This page is a copy of the Archive.org copy of the now no longer availabel http://www.acsu.buffalo.edu/~charngda/elf.html. It is kept here online as a reference only.

Acronyms relevant to Executable and Linkable Format (ELF)

Useful books and references

ELF man page

System V Application Binary Interface

AMD64 System V Application Binary Interface

The gen on function calling conventions

Section II of Linux Standard Base 4.0 Core Specification

Self-Service Linux: Mastering the Art of Problem Determination by Mark Wilding and Dan Behman

Solaris Linker and Libraries Guide

Linkers and Loaders by John Levine

Understanding Linux ELF RTLD internals by mayhem (this article gives you an idea how the runtime linker ld.so works)

ld.so man page

Prelink by Jakub Jelinek (and prelink man page)

Executable and Linkable Format

An ELF executable binary contains at least two kinds of headers: ELF file header (see struct Elf32_Ehdr/struct Elf64_Ehdr in /usr/include/elf.h) and one or more Program Headers (see struct Elf32_Phdr/struct Elf64_Phdr in /usr/include/elf.h)

Usually there is another kind of header called Section Header, which describe attributes of an ELF section (e.g. .text, .data, .bss, etc) The Section Header is described by struct Elf32_Shdr/struct Elf64_Shdr in /usr/include/elf.h

The Program Headers are used during execution (ELF’s “execution view”); it tells the kernel or the runtime linker ld.so what to load into memory and how to find dynamic linking information.

The Section Headers are used during compile-time linking (ELF’s “linking view”); it tells the link editor ld how to resolve symbols, and how to group similar byte streams from different ELF binary objects.

Conceptually, the two ELF’s “views” are as follows (borrowed from Shaun Clowes’s Fixing/Making Holes in Binaries slides):

              +-----------------+
         +----| ELF File Header |----+
         |    +-----------------+    |
         v                           v
 +-----------------+      +-----------------+
 | Program Headers |      | Section Headers |
 +-----------------+      +-----------------+
      ||                               ||
      ||                               ||
      ||                               ||
      ||   +------------------------+  ||
      +--> | Contents (Byte Stream) |<--+
           +------------------------+

In reality, the layout of a typical ELF executable binary on a disk file is like this:

    +-------------------------------+
    | ELF File Header               |
    +-------------------------------+
    | Program Header for segment #1 |
    +-------------------------------+
    | Program Header for segment #2 |
    +-------------------------------+
    | ...                           |
    +-------------------------------+
    | Contents (Byte Stream)        |
    | ...                           |
    +-------------------------------+
    | Section Header for section #1 |
    +-------------------------------+
    | Section Header for section #2 |
    +-------------------------------+
    | ...                           |
    +-------------------------------+
    | ".shstrtab" section           |
    +-------------------------------+
    | ".symtab"   section           |
    +-------------------------------+
    | ".strtab"   section           |
    +-------------------------------+

The ELF File Header contains the file offsets of the first Program Header, the first Section Header, and .shstrtab section which contains the section names (a series of NULL-terminated strings)

The ELF File Header also contains the number of Program Headers and the number of Section Headers.

Each Program Header describes a “segment”: It contains the permissions (Readable, Writeable, or Executable) , offset of the “segment” (which is just a byte stream) into the file, and the size of the “segment”. The following table shows the purposes of special segments. Some information can be found in GNU Binutil’s source file include/elf/common.h:

Likewise, each Section Header contains the file offset of its corresponding “content” and the size of the “content”. The following table shows the purposes of some special sections. Most information here comes from LSB specification. Some information can be found in GNU Binutil’s source file bfd/elf.c (look for bfd_elf_special_section) and bfd/elflink.c (look for double-quoted section names such as ".got.plt")

How is an executable binary in Linux being executed ?

First, the operating system must recognize executable binaries. For example, zcat /proc/config.gz | grep CONFIG_BINFMT_ELF can show whether the Linux kernel is compiled to support ELF executable binary format (if /proc/config.gz does not exist, try /lib/modules/`uname -r`/build/.config)

When the shell makes an execvc system call to run an executable binary, the Linux kernel responds as follows (see here and here for more details) in sequence:

sys_execve function (in arch/x86/kernel/process.c) handles the execvc system call from user space. It calls do_execve function.
do_execve function (in fs/exec.c) opens the executable binary file and does some preparation. It calls search_binary_handler function.
search_binary_handler function (in fs/exec.c) finds out the type of executable binary and calls the corresponding handler, which in our case, is load_elf_binary function.
load_elf_binary (in fs/binfmt_elf.c) loads the user’s executable binary file into memory. It allocates memory segments and zeros out the BSS section by calling the padzero function.

load_elf_binary also examines whether the user’s executable binary contains an INTERP segment or not.
If the executable binary is dynamically linked, then the compiler will usually creates an INTERP segment (which is usually the same as .interp section in ELF’s “linking view”), which contains the full pathname of an “interpreter”, usually is the Glibc runtime linker ld.so.

To see this, use command readelf -p .interp a.out

According to AMD64 System V Application Binary Interface, the only valid interpreter for programs conforming to AMD64 ABI is /lib/ld64.so.1 and on Linux, GCC usually uses /lib64/ld-linux-x86-64.so.2 or /lib/ld-linux-x86-64.so.2 instead:
```
$ gcc -dumpspecs
....

*link:
...
  %{!m32:%{!dynamic-linker:-dynamic-linker %{muclibc:%{mglibc:%e-mglibc and -muclibc used
together}/lib/ld64-uClibc.so.0;:/lib/ld-linux-x86-64.so.2}}}}
...
```
To change the runtime linker, compile the program using something like
```
gcc foo.c -Wl,-I/my/own/ld.so
```
The System V Application Binary Interface specifies, the operating system, instead of running the user’s executable binary, should run this “interpreter”. This interpreter should complete the binding of user’s executable binary to its dependencies.
Thus, if the ELF executable binary file contains an INTERP segment, load_elf_binary will call load_elf_interp function to load the image of this interpreter as well.
Finally, load_elf_binary calls start_thread (in arch/x86/kernel/process_64.c) and passes control to either the interpreter or the user program.

What about `ld.so` ?

ld.so is the runtime linker/loader (the compile-time linker ld is formally called “link editor”) for dynamic executables. It provides the following services:

Analyzes the user’s executable binary’s DYNAMIC segment and determines what dependencies are required. (See below)
Locates and loads these dependencies, analyzes their DYNAMIC segments to determine if more dependencies are required.
Performs any necessary relocations to bind these objects.
Calls any initialization functions (see below) provided by these dependencies.
Passes control to user’s executable binary.

Compile your own `ld.so`

The internal working of ld.so is complex, so you might want to compile and experiment your own ld.so. The source code of ld.so can be found in Glibc. The main files are elf/rtld.c, elf/dl-reloc.c, and sysdeps/x86_64/dl-machine.h.

This link provides general tips for building Glibc. Glibc’s own INSTALL and FAQ documents are useful too.

To compile Glibc (ld.so cannot be compiled independently) download and unpack Glibc source tarball.

Make sure the version of Glibc you downloaded is the same as the system’s current one.
Make sure the environmental variable LD_RUN_PATH is not set.
Read the INSTALL and make sure all necessary tool chains (Make, Binutils, etc) are up-to-date.

Make sure the file system you are doing the compilation is case sensitive, or you will see weird errors like

/scratch/elf/librtld.os: In function `process_envvars':
/tmp/glibc-2.x.y/elf/rtld.c:2718: undefined reference to `__open'
...

ld.so should be compiled with the optimization flag on (-O2 is the default). Failing to do so will end up with weird errors (see Question 1.23 in FAQ)
Suppose Glibc is unpacked at
```
/tmp/glibc-2.x.y/
```
Then edit /tmp/glibc-2.x.y/Makefile.in: Un-comment the line
```
# PARALLELMFLAGS = -j 4
```
and change 4 to an appropriate number.
Since we are only interested in ld.so and not the whole Glibc, we only want to build the essential source files needed by ld.so. To do so, edit /tmp/glibc-2.x.y/Makeconfig: Find the line started with
```
all-subdirs = csu assert ctype locale intl catgets math setjmp signal \
  ...
```
and change it to
```
all-subdirs = csu elf gmon io misc posix setjmp signal stdlib string time
```

Find a scratch directory, say /scratch. Then

$ cd /scratch
$ /tmp/glibc-2.x.y/configure --prefix=/scratch --disable-profile
$ gmake

Since we are not building the entire Glibc, when the gmake stops (probably with some errors), check if /scratch/elf/ld.so exists or not.
ld.so is a static binary, which means it has its own implementation of standard C routines (e.g. memcpy, strcmp, etc) It has its own printf-like routine called _dl_debug_printf.

_dl_debug_printf is not the full-blown printf and has very limited capabilities. For example, to print the address, one would need to use
```
_dl_debug_printf("0x%0*lx\n", (int)sizeof (void*)*2, &foo);
```

How does `ld.so` work ?

ld.so, by its nature, cannot be a dynamic executable itself. The entry point of ld.so is _start defined in the macro RTLD_START (in sysdeps/x86_64/dl-machine.h). _start is placed at the beginning of .text section, and the default ld script specifies “Entry point address” (in ELF header, use readelf -h ld.so|grep Entry command to see) to be the address of _start (use ld -verbose | grep ENTRY command to see). One can set the entry point to a different address at compile time by -e option) so ld.so is executed from here. The very first thing it does is to call _dl_start of elf/rtld.c. To see this, run gdb on some ELF executable binary, and do

(gdb) break _dl_start
Function "_dl_start" not defined.
Make breakpoint pending on future shared library load? (y or [n]) y
Breakpoint 1 (_dl_start) pending.
(gdb) run
Starting program: a.out

Breakpoint 1, 0x0000003433e00fa0 in _dl_start () from /lib64/ld-linux-x86-64.so.2
(gdb) bt
#0  0x0000003433e00fa0 in _dl_start () from /lib64/ld-linux-x86-64.so.2
#1  0x0000003433e00a78 in _start () from /lib64/ld-linux-x86-64.so.2
#2  0x0000000000000001 in ?? ()
#3  0x00007fffffffe4f2 in ?? ()
#4  0x0000000000000000 in ?? ()
...
(gdb) x/10i $pc
   0x3433e00a70 <_start>:       mov    %rsp,%rdi
   0x3433e00a73 <_start+3>:     callq  0x3433e00fa0 <_dl_start>
   0x3433e00a78 <_dl_start_user>:       mov    %rax,%r12
   0x3433e00a7b <_dl_start_user+3>:     mov    0x21b30b(%rip),%eax        # 0x343401bd8c <_dl_skip_args>
...

