| |
About Introducing Your Own
Code |
|
|
by SiuL+Hacky |
 |
|
Courtesy of Fravia's page of
reverse engineering |
|
| |
hi +fravia,
it's curious, linux surprises myself. Two or three essays ago i was
convinced they would be the last ones commenting general facts. I was
thinking that new articles would be just explanations of some
interesting
protection scheme, but nothing else. Well, my friend there are lot of
things left :-).
And think that protectors have not started to think, this
is
just a kid playing with the system.
Hope it'll like you,
SiuL+Hacky
ps. by the way, have you ever think of how incredibly well protected is
the whole fucking machine outthere. The problem is that you need a
good deal of technical background or some kind of Richard Feynman close
to
you :-). It's a good job we have not lost our chance to learn !
|
|
| |
There is a crack, a crack
in everything That's
how the light gets in |
|
Rating
|
( )Beginner
( )Intermediate (X)Advanced ( )Expert
|
|
About Introducing Your Own
Code
Written by SiuL+Hacky
I must admit one of the funny ways to crack applications is using
dlls from the
program; it is what we've called Object Oriented Cracking. In
my opinion that's
one of the more powerful and witty approaches i've seen. Despite this
fact, i faced in my
last windows-cracking-times, the problem of calling routines that were
not present in
dlls, but executables.
The general problem is: imagine a program that validates some code.
As a matter of
fact, what the interesting routine does is not important, but think
about you have to
modify the execution flow in order to test different things. Sometimes
you can just patch
the code, but many times the amount of foreign code you must enter is
larger that the code
you can fit into "dead areas". You can always, of course,
overwrite code, but i
always thought that's not a quite elegant approach. One example is a
validation routine
that it's not easy to clone, because some kind of data is initialized,
and you have to
call just in the right moment, neither before nor after. Now i want to
make a brute force
attack and i need to give room for my code that generates new
alternatives, test them, and
show me somehow that i've been successful.
In this essay you'll need no
particular tool that is not available in any distribution:
gcc, ld-so, binutils ... You may need for aditional information the
source code of some
programs. If it does not come with your distribution, at least Debian
and RedHat have
source codes available in their respective ftp sites and
mirrors.
In this essay i'm gonna talk about general technics and ideas that
are not applied to
any particular target.
Now we're gonna see different ways to do it. Some of them proved to
be impossible,
others were rather complicated and others were finally good. You'll see
if some of them
could deserve further development, and it would be happy if someone can
do useful things
with options i refused (or even better, new ones).
1. MODIFYING ELF HEADERS
The first idea was to simply ADD the code. As i briefly explained in
my very first
linux essay, linux uses mainly ELF format executables, whose structure
is really modular,
so one could think about just appending your code to some part of the
executable. There
two documents about ELF format (that i know of), that will answer many
of the question
that can arise:
elf.hps -> Format
gory details
elf.ps.gz ->
Programmers guide
Firstly a small review about ELF format. Three elements are
important:
- Elf header: is located at the beginning of the
file, it carries
information about the size and location of the other two parts
- Program headers: it's an array where the
different segments that will
be loaded to memory are defined. Usually the array has 5 members.
Here's an example:
(vaddrs and paddrs stand for virtual and physical address)
Program Header:
PHDR off 0x00000034 vaddr 0x08048034 paddr 0x08048034 align 2**2
filesz 0x000000a0 memsz 0x000000a0 flags r-x
INTERP off 0x000000d4 vaddr 0x080480d4 paddr 0x080480d4 align 2**0
filesz 0x00000013 memsz 0x00000013 flags r--
LOAD off 0x00000000 vaddr 0x08048000 paddr 0x08048000
align 2**12
filesz 0x0000055c memsz 0x0000055c flags r-x
LOAD off 0x0000055c vaddr 0x0804955c paddr 0x0804955c align 2**12
filesz 0x000000d4 memsz 0x000000d8 flags rw-
DYNAMIC off 0x00000598 vaddr 0x08049598 paddr 0x08049598 align 2**2
filesz 0x00000098 memsz 0x00000098 flags rw-
The code segment is present in the third element. Note that
alignment is 4096, the size
of each page of virtual memory.
- Section headers: it's an array of more or less
standard elements, that
makes more comprehensive the distribution of data and code. For
instance, the general
"data" set can be split into read-only data, initialized
data, uninitialized
data, etc ... But you can define non-standard sections if you feel
like, regarding the
section should be mapped in one of the segments specified in the
program headers array.
Each element is a structure that references the data/code by means
of a file offset.
This is just a reminding if someone have already read about elf
structure, i'm not
teaching elf structure. If you're lazy, as me, you will not want to
read the docs i
pointed before, but do not expect to understand the what i'm gonna
discuss now.
