Mac OS X 64 bit Assembly System Calls

After reading about shellcode in Chapter 5 of Hacking: The Art of Exploitation, I wanted to go back through some of the examples and try them out. The first example was a simple Hello World program in Intel assembly. I followed along in the book and had no problems reproducing results on a 32 bit Linux VM using nasm with elf file format and ld for linking.

Then I decided I wanted to try something similar but with a little bit of a challenge: write a Mac OS X 64 bit “hello world” program using the new fast ‘syscall’ instruction instead of the software interrupt based (int 0x80) system call, this is where things got interesting. First and foremost, the version of Nasm that comes with Mac OS X is a really old version. If you want to assemble macho64 code, you’ll need to download the lastest version.

[email protected]:~$ nasm -v
NASM version 2.09.03 compiled on Oct 27 2010

I figured I could replace the extended registers with the 64 bit registers and the int 0x80 call with a syscall instruction so my first attempt was something like this

section .data
hello_world     db      "Hello World!", 0x0a

section .text
global _start

mov rax, 4              ; System call write = 4
mov rbx, 1              ; Write to standard out = 1
mov rcx, hello_world    ; The address of hello_world string
mov rdx, 14             ; The size to write
syscall                 ; Invoke the kernel
mov rax, 1              ; System call number for exit = 1
mov rbx, 0              ; Exit success = 0
syscall                 ; Invoke the kernel

After assembling and linking, I got this

[email protected]:~$ nasm -f macho64 helloworld.s
[email protected]:~$ ld helloworld.o 
ld: could not find entry point "start" (perhaps missing crt1.o) for inferred architecture x86_64

Apparently Mac OS X doesn’t use _start for linking, instead it just uses start. After removing the underscore prefix from start, I was able to link but after running, I got this

[email protected]:~$ ./a.out
Bus error

I was pretty stumped at this point so I headed off to Google to figure out how I was supposed to use the syscall instruction. After a bunch of confusion, I stumbled upon the documentation and realized that x86_64 uses entirely different registers for passing arguments. From the documentation:

The number of the syscall has to be passed in register %rax.
rdi - used to pass 1st argument to functions
rsi - used to pass 2nd argument to functions
rdx - used to pass 3rd argument to functions
rcx - used to pass 4th argument to functions
r8 - used to pass 5th argument to functions
r9 - used to pass 6th argument to functions
A system-call is done via the syscall instruction. The kernel destroys registers rcx and r11.

So I tweaked the code with this new information

mov rax, 4              ; System call write = 4
mov rdi, 1              ; Write to standard out = 1
mov rsi, hello_world    ; The address of hello_world string
mov rdx, 14             ; The size to write
syscall                 ; Invoke the kernel
mov rax, 1              ; System call number for exit = 1
mov rdi, 0              ; Exit success = 0
syscall                 ; Invoke the kernel

And with high hopes that I’d see “Hello World!” on the console, I still got the exact same ‘Bus error’ after assembling and linking.

Back to Google to see if others had tried a write syscall on Mac OS X. I found a few posts of people having success with the syscall number 0x2000004 so I thought I’d give it a try. Similarly, the exit syscall number was 0x2000001. I tweaked the code and BINGO! I was now able to see “Hello World” output on my console but I was seriously confused at this point; what was this magic number 0x200000 that is being added to the standard syscall numbers?

I looked in syscall.h to see if this was some sort of padding (for security?) I greped all of /usr/include for 0x2000000 with no hints what-so-ever. I looked into the Mach-o file format to see if it was related to that with no luck.

After about an hour and a half of looking, I spotted what I was looking for in ‘syscall_sw.h’

 * Syscall classes for 64-bit system call entry.
 * For 64-bit users, the 32-bit syscall number is partitioned
 * with the high-order bits representing the class and low-order
 * bits being the syscall number within that class.
 * The high-order 32-bits of the 64-bit syscall number are unused.
 * All system classes enter the kernel via the syscall instruction.
 * These are not #ifdef'd for x86-64 because they might be used for
 * 32-bit someday and so the 64-bit comm page in a 32-bit kernel
 * can use them.