At this breakpoint, we can use pmap to see the memory map of a.out, which would look like this:

0000000000400000      8K r-x--  a.out
0000000000601000      4K rw---  a.out
0000003433e00000    112K r-x--  /lib64/ld-2.5.so
000000343401b000      8K rw---  /lib64/ld-2.5.so
00007ffffffea000     84K rw---    [ stack ]
ffffffffff600000   8192K -----    [ anon ]
 total             8408K

The memory segment of /lib64/ld-2.5.so indeed starts at 3433e00000 (page aligned) and this can be verified by running readelf -t /lib64/ld-2.5.so.

If we put another breakpoint at main and continue, then when it stops, the memory map would change to this:

0000000000400000      8K r-x--  a.out
0000000000601000      4K rw---  a.out
0000003433e00000    112K r-x--  /lib64/ld-2.5.so
000000343401b000      4K r----  /lib64/ld-2.5.so
000000343401c000      4K rw---  /lib64/ld-2.5.so
0000003434200000   1336K r-x--  /lib64/libc-2.5.so     <-- The first "LOAD" segment, which contains .text and .rodata sections
000000343434e000   2044K -----  /lib64/libc-2.5.so     <-- "Hole"
000000343454d000     16K r----  /lib64/libc-2.5.so     <-- Relocation (GNU_RELRO) info  -+---- The second "LOAD" segment
0000003434551000      4K rw---  /lib64/libc-2.5.so     <-- .got.plt .data sections      -+
0000003434552000     20K rw---    [ anon ]             <-- The remaining zero-filled sections (e.g. .bss)
0000003434e00000     88K r-x--  /lib64/libpthread-2.5.so     <-- The first "LOAD" segment, which contains .text and .rodata sections
0000003434e16000   2044K -----  /lib64/libpthread-2.5.so     <-- "Hole"
0000003435015000      4K r----  /lib64/libpthread-2.5.so     <-- Relocation (GNU_RELRO) info  -+---- The second "LOAD" segment
0000003435016000      4K rw---  /lib64/libpthread-2.5.so     <-- .got.plt .data sections      -+
0000003435017000     16K rw---    [ anon ]                   <-- The remaining zero-filled sections (e.g. .bss)
00002aaaaaaab000      4K rw---    [ anon ]
00002aaaaaac6000     12K rw---    [ anon ]
00007ffffffea000     84K rw---    [ stack ]
ffffffffff600000   8192K -----    [ anon ]
 total            14000K

Indeed, ld.so has brought in all the required dynamic libraries.

Note that there are two memory regions of 2044KB with null permissions. As mentioned earlier, the ELF’s ‘execution view’ is concerned with how to load an executable binary into memory. When ld.so brings in the dynamic libraries, it looks at the segments labelled as LOAD (look at “Program Headers” and “Section to Segment mapping” from readelf -a xxx.so command.) Usually there are two LOAD segments, and there is a “hole” between the two segments (look at the VirtAddr and MemSiz of these two segments), so ld.so will make this hole inaccessible deliberately: Look for the PROT_NONE symbol in _dl_map_object_from_fd in elf/dl-load.c

Also note that each of libc-2.5.so and libpthread-2.5.so has a read-only memory region (at 0x343454d000 and 0x3435015000, respectively). This is a for elf/dl-reloc.c. The GNU_RELRO segment is contained in the the second LOAD segment, which contains the following sections (look at “Program Headers” and “Section to Segment mapping” from readelf -l xxx.so command): .tdata, .fini_array, .ctors, .dtors, __libc_subfreeres, __libc_atexit, __libc_thread_subfreeres, .data.rel.ro, .dynamic, .got, .got.plt, .data, and .bss. Except for .got.plt, .data, and .bss, all sections in the the second LOAD segment are also in the GNU_RELRO segment, and they are thus made read-only.

The two [anon] memory segments at 0x3434552000 and 0x3435017000 are for sections which do not take space in the ELF binary files. For example, readelf -t xxx.so will show that .bss section has NOBITS flag, which means that section takes no disk space. When segments containing NOBITS sections are mapped into memory, ld.so allocates extra memory pages to accomodate these NOBITS sections. A LOAD segment is usually structured as a series of contiguous sections, and if a segment contains NOBITS sections, these NOBITS sections will be grouped together and placed at the tail of the segment.

So what does _dl_start do ?

Allocate the initial TLS block and initialize the Thread Pointer if needed (these are for ld.so’s own, not for the user program)
Call _dl_sysdep_start, which will call dl_main
dl_main does the majority of the hard work, for example:

It calls process_envvars to handle these LD_ prefix environmental variables such as LD_PRELOAD, LD_LIBRARY_PATH.

It examines the NEEDED field(s) in the user executable binary’s DYNAMIC segment section (see below) to determine the dependencies.

It calls _dl_init_paths (in elf/dl-load.c) to initialize the dynamic libraries search paths. According to ld.so man page and this page, the dynamic libraries are searched in the following order:
1. The RPATH in the DYNAMIC segment if there is no RUNPATH in the DYNAMIC segment.
  
  RPATH can be specified when the code is compiled with gcc -Wl,-rpath=...
  
  Use of RPATH is deprecated because it has an obvious drawback: There is no way to override it except using LD_PRELOAD environmental variable or removing it from the DYNAMIC segment.
  
  Both RPATH and RUNPATH can contain $ORIGIN (or equivalently ${ORIGIN}), which will be expanded to the value of environmental variable LD_ORIGIN_PATH or the full path of the loaded object (unless the programs use setuid or setgid)
2. The LD_LIBRARY_PATH environmental variable (unless the programs use setuid or setgid)
3. The RUNPATH in the DYNAMIC segment.
  RUNPATH can be specified when the code is compiled with gcc -Wl,-rpath=...,--enable-new-dtags
  One can use chrpath tool to manipulate RPATH and RUNPATH settings.
4. /etc/ld.so.cache
5. /lib
6. /usr/lib
It calls _dl_map_object_from_fd (in elf/dl-load.c) to load the dynamic libraries, sets up the right read/write/execute permissions for the memory segments, (within _dl_map_object_from_fd, look at calls to mmap, mprotect and symbols such as PROT_READ, PROT_WRITE, PROT_EXEC, PROT_NONE), zeroes out BSS sections of dynamic libraries (inside _dl_map_object_from_fd function, look at calls to memset), updates the link map, and performs relocations.

It calls _dl_relocate_object (in elf/dl-reloc.c) to perform runtime relocations (see details below).
When _dl_start returns, it continues to execute code in _dl_start_user (see sysdeps/x86_64/dl-machine.h)
_dl_start_user will call _dl_init_internal, which will call call_init to invoke initialization function of each dynamic library loaded.

Note that _dl_init_internal is defined in elf/dl-init.c as:
```
void
internal_function
_dl_init (struct link_map *main_map, int argc, char **argv, char **env)
```
call_init is also in elf/dl-init.c
The initialization function of a dynamic library, say libfoo.so, is located at the address marked with type “INIT” in the output of readelf -d libfoo.so For Glibc, its initialization function is named _init (not to be confused with the _init inside the user’s executable binary) and its source code is in sysdeps/unix/sysv/linux/x86_64/init-first.c.

_init will do the following things:
- Save argc, argv, envp to hidden variables __libc_argc, __libc_argv, __environ
- Call VDSO_SETUP to set up Virtaul Dynamic Shared Objects (see here) VDSO_SETUP is a platform-dependent macro. For x86_64, this macro is defined as _libc_vdso_platform_setup in sysdeps/unix/sysv/linux/x86_64/init-first.c
- Call __init_misc (in misc/init-misc.c) which saves argv[0] to two global variables: program_invocation_name and program_invocation_short_name
- Call __libc_global_ctors (in elf/soinit.c) to invoke each function listed in the .ctors section (see below).
  
  For x86_64, .ctors section contains only one function: init_cacheinfo
At the end of _dl_start_user, the control transfers to user program’s entry point address (use readelf -h a.out|grep Entry to see) which is usually the initial address of .text section and contains the entry of a function named _start, and in the control transfer, the finalizer function _dl_fini is passed as an argument, and the stack frames are completely clobbered, as if the user program is run without any ld.so intervention. The latter is done by manipulating the stack (see the on-stack auxiliary vector adjustment code and HAVE_AUX_VECTOR in dl_main)

Here is the call graph, which is worth a thousand words

and see here on how it is generated.

To see ld.so in action, set the environmental variable LD_DEBUG to all and then run a user program.

