  0   0  4
  [27] .comment          PROGBITS        00000000 000968 00018f 00      0   0  1
  [28] .debug_aranges    PROGBITS        00000000 000af8 000078 00      0   0  8
  [29] .debug_pubnames   PROGBITS        00000000 000b70 000025 00      0   0  1
  [30] .debug_info       PROGBITS        00000000 000b95 0002e5 00      0   0  1
  [31] .debug_abbrev     PROGBITS        00000000 000e7a 000076 00      0   0  1
  [32] .debug_line       PROGBITS        00000000 000ef0 0001de 00      0   0  1
  [33] .debug_str        PROGBITS        00000000 0010ce 0000f0 01  MS  0   0  1
  [34] .shstrtab         STRTAB          00000000 0011be 00013b 00      0   0  1
  [35] .symtab           SYMTAB          00000000 0018c4 000c90 10     36  65  4
  [36] .strtab           STRTAB          00000000 002554 000909 00      0   0  1
Key to Flags:
  W (write), A (alloc), X (execute), M (merge), S (strings)
  I (info), L (link order), G (group), x (unknown)
  O (extra OS processing required) o (OS specific), p (processor specific)

There are no section groups in this file.

Symbol table '.symtab' contains 201 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
    99: 100109cc 0x2800000 OBJECT  GLOBAL DEFAULT   26 big_big_array
   110: 10010960     4 OBJECT  GLOBAL DEFAULT   23 a_string
   130: 10010964     4 OBJECT  GLOBAL DEFAULT   23 a_var_with_value
   144: 10000430    96 FUNC    GLOBAL DEFAULT   11 main
Example Sections readelf output

Firstly, let us look at the variable big_big_array, which as the name suggests is a fairly large global array. If we skip down to the symbol table we can see that the variable is at location 0x100109cc which we can correlate to the .bss section in the section listing, since it starts just below it at 0x100109c8. Note the size, and how it is quite large. We mentioned that BSS is a standard part of a binary image since it would be silly to require that binary on disk have 10 megabytes of space allocated to it, when all of that space is going to be zero. Note that this section has a type of NOBITS meaning that it does not have any bytes on disk.

Thus the .bss section is defined for global variables whose value should be zero when the program starts. We have seen how the memory size can be different to the on disk size in our discussion of segments; variables being in the .bss section are an indication that they will be given zero value on program start.

The a_string variable lives in the .sdata section, which stands for small data. Small data (and the corresponding .sbss section) are sections available on some architectures where data can be reached by an offset from some known pointer. This means a fixed-value can be added to the base-address, making it faster to get to data in the sections as there are no extra lookups and loading of addresses into memory required. Most architectures are limited to the size of immediate values you can add to a register (e.g. if performing the instruction r1 = add r2, 70;, 70 is an immediate value, as opposed to say, adding two values stored in registers r1 = add r2,r3) and can thus only offset a certain "small" distance from an address. We can also see that our a_var_with_value lives in the same place.

main however lives in the .text section, as we expect (remember the name "text" and "code" are used interchangeably to refer to a program in memory.

3.3.3 Sections and Segments together

$ readelf --segments /bin/ls

Elf file type is EXEC (Executable file)
Entry point 0x100026c0
There are 8 program headers, starting at offset 52

Program Headers:
  Type           Offset   VirtAddr   PhysAddr   FileSiz MemSiz  Flg Align
  PHDR           0x000034 0x10000034 0x10000034 0x00100 0x00100 R E 0x4
  INTERP         0x000154 0x10000154 0x10000154 0x0000d 0x0000d R   0x1
      [Requesting program interpreter: /lib/ld.so.1]
  LOAD           0x000000 0x10000000 0x10000000 0x14d5c 0x14d5c R E 0x10000
  LOAD           0x014d60 0x10024d60 0x10024d60 0x002b0 0x00b7c RWE 0x10000
  DYNAMIC        0x014f00 0x10024f00 0x10024f00 0x000d8 0x000d8 RW  0x4
  NOTE           0x000164 0x10000164 0x10000164 0x00020 0x00020 R   0x4
  GNU_EH_FRAME   0x014d30 0x10014d30 0x10014d30 0x0002c 0x0002c R   0x4
  GNU_STACK      0x000000 0x00000000 0x00000000 0x00000 0x00000 RWE 0x4

 Section to Segment mapping:
  Segment Sections...
   01     .interp
   02     .interp .note.ABI-tag .hash .dynsym .dynstr .gnu.version .gnu.version_ r .rela.dyn .rela.plt .init .text .fini .rodata .eh_frame_hdr
   03     .data .eh_frame .got2 .dynamic .ctors .dtors .jcr .got .sdata .sbss .p lt .bss
   04     .dynamic
   05     .note.ABI-tag
   06     .eh_frame_hdr
Example Sections and Segments

Example, Sections and Segments shows how readelf shows us the segments and section mappings in the ELF file for the binary /bin/ls.

Skipping to the bottom of the output, we can see what sections have been moved into what segments. So, for example the .interp section is placed into an INTERP flagged segment. Notice that readelf tells us it is requesting the interpreter /lib/ld.so.1; this is the dynamic linker which is run to prepare the binary for execution.

Looking at the two LOAD segments we can see the distinction between text and data. Notice how the first one has only "read" and "execute" permissions, whilst the next one has read, write and execute permissions? These describe the code (r/w) and data (r/w/e) segments.

But data should not need to be executable! Indeed, on most architectures (for example, the most common x86) the data section will not be marked as having the data section executable. However, the example output above was taken from a PowerPC machine which has a slightly different programming model (ABI, see below) requiring that the data section be executable 1. Such is the life of a systems programmer, where rules were made to be broken!

The other interesting thing to note is that the file size is the same as the memory size for the code segment, however memory size is greater than the file size for the data segment. This comes from the BSS section which holds zeroed global variables.