Now let's suppose i have a piece of code i want to add to the
program. I could think
about creating a new code section and append it. Ok, that's not
complicated, as you can do
it with a simple program named objcopy. The program comes with
the binutils
package, that every distribution carries. One of the features of
objcopy is to add a
section given a file name.
That sounds good, let's suppose i program my new code in C, i
compile it (i can link it
statically to ease the task), and then objcopy glues it to the
protected crap.
Well it works, but ... objcopy does nothing to map the new
section onto the
executable segment, where all the code is loaded. In other words my new
code is
unreachable, because it is never loaded.
A pity, but now you'll begin to realize the difficulties of the job,
you must edit by
yourself the elf headers and try to make your new section reachable
from the rest of the
code. Our code is not located physically contiguous to the rest of the
executable code:
| ------------ elf headers
and so on ------------ |
| ------------ data from
different kinds ------------ |
| ------------ native code
------------ |
| ------------ more
data/code ------------ |
| ------------ our code
------------ |
I'm locating the new section as the last one, not only because it's
the easiest
solution, but it's the way objcopy does. The problem here is
that we can't
overlap segments (nor sections), so due to the fact you can't (AFAIK)
define two code
segments, you'll get nothing. The only thing you could do is expanding
the current code
section (.text section) and append at the end the new piece of
code. Of course
you'll have to expand the segment too, and other things i'll tell you
soon (that makes the
task a nightmare).
But before tackling the problems of expanding a .text
section, let's face the
problem of modifying elf files. Objcopy can handle section
attributes and so on,
but do not even think about modifying attributes from program headers:
- By hand: The structures of elf files are defined
in the file /usr/include/elf.h.
Knowing the meaning of each element (and you can achieve that
reading the document i told
you above), it's just a programming problem. On the other hand is
really difficult to be
aware of all the details a small modification implies in the rest
of the elements. I built
small programs that list elf elements (that information is
available through objdump too),
and modify particular elements.
Now if somebody's interested on writing X-Windows apps (with some of
the new and
attractive graphic libraries for instance), a VERY USEFUL tool could be
to program a
graphical ELF editor, that could have as a "maximum"
challenge: insert data/code
into sections. If anybody has the guts to try it i could help him with
ELF details (if
necessary, of course).
- Using Libelf: yes a library to edit elf file is
available. That sounds
good, but the rest is not so good. The lib was written by Michael
"Tired" Riepe,
and it's a good work, a good incomplete work. The last version
(last update lot of months
ago) i knew of was libelf-0.6.4. The library provides a fairly
complete API to deal with
elf files ... but the problems is that there's no documentation at
all. Of course after
delving a bit inside elf details, you can "guess" some of
the features of the
functions, but you'll realize it's not the best way to accomplish
things. As a personal
opinion i prefer to control by myself what i'm doing with the elf
file, especially if are
going to read simple attributes. It doesn't mean libelf should not
be considered in some
cases, in fact, reading elf headers, section headers and program
headers is really simple.
But when you face more complicated jobs, as adding parts, you must
know for sure what the
program is supposed to do (behind the scene) and what is not. Well,
you can always look up
the source code ... :-)
- Using BFD. What's that ? Well, you could find
that name playing with
binutils source code, as it is used by the most important GNU
elements in your system (for
instance debugger and linker). I don't know much about BFD but if
it is used in such
packages is really a guaranty. BFD lets you manipulate different
object file formats
(there are a lot if you wondered) in a common way. You work with a
common
object-bfd-format and you must not care about elf details. I'm sure
mastering BFD you can
do what you want with executables, but it seemed to me i had to
spend too much hours on it
and I preferred to think about other ways. Again the docs are
partial, not updated and
probably wrong :).
Ok, unless you could rebuild completely an executable with BFD,
you'll need
modifications in the elf header in order to expand .text
section. I think there
are two different groups of modifications you have to do, one more
obvious, and the second
more complicated:
- Let's suppose there are N sections, and
text section is number i.
We are going to append our code, so you have to physically insert
it. You have to make
modifications regarding program headers and section headers, in
order to make the new code
"fit":
- All the references to file offsets greater that the insertion
point would have to be
increased (p_offset, sh_offset if we talk elf
language).
- The .text section and code segment must be larger. As much as
the size of the new code (sh_size,
p_filesz).
- The new code occupies virtual addresses that were previously
assigned to other stuff, so
the starting virtual and physical address of sections
i+1 to N
(and the suitable segments at program headers) must be
appropriately increased (p_vaddr,
p_paddr, sh_addr).
- That was the easy part, i did it (don't include it because it's
not a very educational
material, but you may ask for it if you want), but it's not enough
... it does not work.