The above debugging information does not show mmap and mprotect calls. However, we can use strace. If we run the user program again with

strace -e trace=mmap,mprotect,munmap,open a.out

we should see something like the following:

mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x2ae62c0d1000

  .... (a lot of failed attempts to open 'libpthread.so.0' using LD_LIBRARY_PATH)

open("/etc/ld.so.cache", O_RDONLY) = 3
mmap(NULL, 104801, PROT_READ, MAP_PRIVATE, 3, 0) = 0x2ae62c0d2000
open("/lib64/libpthread.so.0", O_RDONLY) = 3
mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x2ae62c0ec000
mmap(0x3434e00000, 2204528, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x3434e00000     <-- Bring in the first "LOAD" segment
mprotect(0x3434e16000, 2093056, PROT_NONE) = 0     <-- Make the "hole" inaccessible
mmap(0x3435015000, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x15000) = 0x3435015000     <-- Bring in the second "LOAD" segment
mmap(0x3435017000, 13168, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x3435017000
       (note: 0x3435017000 is the [anon] part which follows immediately after libpthread-2.5.so)
  ...
  .... (a lot of failed attempts to open 'libc.so.6' using LD_LIBRARY_PATH)

open("/lib64/libc.so.6", O_RDONLY) = 3
mmap(0x3434200000, 3498328, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x3434200000     <-- Bring in the first "LOAD" segment
mprotect(0x343434e000, 2093056, PROT_NONE) = 0     <-- Make the "hole" inaccessible
mmap(0x343454d000, 20480, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x14d000) = 0x343454d000     <-- Bring in the second "LOAD" segment
mmap(0x3434552000, 16728, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_ANONYMOUS, -1, 0) = 0x3434552000
       (note: 0x3434552000 is the [anon] part which follows immediately after libc-2.5.so)
mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x2ae62c0ed000
mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x2ae62c0ee000
mprotect(0x343454d000, 16384, PROT_READ) = 0    <-- Make the GNU_RELRO segment read-only
mprotect(0x3435015000, 4096, PROT_READ) = 0     <-- Make the GNU_RELRO segment read-only
mprotect(0x343401b000, 4096, PROT_READ) = 0
munmap(0x2ae62c0d2000, 104801)= 0
mmap(NULL, 10489856, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS|MAP_32BIT, -1, 0) = 0x40dc7000
mprotect(0x40dc7000, 4096, PROT_NONE)   = 0
mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x2aaaaaaab000

`.plt` section

This section contains trampolines for functions defined in dynamic libraries. A sample disassembly (run the command objdump -M intel -dj .plt a.out) will show the following:

4003c0 <printf@plt-0x10>:
  4003c0: push   QWORD PTR [RIP+0x2004d2]        # 600898 <_GLOBAL_OFFSET_TABLE_+0x8>
  4003c6: jmp    QWORD PTR [RIP+0x2004d4]        # 6008a0 <_GLOBAL_OFFSET_TABLE_+0x10>
  4003cc: nop    DWORD PTR [RAX+0x0]

4003d0 <printf@plt>:
  4003d0: jmp    QWORD PTR [RIP+0x2004d2]        # 6008a8 <_GLOBAL_OFFSET_TABLE_+0x18>
  4003d6: push   0
  4003db: jmp    4003c0 <printf@plt-0x10>

4003e0 <__libc_start_main@plt>:
  4003e0: jmp    QWORD PTR [RIP+0x2004ca]        # 6008b0 <_GLOBAL_OFFSET_TABLE_+0x20>
  4003e6: push   1
  4003eb: jmp    4003c0 <printf@plt-0x10>

The _GLOBAL_OFFSET_TABLE_ (labeled as R_X86_64_JUMP_SLOT and starts at address 0x600890) is located in .got.plt section (to see this, run the command objdump -h a.out |grep -A 1 600890 or the command readelf -r a.out) The data in .got.plt section look like the following during runtime (use gdb to see them)

(gdb) b *0x4003d0
(gdb) run
(gdb) x/6a 0x600890
0x600890: 0x6006e8 <_DYNAMIC>     0x32696159a8
0x6008a0: 0x326950aa20 <_dl_runtime_resolve>     0x4003d6 <printf@plt+6>
0x6008b0: 0x326971c3f0 <__libc_start_main>       0x0

When printf is called the first time in the user program, the jump at 4003d0 will jump to 4003d6, which is just the next instruction (push 0) The it jumps to 4003c0, which does not have a function name (so it is shown as <printf@plt-0x10>). At 4003c6, it will jumps to _dl_runtime_resolve. This function (in Glibc’s source file sysdeps/x86_64/dl-trampoline.S) is a trampoline to _dl_fixup (in Glibc’s source file elf/dl-runtime.c). _dl_fixup again, is part of Glibc runtime linker ld.so. In particular, it will change the address stored at 6008a8 to the actual address of printf in libc.so.6. To see this, set up a hardware watchpoint

(gdb) watch *0x6008a8
(gdb) cont
Continuing.
Hardware watchpoint 2: *0x6008a8

Old value = 4195286
New value = 1769244016
0x000000326950abc2 in fixup () from /lib64/ld-linux-x86-64.so.2

If we continue execution, printf will be called, as expected. When printf is called again in the user program, the jump at 4003d0 will bounce directly to printf:

(gdb) x/6a 0x600890
0x600890: 0x6006e8 <_DYNAMIC>     0x32696159a8
0x6008a0: 0x326950aa20 <_dl_runtime_resolve>     0x3269748570 <printf>
0x6008b0: 0x326971c3f0 <__libc_start_main>       0x0

`.init`, `.fini`, `.preinit_array`, `.init_array` and `.fini_array` sections

.init and .fini sections contain code to do initialization and termination, as specified by the System V Application Binary Interface. If the code is compiled by GCC, then one will see the following code in .init and .fini sections, respectively:

4003a8 <_init>:
  4003a8: sub    RSP, 8
  4003ac: call   call_gmon_start
  4003b1: call   frame_dummy
  4003b6: call   __do_global_ctors_aux
  4003bb: add    RSP, 8
  4003bf: ret

400618 <_fini>:
  400618: sub    RSP, 8
  40061c: call   __do_global_dtors_aux
  400621: add    RSP, 8
  400625: ret

There is only one function: _init, in .init section, and likewise, only one function: _fini in .fini section. Both _init and _fini are synthesized at compile time by the compiler/linker. Glibc provides its own prolog and epilog for _init and _fini, but the compiler is free to choose how to use them and add more code into _init and _fini.

In Glibc, the source file sysdeps/generic/initfini.c (and some system dependent ones, such as sysdeps/x86_64/elf/initfini.c) is compiled into two files: /usr/lib64/crti.o for prolog and /usr/lib64/crtn.o for epilog.

For the compiler part, GCC uses different prolog and epilog files, depending on the compiler command-line options. To see them, execute gcc -dumpspec, and one can see

...

*endfile:
  %{ffast-math|funsafe-math-optimizations:crtfastmath.o%s}
  %{mpc32:crtprec32.o%s}
  %{mpc64:crtprec64.o%s}
  %{mpc80:crtprec80.o%s}
  %{shared|pie:crtendS.o%s;:crtend.o%s}
  crtn.o%s

...

*startfile:
  %{!shared: %{pg|p|profile:gcrt1.o%s;pie:Scrt1.o%s;:crt1.o%s}}
  crti.o%s
  %{static:crtbeginT.o%s;shared|pie:crtbeginS.o%s;:crtbegin.o%s}

...

The detailed explanation of GCC spec file is here. For above snippet, it means, for example, if compiler command-line option -ffast-math is used, include GCC’s crtfastmath.o file (this file can be found under /usr/lib/gcc/<arch>/<version>/) at the end of the linking process. Glibc’s crtn.o is always included at the end of linking. The %s means this preceding file is a startup file. (GCC allows to skip startup files during linking using -nostartfiles compiler option)

Similarly, if -shared compiler command-line option is not used, then always include Glibc’s crt1.o at the start of the linking process. crt1.o contains the function _start in .text section (not .init section!) _start is the function that is executed before anything else… see below. Next, include Glibc’s crti.o in the linking. Finally, include either crtbeginT.o, crtbeginS.o, or crtbegin.o (both are part of GCC, of course), depending on whether -static or -shared (or neither) is used.

So, for example, if a program is compiled using dynamic linking (which is default), no profiling, no fast math optimizations, then the linking will include the following files in the following order:

crt1.o (part of Glibc)
crti.o (part of Glibc and contributes the code at 4003a8, 4003ac, 400618, and the body of call_gmon_start)
crtbegin.o (part of GCC and contributes the code at 4003b1 and 40061c, and the body of frame_dummy and __do_global_dtors_aux)
user’s code
crtend.o (part of GCC and contributes the code at 4003b6 and the body of __do_global_ctors_aux)
crtn.o (part of Glibc and contributes the code at 4003bb, 4003bf, 400621, 400625)

Why __do_global_ctors_aux is in crtend*.o and __do_global_dtors_aux is in crtbegin*.o ? Recall the order of invocation of destructors should be the reverse order of invocation of constructors. Therefore, GCC doing so will ensure __do_global_ctors_aux is called as late as possible in .init section and __do_global_dtors_aux is called as early as possible in .fini section.

Now back to the 4003a8 <_init>.

call_gmon_start is part of the Glibc prolog /usr/lib64/crti.o. It initializes gprof related data structures.

frame_dummy is in GCC code gcc/crtstuff.c and it is used to set up excepion handling and Java class registration (JCR) information.

The most interesting code is __do_global_ctors_aux (in GCC’s gcc/crtstuff.c and gcc/gbl-ctors.h) What it does is to call functions which are marked as __attribute__ ((constructor)) (and static C++ objects’ constructors) one by one:

  __SIZE_TYPE__ nptrs = (__SIZE_TYPE__) __CTOR_LIST__[0];
  unsigned i;

  if (nptrs == (__SIZE_TYPE__)-1)
    for (nptrs = 0; __CTOR_LIST__[nptrs + 1] != 0; nptrs++);

  for (i = nptrs; i >= 1; i--)
    __CTOR_LIST__[i] ();

The array __CTOR_LIST__ is stored in a special section called .ctors. Suppose a function called foo is marked as __attribute__ ((constructor)), then the runtime call stack trace would be

(gdb) break foo
(gdb) run
(gdb) bt
#0  0x00000000004004d8 in foo ()
#1  0x0000000000400606 in __do_global_ctors_aux ()
#2  0x00000000004003bb in _init ()
#3  0x00000000004005a0 in ?? ()
#4  0x0000000000400561 in __libc_csu_init ()
#5  0x000000326971c46f in __libc_start_main ()
#6  0x000000000040041a in _start ()

Similarly, the __do_global_dtors_aux in _fini function will invoke all functions which are marked as __attribute__ ((destructor)). __do_global_dtors_aux code is also in GCC’s source tree at gcc/crtstuff.c. If a function called foo is marked as __attribute__ ((destructor)) (and static C++ objects’ destructors), then the runtime call stack trace would be

(gdb) bt
#0  0x0000000000400518 in foo ()
#1  0x00000000004004ca in __do_global_dtors_aux ()
#2  0x0000000000400641 in _fini ()
#3  0x00000032699367e8 in ?? () from /lib64/tls/libc.so.6
#4  0x0000003269730c95 in exit () from /lib64/tls/libc.so.6
#5  0x000000326971c4d2 in __libc_start_main () from /lib64/tls/libc.so.6
#6  0x000000000040045a in _start ()

The array __DTOR_LIST__ contains the addresses of these destructors and it is stored in a special section called .dtors.