#define SYSCALL_CLASS_NONE	0	/* Invalid */
#define SYSCALL_CLASS_MACH	1	/* Mach */	
#define SYSCALL_CLASS_UNIX	2	/* Unix/BSD */
#define SYSCALL_CLASS_MDEP	3	/* Machine-dependent */
#define SYSCALL_CLASS_DIAG	4	/* Diagnostics */

Mac OS X or likely BSD has split up the system call numbers into several different “classes.” The upper order bits of the syscall number represent the class of the system call, in the case of write and exit, it’s SYSCALL_CLASS_UNIX and hence the upper order bits are 2! Thus, every Unix system call will be (0×2000000 + unix syscall #).

Armed with this information, here’s the final x86_64 Mach-o “Hello World”

section .data
hello_world     db      "Hello World!", 0x0a

section .text
global start

mov rax, 0x2000004      ; System call write = 4
mov rdi, 1              ; Write to standard out = 1
mov rsi, hello_world    ; The address of hello_world string
mov rdx, 14             ; The size to write
syscall                 ; Invoke the kernel
mov rax, 0x2000001      ; System call number for exit = 1
mov rdi, 0              ; Exit success = 0
syscall                 ; Invoke the kernel

And here’s the output

[email protected]:~$ nasm -f macho64 helloworld.s
[email protected]:~$ ld helloworld.o 
[email protected]:~$ ./a.out
Hello World!

How is glibc loaded at runtime?

I’ve been looking into address-space layout randomization (ASLR). ASLR relies on randomizing the base address of things like shared libraries, making return-to-libc attacks more difficult.

I understood the basics of ASLR but I still had a lot of questions. How are shared libraries, like libc, loaded at runtime? What is the global offset table? What is the procedure linkage table? What is a position independent executable? In this post, we’re going to look at all of these.

Back in the Day

In “the olden days” libraries used to be hard coded to be loaded at a fixed address in memory space. Runtime linkers had to deal with relocating conflicting hard coded addresses. Windows, to some extent, still does this.

PIC - Position Independent Code

Then came along Position Independent Code which simply means that the code (usually shared libraries) can be loaded at any address in memory-space and relocations are no longer a problem. In order to do that, binaries added sections for the GOT and the PLT.

Global Offset Table

Every ELF executable has a section called the Global Offset Table or the GOT for short. This table is responsible for holding the absolute address of functions in shared libraries linked dynamically at runtime.

[email protected]:~$ objdump -R ./hello_world

./hello_world:     file format elf32-i386

OFFSET   TYPE              VALUE
08049564 R_386_GLOB_DAT    __gmon_start__
08049574 R_386_JUMP_SLOT   __gmon_start__
08049578 R_386_JUMP_SLOT   __libc_start_main
0804957c R_386_JUMP_SLOT   printf
Procedure Linkage Table

Just like the GOT, every ELF executable also has a section called the Procedure Linkage Table or PLT for short (not to be confused with BLT (Bacon Lettuce Tomato) :-) ). If you’ve read disassembled code, you’ll often see function calls like [email protected]. That’s a call to the printf in the procedure linking table. The PLT is sort of like the spring board that allows us to resolve the absolute addresses of shared libraries at runtime.

[email protected]:~$ objdump -d -j .plt ./hello_world

./hello_world:     file format elf32-i386

Disassembly of section .plt:

08048270 <[email protected]>:
 8048270:       ff 35 6c 95 04 08       pushl  0x804956c
 8048276:       ff 25 70 95 04 08       jmp    *0x8049570
 804827c:       00 00                   add    %al,(%eax)

08048280 <[email protected]>:
 8048280:       ff 25 74 95 04 08       jmp    *0x8049574
 8048286:       68 00 00 00 00          push   $0x0
 804828b:       e9 e0 ff ff ff          jmp    8048270 <_init+0x18>

08048290 <[email protected]>:
 8048290:       ff 25 78 95 04 08       jmp    *0x8049578
 8048296:       68 08 00 00 00          push   $0x8
 804829b:       e9 d0 ff ff ff          jmp    8048270 <_init+0x18>

080482a0 <printf@plt>:
 80482a0:       ff 25 7c 95 04 08       jmp    *0x804957c
 80482a6:       68 10 00 00 00          push   $0x10
 80482ab:       e9 c0 ff ff ff          jmp    8048270 <_init+0x18>
The GOT, The PLT, and the Linker

How do these all work together to load a shared library at runtime? Well it’s actually pretty cool. Lets walk through the first call to printf. [email protected], which is not really printf but a location in the PLT, is called and the first jump is executed.