Why ? let's see a usual distribution of sections in an easy
program:
Sections:
Idx Name Size VMA LMA File off Algn
0 .interp 00000013 080480d4 080480d4 000000d4 2**0
CONTENTS, ALLOC, LOAD, READONLY, DATA
1 .hash 00000098 080480e8 080480e8 000000e8 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
2 .dynsym 00000130 08048180 08048180 00000180 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
3 .dynstr 000000c7 080482b0 080482b0 000002b0 2**0
CONTENTS, ALLOC, LOAD, READONLY, DATA
4 .rel.got 00000008 08048378 08048378 00000378 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
5 .rel.bss 00000008 08048380 08048380 00000380 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
6 .rel.plt 00000030 08048388 08048388 00000388 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA
7 .init 0000002c 080483c0 080483c0 000003c0 2**4
CONTENTS, ALLOC, LOAD, READONLY, CODE
8 .plt 00000070 080483ec 080483ec 000003ec 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
9 .text 000000c8 0804845c 0804845c 0000045c 2**2
CONTENTS, ALLOC, LOAD, READONLY, CODE
10 .fini 0000001c 08048530 08048530 00000530 2**4
CONTENTS, ALLOC, LOAD, READONLY, CODE
11 .rodata 00000010 0804854c 0804854c 0000054c 2**0
CONTENTS, ALLOC, LOAD, READONLY, DATA
12 .data 00000004 0804955c 0804955c 0000055c 2**2
CONTENTS, ALLOC, LOAD, DATA
13 .ctors 00000008 08049560 08049560 00000560 2**2
CONTENTS, ALLOC, LOAD, DATA
14 .dtors 00000008 08049568 08049568 00000568 2**2
CONTENTS, ALLOC, LOAD, DATA
15 .got 00000028 08049570 08049570 00000570 2**2
CONTENTS, ALLOC, LOAD, DATA
16 .dynamic 00000098 08049598 08049598 00000598 2**2
CONTENTS, ALLOC, LOAD, DATA
17 .bss 00000004 08049630 08049630 00000630 2**2
ALLOC
18 .comment 00000064 00000000 00000000 00000630 2**0
CONTENTS, READONLY
19 .note 00000064 00000064 00000064 00000694 2**0
CONTENTS, READONLY
If you look at some function present in the symbol table that is
dynamically loaded (printf
for instance), you'll get an address:
080483fc DF *UND* 00000024
printf
so 80483fc is the supposed address for
printf
in our program. There you'll not find printf's code,
but a jmp
[index_table]. The contents of the table are initialized by dynamic
loader. We see, we're
lucky because that address (80483fc) is
from .plt
section, and that address is not modified by the insertion of the new
code. But if we go
on, we see:
080483fc jmp *0x804957c
so the data table that keeps the real address of dynamic function,
located in section .got,
is disturbed by our patch. If we add, say 10k of code, *0x804957c
is not pointing .got section anymore. It is pointing some
innocent instruction
inside .text section.
You would need to patch .plt section, that could be big,
and well, i gave up
here. The problem is not unsolvable, may be patient programmer could do
:-). Anyway you'll
agree the procedure is getting more confused and less elegant.
2. TRACING A PROCESS
The idea of this approach is taken from ltrace written by
Juan CÈspedes. What
he does is read symbol table and place breakpoints before calling
dynamically linked
functions. Read the stack, call the function and keep the returned
value. That job is
accomplished by ptrace, a system call that let you "debug" a
process.
What i want to do here is to patch code and data on the fly in order
to suit my needs:
change the calling parameter, create calling loops to validating
functions, ... The
program will know nothing :). Ptrace provides elements to modify
code/data, read
code/data, insert breakpoints, resume execution, trap signals, trap
system calls, etc ...
almost all you want. I'd like to remark that, WITH PTRACE YOU HAVE GOT
ALL THE POWER.
The main lacks for this approach are:
- It's slow. If you're trying a brute force approach you
need all the speed from
you computer, and ptrace calls are not fast. Don't misunderstand,
may be just
microseconds, but in some scenario it could be important.
- All my access to victim's address space is done via
ptrace. I cannot modify
data/code directly, moreover i cannot jump directly to foreign code
nor return from it. Of
course, i can give my own functions to be called on trap signals
and so one, but it's like
a firewall that separates both processes.
It would be nice to manipulate the kernel and make the trick, but
that's far beyond my
available time ... and you'll see there are other alternatives. To get
this approach
closer to you let's see a simple piece of code dealing with ptrace. It
creates a child
process that will be debugged by its parent, then the child process
executes the victim
program. I will not explain what is a child/parent process, or how a
fork call works, but
there are thousands of books that will inform you. On the other hand i
would like to tell
you about a good guide for ptrace's call, but sadly i did not find it,
so consult man
page.