What user functions will be executed before `main` and at program exit?

As above call strack trace shows, _init is NOT the only function to be called before main. It is __libc_csu_init function (in Glibc’s source file csu/elf-init.c) that determines what functions to be run before main and the order of running them. Its code is like this

   void __libc_csu_init (int argc, char **argv, char **envp)
   {
   #ifndef LIBC_NONSHARED
     {
       const size_t size = __preinit_array_end - __preinit_array_start;
       size_t i;
       for (i = 0; i < size; i++)
         (*__preinit_array_start [i]) (argc, argv, envp);
     }
   #endif

     _init ();

     const size_t size = __init_array_end - __init_array_start;
     for (size_t i = 0; i < size; i++)
         (*__init_array_start [i]) (argc, argv, envp);
   }

(Symbols such as __preinit_array_start, __preinit_array_end, __init_array_start, __init_array_end are defined by the default ld script; look for PROVIDE and PROVIDE_HIDDEN keywords in the output of ld -verbose command.)

The __libc_csu_fini function has similar code, but what functions to be executed at program exit are actually determined by exit:

    void __libc_csu_fini (void)
    {
    #ifndef LIBC_NONSHARED
      size_t i = __fini_array_end - __fini_array_start;
      while (i-- > 0)
        (*__fini_array_start [i]) ();

      _fini ();
    #endif
    }

To see what’s going on, consider the following C code example:

   #include <stdio.h>
   #include <stdlib.h>

   void preinit(int argc, char **argv, char **envp) {
     printf("%s\n", __FUNCTION__);
   }

   void init(int argc, char **argv, char **envp) {
     printf("%s\n", __FUNCTION__);
   }

   void fini() {
     printf("%s\n", __FUNCTION__);
   }

   __attribute__((section(".init_array"))) typeof(init) *__init = init;
   __attribute__((section(".preinit_array"))) typeof(preinit) *__preinit = preinit;
   __attribute__((section(".fini_array"))) typeof(fini) *__fini = fini;

   void  __attribute__ ((constructor)) constructor() {
     printf("%s\n", __FUNCTION__);
   }

   void __attribute__ ((destructor)) destructor() {
     printf("%s\n", __FUNCTION__);
   }

   void my_atexit() {
     printf("%s\n", __FUNCTION__);
   }

   void my_atexit2() {
     printf("%s\n", __FUNCTION__);
   }

   int main() {
     atexit(my_atexit);
     atexit(my_atexit2);
   }

The output will be

   preinit
   constructor
   init
   my_atexit2
   my_atexit
   fini
   destructor

The .preinit_array and .init_array sections must contain function pointers (NOT code!) The prototype of these functions must be

void func(int argc,char** argv,char** envp)

__libc_csu_init execute them in the following order:

Function pointers in .preinit_array section
Functions marked as __attribute__ ((constructor)), via _init
Function pointers in .init_array section

The .fini_array section must also contain function pointers and the prototype is like the destructor, i.e. taking no arguments and returning void. If the program exits normally, then the exit function (Glibc source file stdlib/exit.c) is called and it will do the following:

In reverse order, functions registered via atexit or on_exit
Function pointers in .fini_array section, via __libc_csu_fini
Functions marked as __attribute__ ((destructor)), via __libc_csu_fini (which calls _fini after Step 2)
stdio cleanup functions

It is not advisable to put a code in .init section, e.g.

void __attribute__((section(".init"))) foo() {
  ...
}

because doing so will cause __do_global_ctors_aux NOT to be called. The .init section will now look like this:

4003a0 <_init>:
  4003a0: sub    RSP, 8
  4003a4: call   call_gmon_start
  4003a9: call   frame_dummy

4003ae <foo>:
  4003ae: push   RBP
  4003af: mov    RBP, RSP

     ....  (foo's body)

  4003b2: leave
  4003b3: ret
  4003b4: call   __do_global_ctors_aux
  4003b9: add    RSP, 8
  4003bd: ret

Now .init section contains more than one function, but the epilog of _init is distorted by the insertion of foo

Similarly, it is not advisable to put a code in .fini section, because otherwise the code will look like this:

4006d8 <_fini>:
  4006d8: sub    RSP, 8
  4006dc: call   __do_global_dtors_aux

4006e1 <foo>:
  4006e1: push   RBP
  4006e2: mov    RBP, RSP

     ....  (foo's body)

  4006ef: leave
  4006f0: ret
  4006f1: add    RSP, 8
  4006f5: ret

Now the epilog of _fini is distorted by the insertion of foo, so the stack frame pointer will not be adjusted (add RSP, 8 is not executed), causing segmentation fault.

What do `_start` and `__libc_start_main` do?

The above call stack traces show that _start calls __libc_start_main, which runs all of the code before main.

_start is part of Glibc code, as in sysdeps/x86_64/elf/start.S. As mentioned earlier, it is compiled as /usr/lib64/crt1.o and is statically linked to user’s executable binary during compilation. To see this, run gcc with -v command, and the last line would be something like:

.../collect2 ... /usr/lib64/crt1.o /usr/lib64/crti.o ...  /usr/lib64/crtn.o

_start is always placed at the beginning of .text section, and the default ld script specifies “Entry point address” (in ELF header, use readelf -h ld.so|grep Entry command to see) to be the address of _start (use ld -verbose | grep ENTRY command to see), so _start is guaranteed to be run before anything else. (This is changeable, however, at compile time one can specify a different initial address by -e option)

_start does only one thing: It sets up the arguments needed by __libc_start_main and then call it. __libc_start_main’s source code is csu/libc-start.c and its function prototype is:

__libc_start_main (int (*main) (int, char **, char **),
                   int argc,
                   char *argv,
                   int  (*init) (int, char **, char **),
                   void (*fini) (void),
                   void (*rtld_fini) (void),
                   void *stack_end)
                  )

__libc_start_main does quite a lot of work in addition to kicking off __libc_csu_init:

Set up argv and envp
Initialize the thread local storage by calling __pthread_initialize_minimal (which only calls __libc_setup_tls).

__libc_setup_tls will initialize Thread Control Block and Dynamic Thread Vector.
Set up the thread stack guard
Register the destructor (i.e. the rtld_fini argument passed to __libc_start_main) of the dynamic linker (by calling __cxa_atexit) if there is any
Initialize Glibc inself by calling __libc_init_first
Register __libc_csu_fini (i.e. the fini argument passed to __libc_start_main) using __cxa_atexit
Call __libc_csu_init (i.e. the init argument passed to __libc_start_main)
1. Call function pointers in .preinit_array section
2. Execute the code in .init section, which is usually _init function. What _init function does is compiler-specific. For GCC, _init executes user functions marked as __attribute__ ((constructor)) (in __do_global_dtors_aux)
3. Call function pointers in .init_array section
Set up data structures needed for thread unwinding/cancellation
Call main of user’s program.
Call exit

So if the last line of user program’s main is return XX, then the XX will be passed to exit at Step #11 above. If the last line is not return XX or is simply return, then the value passed to exit would be undefined.

Of course, if the user program calls exit or abort, then exit will gets called.

Here is the call graph, which is worth a thousand words

and see here on how it is generated.

If one tries to build a program which does not contain main, then one should see the following error:

/usr/lib/crt1.o: In function `_start': (.text+0x20): undefined reference to `main'
collect2: ld returned 1 exit status

As mentioned earlier, crt1.o (part of Glibc) contains the function _start, which will call __libc_start_main and pass main (a function pointer) as one of the arguments. If one uses

nm -u /usr/lib/crt1.o

then it will show main is a undefined symbol in crt1.o. Now let’s disassemble crt1.o:

$ objdump -M intel -dj .text /usr/lib/crt1.o

crt1.o:     file format elf64-x86-64

Disassembly of section .text:

0000000000000000 <_start>:
   0:   31 ed                   xor    ebp,ebp
   2:   49 89 d1                mov    r9,rdx
   5:   5e                      pop    rsi
   6:   48 89 e2                mov    rdx,rsp
   9:   48 83 e4 f0             and    rsp,0xfffffffffffffff0
   d:   50                      push   rax
   e:   54                      push   rsp
   f:   49 c7 c0 00 00 00 00    mov    r8,0x0
  16:   48 c7 c1 00 00 00 00    mov    rcx,0x0
  1d:   48 c7 c7 00 00 00 00    mov    rdi,0x0
  24:   e8 00 00 00 00          call   29 <_start+0x29>
  29:   f4                      hlt
  ...

Above shows .text+0x20 refers to the 4 bytes of an mov instruction. This means during the linking, the address of main should be resolved and then inserted at the right memory location: .text+0x20. Now let’s cross reference the relocation table:

$ readelf -p /usr/lib/crt1.o

Relocation section '.rela.text' at offset 0x410 contains 4 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000000012  00090000000b R_X86_64_32S      0000000000000000 __libc_csu_fini + 0
000000000019  000b0000000b R_X86_64_32S      0000000000000000 __libc_csu_init + 0
000000000020  000c0000000b R_X86_64_32S      0000000000000000 main + 0
000000000025  000f00000002 R_X86_64_PC32     0000000000000000 __libc_start_main - 4

Above shows where 0x20 comes from.

How to find the address of `main` of an executable binary ?

When an ELF executable binary is stripped off symbolic information, it is not clear where the main is located.

From above analysis, it’s possible to find out the address of main (which is NOT the “Entry point address” seen from the output of readelf -h a.out | grep Entry command. “Entry point address” is the address of _start)

Since the address of main is the first argument to the call to __libc_start_main, we can extract the value of the first argument as follows.

On 64-bit x86, the calling convention requires that the first argument goes to RDI register, so the address can be extracted by

objdump -j .text -d a.out | grep -B5 'call.*__libc_start_main' | awk '/mov.*%rdi/ { print $NF }'

On 32-bit x86, the C calling convention (“cdecl”) is that the first argument is the last item to be pushed onto the stack before the call, so the address can be extracted by

objdump -j .text -d a.out | grep -B2 'call.*__libc_start_main' | awk '/push.*0x/ { print $NF }'

PIC, TLS, and AMD64 code models

Relocation is the process of connecting symbolic references with symbolic definitions. The runtime relocation is done by ld.so, as in elf_machine_rela function in Glibc’s source file sysdeps/x86_64/dl-machine.h. The link-time relocation is done by the link-editor ld, which uses the relocation table in the object file (.rela.text section). Each symbolic reference has an entry in the relocation table, and each entry contains three fields: offset, info (relocation type, symbol table index), and addend. The relocation types are:

Relocation type	Meaning	Used when
`R_X86_64_16`	Direct 16 bit zero extended
`R_X86_64_32`	Direct 32 bit zero extended
`R_X86_64_32S`	Direct 32 bit sign extended
`R_X86_64_64`	Direct 64 bit	Large code model
`R_X86_64_8`	Direct 8 bit sign extended
`R_X86_64_COPY`	Copy symbol at runtime
`R_X86_64_DTPMOD64`	ID of module containing symbol	TLS
`R_X86_64_DTPOFF32`	Offset in TLS block	TLS
`R_X86_64_DTPOFF64`	Offset in module’s TLS block	TLS
`R_X86_64_GLOB_DAT`	`.got` section, which contains addresses to the actual functions in DLL
`R_X86_64_GOT32`	32 bit GOT entry
`R_X86_64_GOT64`	64-bit GOT entry offset	PIC & Large code model
`R_X86_64_GOTOFF64`	64-bit GOT offset	PIC & Large code model
`R_X86_64_GOTPC32`	32-bit PC relative offset to GOT
`R_X86_64_GOTPC32_TLSDESC`	32-bit PC relative to TLS descriptor in GOT	TLS
`R_X86_64_GOTPC64`	64-bit PC relative offset to GOT	PIC & Large code model
`R_X86_64_GOTPCREL`	32 bit signed PC relative offset to GOT	PIC
`R_X86_64_GOTPCREL64`	64-bit PC relative offset to GOT entry	PIC & Large code model
`R_X86_64_GOTPLT64`	Like GOT64, indicates that PLT entry needed	PIC & Large code model
`R_X86_64_GOTTPOFF`	32 bit signed PC relative offset to GOT entry for IE symbol	TLS
`R_X86_64_JUMP_SLOT`	`.got.plt` section, which contains addresses to the actual functions in DLL	DLL
`R_X86_64_PC16`	16 bit sign extended PC relative
`R_X86_64_PC32`	PC relative 32 bit signed
`R_X86_64_PC64`	64-bit PC relative	Large code model
`R_X86_64_PC8`	8 bit sign extended PC relative
`R_X86_64_PLT32`	32 bit PLT address
`R_X86_64_PLTOFF64`	64-bit GOT relative offset to PLT entry	PIC & Large code model
`R_X86_64_RELATIVE`	Adjust by program base
`R_X86_64_SIZE32`
`R_X86_64_SIZE64`
`R_X86_64_TLSDESC`	2 by 64-bit TLS descriptor	TLS
`R_X86_64_TLSDESC_CALL`	Relaxable call through TLS descriptor	TLS
`R_X86_64_TLSGD`	32 bit signed PC relative offset to two GOT entries for GD symbol	TLS & PIC
`R_X86_64_TLSLD`	32 bit signed PC relative offset to two GOT entries for LD symbol	TLS
`R_X86_64_TPOFF32`	Offset in initial TLS block	TLS
`R_X86_64_TPOFF64`	Offset in initial TLS block	TLS & Large code model

According to Chapter 3.5 of AMD64 System V Application Binary Interface, there are four code models and they differ in addressing modes (absolute versus relative):

Small: All compile- and link-time addresses and symbols are assumed to fit in 32-bit immediate operands. This model restricts code and global data to the low 2 GB of the address space (to be exact, between 0x0 and 0x7eff ffff, which is 2031 MB)

The compiler can encode symbolic references
- In sign-extended immediate operands for offsets in the range of 0x8000 0000 (-231) to 0x100 0000 (224)
- In zero-extended immediate operands for offsets in the range of 0x0 to 0x7f00 0000 (231 - 224)
- In RIP relative addressing mode for offsets in the range 0xff00 0000 (-224 = -16 MB) to 0x100 0000 (224 = 16 MB)
This mode is the default mode for most compilers.
Large: No restrictions are placed on the size or placement of code and data. The max virtual memory space is 48 bits (256 TB).
Medium: Like the Small code model, except the data sections are split into two parts, e.g. instead of having just .data, .rodata, .bss sections, there would also be .ldata, .lrodata, .lbss sections. The smaller .data, etc are still the same as in the Small code model, and the larger .ldata, etc are as in the Large code model.
Kernel: Like the Small code model, but the 2 GB address space spans from 0xffff ffff 8000 0000 (264-231) to 0xffff ffff ff00 0000 (264-224) Because of this, the offsets which can be encoded using sign-extended and zero-extended immediate operands change.

Now consider the following C code

extern int esrc[100];
       int gsrc[100];
static int ssrc[100];

void foo() {
  int k;
  k = esrc[5];
  k = gsrc[5];
  k = ssrc[5];
}

Small code model, no PIC (i.e. compiled with just gcc -c):

k = esrc[5];     mov EAX, DWORD PTR[RIP+0x0]
                 mov DWORD PTR[RBP-0x4], EAX
k = gsrc[5];     mov EAX, DWORD PTR[RIP+0x0]
                 mov DWORD PTR[RBP-0x4], EAX
k = ssrc[5];     mov EAX, DWORD PTR[RIP+0x0]
                 mov DWORD PTR[RBP-0x4], EAX

and the relocation table is (use readelf -r foo.o command)

type           Sym. Name + Addend
R_X86_64_PC32  esrc + 10
R_X86_64_PC32  gsrc + 10
R_X86_64_PC32  .bss + 10

All of the 0x0’s in the generated assembly will be filled at link-time with their relative offsets in respective sections, as indicated by the relocation table.

Large code model, no PIC (i.e. compiled with gcc -c -mcmodel=large)

k = esrc[5];     mov RAX, 0x0
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX
k = gsrc[5];     mov RAX, 0x0
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX
k = ssrc[5];     mov RAX, 0x0
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX

and the relocation table is:

type         Sym. Name + Addend
R_X86_64_64  esrc + 0
R_X86_64_64  gsrc + 0
R_X86_64_64  .bss + 0

All of the 0x0’s in the generated assembly will be filled at link-time with their (64-bit) absolute addresses.

Small code model, PIC (i.e. compiled with gcc -c -fPIC. Note that adding -shared or not will not affect the generated code)

k = esrc[5];     mov RAX, QWORD PTR[RIP+0x0]
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX
k = gsrc[5];     mov RAX, QWORD PTR[RIP+0x0]
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX
k = ssrc[5];     mov EAX, DWORD PTR[RIP+0x0]
                 mov DWORD PTR[RBP-0x4], EAX

and the relocation table is:

type              Sym. Name + Addend
R_X86_64_GOTPCREL esrc - 4
R_X86_64_GOTPCREL gsrc - 4
R_X86_64_PC32     .bss + 10

The first two 0x0’s in the generated assembly will be filled with the relative offset of _GLOBAL_OFFSET_TABLE_ (i.e. the .got.plt section)

Large code model, PIC (i.e. compiled with gcc -c -fPIC -mcmodel=large)

                 lea RBX, [RIP-0x7]
                 mov R11, 0x0
                 add RBX, R11
k = esrc[5];     mov RAX, 0x0
                 mov RAX, QWORD PTR[RBX+RAX*1]
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX
k = gsrc[5];     mov RAX, 0x0
                 mov RAX, QWORD PTR[RBX+RAX*1]
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX
k = ssrc[5];     mov RAX, 0x0
                 mov RAX, QWORD PTR[RBX+RAX*1]
                 mov EAX, DWORD PTR[RAX+0x10]
                 mov DWORD PTR[RBP-0x4], EAX

The first 0x0 is in the generated assembly will be filled with the absolute address of _GLOBAL_OFFSET_TABLE_

`_GLOBAL_OFFSET_TABLE_`, `.got.plt` section, and `DYNAMIC` segment

Earlier we see that the _GLOBAL_OFFSET_TABLE_ is located in .got.plt section:

(gdb) b *0x4003d0
(gdb) run
(gdb) x/6a 0x600890
0x600890: 0x6006e8 <_DYNAMIC>     0x32696159a8
0x6008a0: 0x326950aa20 <_dl_runtime_resolve>     0x4003d6 <printf@plt+6>
0x6008b0: 0x326971c3f0 <__libc_start_main>       0x0

According to Chapter 5.2 of AMD64 System V Application Binary Interface, the first 3 entries of this table are reserved for special purposes. The first entry is set up during compilation by the link editor ld. The second and third entries are set up during runtime by the runtime linker ld.so (see function _dl_relocate_object in Glibc source file elf/dl-reloc.c and in particular, notice the ELF_DYNAMIC_RELOCATE macro, which calls function elf_machine_runtime_setup in sysdeps/x86_64/dl-machine.h)

The first entry _DYNAMIC has value 6006e8, and this is exactly the starting address of .dynamic section (or DYNAMIC segment, in ELF’s “execution view”.) The runtime linker ld.so uses this section to find the all necessary information needed for runtime relocation and dynamic linking.