#include
#include
#include
#include
#include
#include
#include
void main(int ARGC, char *ARGV[]){
int pid, status;
pid=fork();
if (pid==-1) perror("fork");
else if (pid==0){
/* this code will be executed ONLY by the child process */
signal(SIGTRAP, SIG_IGN);
if ( ptrace(PTRACE_TRACEME, 0, 0, 0) == -1)
perror("ptrace");
printf("\nProcess controlled\n");
if (execlp("./the_victim_program",
"./the_victim_program", NULL)== -1)
perror("execlp");
}
else{
/* this code will be executed ONLY by the parent process
*/
/* it will wait until receives a signal from the child process */
wait(&status);
if ( WIFSTOPPED(status) ) printf ("\nSignal: %i\n",
WSTOPSIG(status) );
/* the debugged process is stopped (with sigtrap) and the
parent */
/* receives the signal. The process is stopped before starting. Here */
/* you can change whatever you want. With "sigaction" you
can handle */
/* signals (for instance sigtrap) with the function you fancy */
/* After changing whatever the execution is continued */
if (ptrace(PTRACE_CONT, pid, 0, WSTOPSIG(status))==-1)
perror("ptrace");
wait(&status);
/* wait until child process exits */
}
}
3. MANIPULATING DYNAMIC LINKER
What we are trying to do is to write our own code and make it
visible from our victim's
code. We are trying to share our code, if you want to put that way. But
that is done
everyday with no artificial method: dynamic linker. So we are going to
try some trick or
otherwise try-to-emulate what the linker does.
The most obvious and easy trick, but not less powerful is the well
known LD_PRELOAD
trick. You can find it documented (though the procedure has to be tuned
:-) together with
dynamic liker (ld-so) source code. It was mentioned in
zlib package and
other docs too. The trick was also explained by
adq in a
previous essay. Anyway, as far as i've read it is always used to fake
some function
dynamically loaded, but never to access victim's code from the faked
function.
Example:
Suppose our evil program uses some common function as sin,
you could preload
it (via LD_PRELOAD env. variable) and load in memory
your
"holy" code. You could always resolve the problem of the
actual function along
with cos function, for example ;). Now pass (or hardwire) the
address of a
validating function to our preloaded library (or to our preloaded sin
function if you feel
like that), and patch the executable (physically or ptracing it) to
call the faked sin
function instead of the validating function. Then from your faked
function you
"see" the calling address space and you can call the
validating function a
zillion times until you get a good value, and you can then return to
the caller code :-).
You get ????
Before
LOCAL CODE
call validating_function
:validating_function
....
ret
check_it
After
LOCAL CODE
LIBRARY CODE
call sin --------------------> :my_preloaded_faked_sin
...
call validating function
ret
check_it
I like the above method especially because it's fairly easy and is
the dynamic loader
the one who makes the dirty job of loading pages into memory. Moreover
the patching is
minimum, but it is good to know other alternatives, that will do more
or less what the
dynamic linker does in background. Keep in mind this easy method can
only be used in
dynamic executables, that although they are the most usual,
sometimes you find
statically linked ones.
Firstly you can try to load a library in a way similar to the way
you do in Windoze.
The most important drawback is that you need not only to overwrite code
from the
executable, but you need that the functions involved were previously
used by the program
(as far that they have to be present in the dynamic table). The
functions involved are:
void *dlopen (const char *filename,
int flag)
void *dlsym (void *handle, char
*symbol)
int dlclose (void *handle)
so you will load by yourself (from the victim's code) a library and
get the address of
a particular function or symbol. The you can jump to it.
Another alternative, more sophisticated but easier to implement is
to introduce the
system call mmap. It is a very interesting service that maps a
file into memory
with the corresponding attributes. If you wondered is the way dynamic
loader
"loads" libraries into memory:
void * mmap (void *start, size_t
length, int prot , int
flags, int fd, off_t offset);
mmap, remember is a system call, no library is necessary.
It's a great gift
from kernel guys. The system call returns a pointer to the loaded area.
If you use in the
"flag" parameter the value MAP_FIXED, the parameter start
should indicate the
address where YOU WANT the code/data to be loaded. I didn't try yet,
but what about if you
overwrite the code previously loaded ... it would not work as it is
read-only data..
Well, that's all. Lots
of material to play with. I have not given ready-to-make-recipes
because if you have the interest you will build by yourself the
procedures you need with
the clues given here. But you can always ask whatever you feel is not
clear enough. Of
course, if you've just installed linux for the first time, this will
not be useful for you
and you better read previous essays
in order learn
something
I WILL bother explaining you that you should TEST
by yourself what is
explained here and try to find new ways to make accesible your code to
a non-colaborating
process.
You are deep inside fravia's page of reverse
engineering, choose your
way out:
homepage
links
+ORC
bots wars
students' essays
counter measures
bots wars
antismut CGI tricks
academy database
tools
javascript tricks
cocktails
search_forms
mail_fravia+
Is software reverse engineering illegal?