To see DYNAMIC segment’s content, use readelf -d a.out command, or objdump -x a.out, or just use x/50a 0x6006e8 in gdb. The readelf -d a.out command will show something like this:

Dynamic section at offset 0x6e8 contains 21 entries:
  Tag        Type                 Name/Value
 0x0000000000000001 (NEEDED)     Shared library: [libc.so.6]  <-- dependent dynamic library name
 0x000000000000000c (INIT)       0x4003a8     <-- address of .init section
 0x000000000000000d (FINI)       0x400618     <-- address of .fini section
 0x0000000000000004 (HASH)       0x400240     <-- address of .hash section
 0x000000006ffffef5 (GNU_HASH)   0x400268     <-- address of .gnu.hash section
 0x0000000000000005 (STRTAB)     0x4002e8     <-- address of .strtab section
 0x0000000000000006 (SYMTAB)     0x400288     <-- address of .symtab section
 0x000000000000000a (STRSZ)      63 (bytes)   <-- size of .strtab section
 0x000000000000000b (SYMENT)     24 (bytes)   <-- size of an entry in .symtab section
 0x0000000000000015 (DEBUG)      0x0          <-- see below
 0x0000000000000003 (PLTGOT)     0x600860     <-- address of .got.plt section
 0x0000000000000002 (PLTRELSZ)   48 (bytes)   <-- total size of .rela.plt section
 0x0000000000000014 (PLTREL)     RELA         <-- RELA or REL ?
 0x0000000000000017 (JMPREL)     0x400368     <-- address of .rela.plt section
 0x0000000000000007 (RELA)       0x400350     <-- address of .rela.dyn section
 0x0000000000000008 (RELASZ)     24 (bytes)   <-- total size of .rela.dyn section
 0x0000000000000009 (RELAENT)    24 (bytes)   <-- size of an entry in .rela.dyn section
 0x000000006ffffffe (VERNEED)    0x400330     <-- address of .gnu.version_r section
 0x000000006fffffff (VERNEEDNUM) 1            <-- number of needed versions
 0x000000006ffffff0 (VERSYM)     0x400328     <-- address of .gnu.version section
 0x0000000000000000 (NULL)       0x0          <-- marks the end of .dynamic section

Each entry in DYNAMIC segment is a struct of only two members: “tag” and “value”. The NEEDED, INIT … above are “tags” (see /usr/include/elf.h)

Other tags of interest are:

BIND_NOW           The same as BIND_NOW in FLAGS. This has been superseded by
                   BIND_NOW in FLAGS

CHECKSUM           The checksum value used by prelink.

DEBUG              At runtime ld.so will fill its value with the runtime
                   address of r_debug structure (see elf/rtld.c)
                   and this info is used by GDB (see elf_locate_base function
                   in GDB's source tree).

FINI               Address of .fini section
FINI_ARRAY         Address of .fini_array section
FINI_ARRAYSZ       Size of .fini_array section

FLAGS              Additional flags, such as BIND_NOW, STATIC_TLS, TEXTREL..

FLAGS_1            Additional flags used by Solaris, such as NOW (the same as BIND_NOW), INTERPOSE..

GNU_PRELINKED      The timestamp string when the binary object is last prelinked.

INIT               Address of .init section
INIT_ARRAY         Address of .init_array section
INIT_ARRAYSZ       Size of .init_array section

INTERP             Address of .interp section

PREINIT_ARRAY      Address of .preinit_array section
PREINIT_ARRAYSZ    Size of .preinit_array section

RELACOUNT          Number of R_X86_64_RELATIVE entries in RELA segment (.rela.dyn
                   section)

RPATH              Dynamic library search path, which has higher precendence than
                   LD_LIBRARY_PATH. RPATH is ignored if RUNPATH is present.

                   Use of RPATH is deprecated.

                   When one uses "gcc -Wl,-rpath=... " to build binaries, the info
                   is stored here.

RUNPATH            Dynamic library search path, which has lower precendence than
                   LD_LIBRARY_PATH.

                   When one uses "gcc -Wl,-rpath=...,--enable-new-dtags"
                   to build binaries, the info is stored here.
                   (See here for details.)

                   One can use chrpath
                   tool to manipulate RPATH and RUNPATH settings.


SONAME             Shared object (i.e. dynamic library) name. When one uses
                   "gcc -Wl,-soname=... " to build binaries, the info is
                   stored here.

TEXTREL            Relocation might modify .text section.

VERDEF             Address of .gnu.version_d section
VERDEFNUM          Number of version definitions.

Runtime Relocation

After exploring DYNAMIC segment, we can explain how ld.so performs runtime relocation.

First, before ld.so loads all dependent libraries of a dynamic executable, it needs to run its own relocation! Even if ld.so is a statically-linked binary, it also has a DYNAMIC segment and thus PLTREL (.rela.dyn section) and JMPREL (.rela.plt section) tags:

$ readelf -a `readelf -p .interp /bin/sh | awk '/ld/ {print $3}'`

 ....

Dynamic section at offset 0x14e18 contains 22 entries:
  Tag        Type                         Name/Value
 0x000000000000000e (SONAME)             Library soname: [ld-linux-x86-64.so.2]
 0x0000000000000004 (HASH)               0x3269500190
 0x0000000000000005 (STRTAB)             0x3269500578
 0x0000000000000006 (SYMTAB)             0x3269500260
 0x000000000000000a (STRSZ)              388 (bytes)
 0x000000000000000b (SYMENT)             24 (bytes)
 0x0000000000000003 (PLTGOT)             0x3269614f98
 0x0000000000000002 (PLTRELSZ)           120 (bytes)
 0x0000000000000014 (PLTREL)             RELA
 0x0000000000000017 (JMPREL)             0x32695009a0
 0x0000000000000007 (RELA)               0x32695007c0
 0x0000000000000008 (RELASZ)             480 (bytes)
 0x0000000000000009 (RELAENT)            24 (bytes)
 0x000000006ffffffc (VERDEF)             0x3269500740
 0x000000006ffffffd (VERDEFNUM)          4
 0x0000000000000018 (BIND_NOW)
 0x000000006ffffffb (FLAGS_1)            Flags: NOW
 0x000000006ffffff0 (VERSYM)             0x32695006fc
 0x000000006ffffff9 (RELACOUNT)          19
 0x000000006ffffdf8 (CHECKSUM)           0x4c4e099e
 0x000000006ffffdf5 (GNU_PRELINKED)      2010-08-26T08:13:28
 0x0000000000000000 (NULL)               0x0

Relocation section '.rela.dyn' at offset 0x7c0 contains 20 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
003269614cf0  000000000008 R_X86_64_RELATIVE                    000000326950dd80
  ....
003269615820  000000000008 R_X86_64_RELATIVE                    0000003269501140
003269614fe0  001e00000006 R_X86_64_GLOB_DAT 0000003269615980 _r_debug + 0

Relocation section '.rela.plt' at offset 0x9a0 contains 5 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
003269614fb0  000b00000007 R_X86_64_JUMP_SLO 000000326950f1b0 __libc_memalign + 0
003269614fb8  000c00000007 R_X86_64_JUMP_SLO 000000326950f2b0 malloc + 0
003269614fc0  001200000007 R_X86_64_JUMP_SLO 000000326950f2c0 calloc + 0
003269614fc8  001800000007 R_X86_64_JUMP_SLO 000000326950f340 realloc + 0
003269614fd0  002000000007 R_X86_64_JUMP_SLO 000000326950f300 free + 0

Note that the ld.so is prelinked. On Fedora and Red Hat Enterprise Linux (RHEL) systems, prelink is run every two weeks. To see if your Linux has similar setup, check /etc/sysconfig/prelink and /etc/prelink.conf

What does this prelink do? It changes the base address of a dynamic library to the actual address in the user program’s address space when it is loaded into memory. Of course, ld.so recognizes GNU_PRELINKED tag and will load a dynamic library to its this base address (recall the first argument of mmap is the preferred address; of course, this is subject to the operating system.)

Normally, a dynamic library is built as position independent code, i.e. the -fPIC compiler command-line option, and thus the base address is 0. For example, a normal libc.so has ELF program header as follows (readelf -l command):

Program Headers:
  Type  Offset             VirtAddr           PhysAddr
        FileSiz            MemSiz              Flags  Align
  LOAD  0x0000000000000000 0x0000000000000000 0x0000000000000000
        0x0000000000179058 0x0000000000179058  R E    200000
  LOAD  0x0000000000179730 0x0000000000379730 0x0000000000379730
        0x0000000000004668 0x00000000000090f8  RW     200000
   ....

And when calling mmap with address 0 (i.e. NULL) the operating system can choose any address it feels appropriate.

A prelinked one, on the other hand, has its ELF program header as follows:

Program Headers:
  Type Offset             VirtAddr           PhysAddr
       FileSiz            MemSiz              Flags  Align
  LOAD 0x0000000000000000 0x0000003433e00000 0x0000003433e00000
       0x000000000001bb80 0x000000000001bb80  R E    200000
  LOAD 0x000000000001bb90 0x000000343401bb90 0x000000343401bb90
       0x0000000000000f58 0x00000000000010f8  RW     200000

What is the advantage of prelinking? ld.so will not process R_X86_64_RELATIVE relocation types since they are already in the “right” place in user program’s address space. The extra benefit of this is the memory regions which ld.so would have written to (if R_X86_64_RELATIVE needs processing) will not incur any Copy-On-Writes and thus can be made Read-Only.

According to this post, for GUI programs, which tend to link against dozens of dynamic libraries and use lengthy C++ demangled names, the speed up can be an order of magnitude.

How to disable prelinking at runtime? Run the user program with LD_USE_LOAD_BIAS environmental variable set to 0.

How does ld.so process its own relocation?

The relocation is done by _dl_relocate_object function in Glibc’s elf/dl-reloc.c, which will call elf_machine_rela function in sysdeps/x86_64/dl-machine.h to do the majority of work.

First to be processed is the .rela.dyn relocation table, which contains a bunch of R_X86_64_RELATIVE types and one R_X86_64_GLOB_DAT type (the variable _r_debug)

If prelink is used, i.e. ld.so is indeed loaded to the desired address, then R_X86_64_RELATIVE relocation types will be ignored. If not, then the address calculation for R_X86_64_RELATIVE types is

Base Address + Value Stored at [Base Address + Offset]

For example, in ld.so’s case, its base address is 2a95556000 (can be obtained from pmap command; inside ld.so, it calls elf_machine_load_address function to get this value)

0000400000    4K r-x--  /tmp/a.out
0000500000    4K rw---  /tmp/a.out
2a95556000   92K r-x--  /lib64/ld.so
2a9556d000    8K rw---  [ anon ]
2a95599000    4K rw---  [ anon ]
2a9566c000    4K r----  /lib64/ld.so
2a9566d000    4K rw---  /lib64/ld.so
3269700000 1216K r-x--  /lib64/libc-2.3.4.so
...

And ld.so’s .rela.dyn relocation table is (no prelinked! If ld.so is prelinked, the offset will be in a much higher address)

Relocation section '.rela.dyn' at offset 0x7c0 contains 20 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000116d50  000000000008 R_X86_64_RELATIVE                    000000000000e250
...

so the relocation for 000000116d50 is processed as

0x2a95556000 + *(0x116d50+0x2a95556000)

and this new value is stored at 0x2a9566cd50 (=0x116d50+0x2a95556000)

As R_X86_64_RELATIVE types do not require symbol lookups, they are handled in a tight loop in elf_machine_rela_relative function in sysdeps/x86_64/dl-machine.h

Any relocation types other than R_X86_64_RELATIVE need to go through symbol resolution first.

So what about R_X86_64_GLOB_DAT relocation type in ld.so ? First, RESOLVE_MAP (a macro defined within elf/dl-reloc.c) is called (with r_type = R_X86_64_GLOB_DAT) to find out which ELF binary (could be the user’s program or its dependent dynamic libraries) contains this symbol. Then R_X86_64_GLOB_DAT relocation type is calculated as

Base Address + Symbol Value + Addend

where Base Address is the base address of ELF binary which contains the symbol, and Symbol Value is the symbol value from the symbol table of ELF binary which contains the symbol.

So for ld.so,

Relocation section '.rela.dyn' at offset 0x7c0 contains 20 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
  ....
000000116fe0  001e00000006 R_X86_64_GLOB_DAT 00000000001179c0 _r_debug + 0

The relocation for 000000116fe0 is processed as

0x2a95556000 + 0x1179c0 + 0

because ld.so determines _r_debug can be found from itself. The calculated value is stored at 0x2a9566cfe0 (=0x116fe0+0x2a95556000).

The next to be processed by ld.so is its own .rela.plt relocation table, which contains a bunch of R_X86_64_JUMP_SLOT types. This reloction type is handled exactly the same way as R_X86_64_GLOB_DAT.

After ld.so finishes its own relocation, it loads user program’s dependent libraries and process their relocations one by one. First, ld.so handles libc.so’s relocation. libc.so has two relocation types we have not covered so far: R_X86_64_64 and R_X86_64_TPOFF64.

R_X86_64_64 relocation type is processed by first looking up the symbol’s runtime absolute address, and then calculating

Absolute Address + Addend

And the R_X86_64_TPOFF64 relocation type is calculated as

Symbol Value + Addend - TLS Offset

which usually results in a negative value.

`R_X86_64_COPY` relocation type

R_X86_64_COPY relocation type is used when a dynamic binary refers to an initialized global variable (not a function!) defined in a dynamic link library. Unlike functions, for variables, there is no lazy binding, and the trampoline trick used in .plt section does not work. Instead, the global variable will actually be allocated in dynamic binary’s .bss section.

To see how R_X86_64_COPY relocation type works, consider the following two code:

foo.c

    int foo=4;

    void foo_access() {
      foo=5;
    }

bar.c

    #include <stdio.h>
    extern int foo;

    int main() {
       printf("foo=%d\n",foo);
    }

Now compile them as follows:

$ gcc -shared -fPIC -Wl,-soname=libfoo.so foo.c -o /tmp/libfoo.so
$ gcc bar.c -o bar -L/tmp -lfoo

And run them as

$ LD_PRELOAD=/tmp/libfoo.so ./bar

Before explaining what happened during runtime, we need to examine the binaries first.

The foo_access in libfoo.so is like this:

69c <foo_access>:
 69c:  push   rbp
 69d:  mov    rbp,rsp
 6a0:  mov    rax,QWORD PTR [rip+0x100269]  # 100910 <_DYNAMIC+0x198>
 6a7:  mov    DWORD PTR [rax],0x5
 6ad:  leave
 6ae:  ret

So for libfoo.so, the address of variable foo is in its .got section, not .data section:

$ readelf -a /tmp/libfoo.so

Section Headers:
  [Nr] Name              Type             Address           Offset
       Size              EntSize          Flags  Link  Info  Align
...
  [18] .got              PROGBITS         0000000000100908  00000908
       0000000000000020  0000000000000008  WA       0     0     8
  [19] .got.plt          PROGBITS         0000000000100928  00000928
       0000000000000020  0000000000000008  WA       0     0     8
...
  [20] .data             PROGBITS         0000000000100948  00000948
       0000000000000014  0000000000000000  WA       0     0     8
...

Relocation section '.rela.dyn' at offset 0x520 contains 6 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000100948  000000000008 R_X86_64_RELATIVE                    0000000000100948
000000100950  000000000008 R_X86_64_RELATIVE                    0000000000100768
000000100908  000f00000006 R_X86_64_GLOB_DAT 0000000000000000 __cxa_finalize + 0
000000100910  001100000006 R_X86_64_GLOB_DAT 0000000000100958 foo + 0
....

But what about the address 0x100958 ? This address is in libfoo.so’s .data section! Well, 0x100958 has the initial value of foo (in our example, 4) At runtime, ld.so will copy this value to bar’s .bss section:

$ objdump -sj .data libfoo.so

libfoo.so:     file format elf64-x86-64

Contents of section .data:
 100948 48091000 00000000 68071000 00000000  H.......h.......
 100958 04000000                             ....

Next, disassemble the main function of bar:

4005f8 <main>:
  4005f8:  push   rbp
  4005f9:  mov    rbp,rsp
  4005fc:  mov    esi,DWORD PTR [rip+0x1003de] # 5009e0 <__bss_start>
  400602:  mov    edi,0x40070c
  400607:  mov    eax,0x0
  40060c:  call   400528 <printf@plt>
  400611:  leave
  400612:  ret

So the variable foo is indeed located in bar’s .bss section. Let’s double check with nm:

$ nm -n bar | grep 5009e0

00000000005009e0 A __bss_start
00000000005009e0 A _edata
00000000005009e0 B foo

(Symbols such as __bss_start and _edata are defined by the default ld script; one can search them in the output of ld -verbose command.)

The dynamic relocation table of bar is:

Relocation section '.rela.dyn' at offset 0x490 contains 2 entries:
  Offset          Info           Type           Sym. Value    Sym. Name + Addend
000000500998  000c00000006 R_X86_64_GLOB_DAT 0000000000000000 __gmon_start__ + 0
0000005009e0  000700000005 R_X86_64_COPY     00000000005009e0 foo + 0

Now what happens during runtime is this: After ld.so loads all dependent dynamic libraries, it starts processing their relocations. When it sees foo of libfoo.so, it calls RESOLVE_MAP with r_type = R_X86_64_GLOB_DAT to get the Base Address, which is 0, and Symbol Value, which is 5009e0. Next it sees foo of libfoo.so has R_X86_64_GLOB_DAT relocation type, so it calculates the new address as 5009e0 = 0 + 5009e0 + 0 (addend) and stores the result somewhere inside .got section.

After ld.so has processed relocations of all dynamic libraries, it starts processing the relocation table of bar. When it sees foo of bar, it calls RESOLVE_MAP again, but with r_type = R_X86_64_COPY. This time, the address returned is the runtime address of foo in libfoo.so’s .data section. As mentioned earlier, this address holds the initial value of foo. Next it sees foo of bar has R_X86_64_COPY relocation type, so it uses memcpy to copy data to 5009e0 (see the Sym. Value of .rela.dyn section of bar above) from the runtime address of foo in libfoo.so’s .data section (see Glibc source file sysdeps/x86_64/dl-machine.h)

The above example also illustrates the difference between .got section and .got.plt section. For the runtime linker ld.so, all it knows is entries in PLTREL segment, i.e. .rela.dyn section, (which corresponds to .got section) must be resolved/relocated immediately, while entries in JMPREL segment, i.e. .rela.plt section, (which corresponds to .got.plt section) can use lazy binding. For x86_64 architecture, the relocation is actually not needed for R_X86_64_JUMP_SLOT relocation types (albeit the symbol resolution is still needed)

PIC or no PIC

When building a dynamic library, we are told to always compile the code with -fPIC option.

What’s the difference then ?

Consider the following simple code:

#include <stdio.h>
int bar;

void foo() {
  printf("%d\n",bar);
}

Compile the above code in 32-bit mode with and without -fPIC:

$ gcc -shared -m32 foo.c -o nopic.so
$ gcc -shared -m32 -fPIC foo.c -o pic.so

(If you try to compile the above in 64-bit mode, GCC will stop and insist you should compile with -fPIC option, i.e. you are going to see error message such as relocation R_X86_64_PC32 against symbol `XXXYYY' can not be used when making a shared object; recompile with -fPIC) The sections and relocation tables of nopic.so and pic.so are shown at left and right hand side, respectively:

Section Headers:                                                 Section Headers:
[Nr] Name           Type     Addr                                [Nr] Name              Type     Addr
[ 0]                NULL     0000                                [ 0]                   NULL     0000
  ...                                                              ...
[ 8] .init          PROGBITS 02f8                                [ 8] .init             PROGBITS 02f0
[ 9] .plt           PROGBITS 0310                                [ 9] .plt              PROGBITS 0308
[10] .text          PROGBITS 0340                                [10] .text             PROGBITS 0350
[11] .fini          PROGBITS 0488                                [11] .fini             PROGBITS 04a8
[12] .rodata        PROGBITS 04a4                                [12] .rodata           PROGBITS 04c4
  ...                                                              ...
[17] .dynamic       DYNAMIC  14c0                                [17] .dynamic          DYNAMIC  14e0
[18] .got           PROGBITS 1590                                [18] .got              PROGBITS 15a8
[19] .got.plt       PROGBITS 159c                                [19] .got.plt          PROGBITS 15b8
[20] .data          PROGBITS 15b0                                [20] .data             PROGBITS 15d0

  ...                                                              ...

Relocation section '.rel.dyn' at offset 0x2b0                    Relocation section '.rel.dyn' at offset 0x2b0
contains 7 entries:                                              contains 5 entries:
 Offset     Info    Type            Sym.Value  Sym. Name         Offset     Info    Type            Sym.Value  Sym. Name
00000439  00000008 R_386_RELATIVE                                000015d0  00000008 R_386_RELATIVE
000015b0  00000008 R_386_RELATIVE                                000015a8  00000106 R_386_GLOB_DAT    000015dc   bar
00000434  00000101 R_386_32          000015bc   bar                     ...
00000445  00000602 R_386_PC32        00000000   printf
 ...

Relocation section '.rel.plt' at offset 0x2e8:                   Relocation section '.rel.plt' at offset 0x2d8
contains 2 entries:                                              contains 3 entries:
 Offset     Info    Type            Sym.Value  Sym. Name         Offset     Info    Type            Sym.Value  Sym. Name
000015a8  00000207 R_386_JUMP_SLOT   00000000   __gmon_start__   000015c4  00000207 R_386_JUMP_SLOT   00000000   __gmon_start__
000015ac  00000a07 R_386_JUMP_SLOT   00000000   __cxa_finalize   000015c8  00000607 R_386_JUMP_SLOT   00000000   printf
                                                                      ...

When we compile with -fPIC we can see the variable bar has the right relocation type (R_386_GLOB_DAT) and the relocation takes place in the right section (.got) The same for printf.

Without -fPIC, the relocations of the format string “\n”, bar and printf all take place inside the .text section! But we know .text section is in a Read-Only LOAD segment, so what ld.so would do ?

As expected, ld.so will make .text section writeable, patch the bytes, and make it Read-Only again. Since the relocation of both bar and printf are in .rel.dyn, their relocations are performed immediately (no lazy binding), so this approach is feasible.

So how does ld.so handle R_386_RELATIVE, R_386_32 and R_386_PC32 relocation types ?

Let’s look at the disassembly:

0000042c <foo>:
 42c: 55                push   ebp
 42d: 89 e5             mov    ebp,esp
 42f: 83 ec 18          sub    esp,0x18
 432: 8b 15 00 00 00 00 mov    edx,DWORD PTR ds:0x0   <-- reference to bar
 438: b8 a4 04 00 00    mov    eax,0x4a4              <-- reference to "%d\n" format string in .rodata
 43d: 89 54 24 04       mov    DWORD PTR [esp+0x4],edx
 441: 89 04 24          mov    DWORD PTR [esp],eax
 444: e8 fc ff ff ff    call   445 <foo+0x19>  <-- reference to printf
 449: c9                leave
 44a: c3                ret

How would the 4 bytes starting at 445 (R_386_PC32 type) be patched ? Suppose at runtime, our nopic.so is loaded into memory with base address 8000, and the 4 bytes to be patched are now at 8000 + 445 = 8445. Furthermore, suppose ld.so has determined the entry address of printf to be 10000, then ld.so calculates the relative offset as follows:

10000 - 8445 + fffffffc = 7bb7

(fffffffc is -4) so ld.so replaces fc ff ff ff with b7 7b 00 00

To patch the 4 bytes starting at 434 (R_386_32 type) is simpler. ld.so will simply overwrite the 4 bytes with the runtime absolute address of bar.

To patch the 4 bytes starting at 439 (R_386_RELATIVE type) ld.so calculates the address as

10000 + 4a4 = 104a4

so ld.so replaces a4 04 00 00 with a4 04 01 00

Finally, what about the R_386_RELATIVE relocation at 15b0 ? 15b0 is the starting address of .data section, and the first 4 bytes of .data section stores its own address, 15b0. So it has to be relocated and patched as 115b0.

In conclusion, R_386_RELATIVE means “32-bit relative to base address”, R_386_PC32 means the “32-bit IP-relative offset” and R_386_32 means the “32-bit absolute.”

Troubleshooting `ld.so`

What is “error while loading shared libraries: requires glibc 2.5 or later dynamic linker” ?

The cause of this error is the dynamic binary (or one of its dependent shared libraries) you want to run only has .gnu.hash section, but the ld.so on the target machine is too old to recognize .gnu.hash; it only recognizes the old-school .hash section.

This usually happens when the dynamic binary in question is built using newer version of GCC. The solution is to recompile the code with either -static compiler command-line option (to create a static binary), or the following option:

-Wl,--hash-style=both

This tells the link editor ld to create both .gnu.hash and .hash sections.

According to ld documentation here, the old-school .hash section is the default, but the compiler can override it. For example, the GCC (which is version 4.1.2) on RHEL (Red Hat Enterprise Linux) Server release 5.5 has this line:

$ gcc -dumpspecs
....
*link:
%{!static:--eh-frame-hdr} %{!m32:-m elf_x86_64} %{m32:-m elf_i386} --hash-style=gnu   %{shared:-shared}   ....
...

For more information, see here.

What is “Floating point exception” ?

The cause of this error is the same as the previous question. On certain systems, e.g. RHEL, the old version ld.so is backported to emit “error while loading shared libraries: requires glibc 2.5 or later dynamic linker”, but this is not always the case, and you will see this error instead.

What is “…/libc.so.6: version `GLIBC_2.4’ not found “ ?

As the error message says, some of the symbols need Glibc version 2.4 or higher. This can also be seen by

$ objdump -x foo | grep 'Version References' -A10

Version References:
  required from libc.so.6:
    0x0d696914 0x00 03 GLIBC_2.4
    0x09691a75 0x00 02 GLIBC_2.2.5

...

The fix is to recompile the code with -static compiler command-line option.

What is “FATAL: kernel too old” ?

Even if you recompile the code with -static compiler command-line option to avoid any dependency on the dynamic Glibc library, you could still encounter the error in question, and your code will exit with Segmentation Fault error.

This kernel version check is done by DL_SYSDEP_OSCHECK macro in Glibc’s sysdeps/unix/sysv/linux/dl-osinfo.h It calls _dl_discover_osversion to get current kernel’s version.

To wit, run your code (suppose it is not stripped) inside gdb,

(gdb) run
Starting program: foo
FATAL: kernel too old

Program received signal SIGSEGV, Segmentation fault.
0x00000000004324a9 in ptmalloc_init ()
(gdb) call _dl_discover_osversion()
$1 = 132617
(gdb) p/x $1
$2 = 0x20609
(gdb)

Here 0x20609 means the current kernel version is 2.6.9.

The fix (or hack) is to add the following function in your code:

int _dl_discover_osversion() { return 0xffffff; }

and compile your code with -static compiler command-line option.

Exploring Glibc’s `pthread_t`

When one creates a thread using the Pthread API, one will get a pthread_t object as a handle. In Glibc, pthread_t is actually a pointer pointing to a pthread struct, which is opaque. Its definition can be found in Glibc’s source tree at nptl/descr.h. The first member of pthread struct is yet another struct called tcbhead_t defined in system-dependent header files such as nptl/sysdeps/x86_64/tls.h. It holds TLS related information. It contains at least an integer member called multiple_threads which indicates if the process is running in multi-thread mode.

The second member of pthread struct is also a struct called list_t defined in nptl/sysdeps/pthread/list.h.

The third and fourth members of pthread struct are thread ID and thread group ID (both are of pid_t type).

Other members of pthread struct which are of interest: int cancelhandling for cancellation information, int flags for thread attributes, start_routine for start position of the code to be executed for the thread, void *arg for the argument to start_routine void *stackblock and size_t stackblock_size for thread-specific stack information.

Since pthread struct is opaque, how can one obtain the above information, or more precisely, how can one obtain the offsets of these members within the pthread struct ? We can use the known information and search for the memory region pointed by pthread_t, as in this code snippet.

Acronyms relevant to Executable and Linkable Format (ELF)

Useful books and references

Executable and Linkable Format

How is an executable binary in Linux being executed ?

What about ld.so ?

Compile your own ld.so

How does ld.so work ?

Here is the call graph, which is worth a thousand words

.plt section

.init, .fini, .preinit_array, .init_array and .fini_array sections

What user functions will be executed before main and at program exit?

What do _start and __libc_start_main do?

Here is the call graph, which is worth a thousand words

How to find the address of main of an executable binary ?

PIC, TLS, and AMD64 code models

_GLOBAL_OFFSET_TABLE_, .got.plt section, and DYNAMIC segment

Runtime Relocation

R_X86_64_COPY relocation type

PIC or no PIC

Troubleshooting ld.so

What is “error while loading shared libraries: requires glibc 2.5 or later dynamic linker” ?

What is “Floating point exception” ?

What is “…/libc.so.6: version `GLIBC_2.4’ not found “ ?

What is “FATAL: kernel too old” ?

Exploring Glibc’s pthread_t

What about `ld.so` ?

Compile your own `ld.so`

How does `ld.so` work ?

`.plt` section

`.init`, `.fini`, `.preinit_array`, `.init_array` and `.fini_array` sections

What user functions will be executed before `main` and at program exit?

What do `_start` and `__libc_start_main` do?

How to find the address of `main` of an executable binary ?

`_GLOBAL_OFFSET_TABLE_`, `.got.plt` section, and `DYNAMIC` segment

`R_X86_64_COPY` relocation type

Troubleshooting `ld.so`

Exploring Glibc’s `pthread_t`