080482a0 <printf@plt>:
 80482a0:       ff 25 7c 95 04 08       jmp    *0x804957c
 80482a6:       68 10 00 00 00          push   $0x10
 80482ab:       e9 c0 ff ff ff          jmp    8048270 <_init+0x18>

Notice that this jump is a pointer to an address. We’re going to jump to the address pointed to by this address. The 0x804957c is an address in the GOT. The GOT will eventually hold the absolute address call to printf, however, on the very first call the address will point back to the instruction after the jump in the PLT - 0x80482a6. We can see this below by looking at the output of the GOT. Essentially we’ll execute all of the instructions of the [email protected] the very first call.

(gdb) x/8x 0x804957c-20
0x8049568 <_GLOBAL_OFFSET_TABLE_>:      0x0804949c      0xb80016e0      0xb7ff92f0      0x08048286
0x8049578 <_GLOBAL_OFFSET_TABLE_+16>:   0xb7eafde0      0x080482a6      0x00000000      0x00000000

In the PLT code, an offset is pushed onto the stack and another jmp is executed

080482a0 <printf@plt>:
 80482a0:       ff 25 7c 95 04 08       jmp    *0x804957c
 80482a6:       68 10 00 00 00          push   $0x10
 80482ab:       e9 c0 ff ff ff          jmp    8048270 <_init+0x18>

This jump is a jump into the eventual runtime linker code that will load the shared library which contains printf. The offset, $0x10, that was pushed onto the stack tells the linker code the offset of the symbol in the relocation table (see objdump -R ./hello_world output above), printf in this case. The linker will then write the address of printf into the GOT at 0x804957c. We can see this if we look at the GOT after the library has been loaded.

(gdb) x/8x 0x804957c-20
0x8049568 <_GLOBAL_OFFSET_TABLE_>:      0x0804949c      0xb80016e0      0xb7ff92f0      0x08048286
0x8049578 <_GLOBAL_OFFSET_TABLE_+16>:   0xb7eafde0      0xb7edf620      0x00000000      0x00000000

Notice that the previous address, 0x80482a6, has been replaced by the linker with 0xb7edf620. To confirm that this indeed is the address for printf, we can start a disassemble at this address

(gdb) disassemble 0xb7edf620
Dump of assembler code for function printf:

Since the library is now loaded and the GOT has been overwritten with the absolute address to printf, subsequent calls to the function [email protected] will jump directly to the address of printf! All of this also has the added benefit that a shared library is not loaded until a function in it’s library is loaded – in other words, a nice form of “lazy-loading!”

Turning off buffer overflow protections in GCC

As I learn more and more about buffer overflow exploitation I’m realizing that it’s actually pretty hard to run the examples that teach you how to exploit buffer overflows. GCC (and other compilers) have built in support for mitigating the simple buffer overflows and it’s turned on by default.

With GCC you have to compile with the -fno-stack-protector option otherwise you get ***stack smashing detected***, this is pretty well known and documented all over the Internet.

However, additionally you’ll need to disable the FORTIFY_SOURCE option otherwise you’ll get “Abort trap” if you try to do a buffer overflow that uses something like strcpy or memcpy.

To disable it, simply compile with the flag -D_FORTIFY_SOURCE=0 (e.g. gcc -g -fno-stack-protector -D_FORTIFY_SOURCE=0 -o overflow_example overflow_example.c)

FORTIFY_SOURCE is enabled by default on a lot of Unix like OSes today. It’s original development dates back to 2004 but I don’t think it was turned on by default originally

#  else
#    define _FORTIFY_SOURCE 2   /* on by default */
#  endif

That means that when you include something like string.h, at the bottom of string.h is the following (at least on Mac OS X):

Security Checking Functions
#if defined (__GNUC__) && _FORTIFY_SOURCE > 0 && !defined (__cplusplus)
/* Security checking functions.  */
#include <secure/_string.h>

So when you compile, instead of seeing a call to strcpy

Standard call to strcpy

0x0000000100000d50 <main+188>:  lea    rdi,[rbp-0x20]
0x0000000100000d54 <main+192>:  call   0x100000dba <dyld_stub_strcpy>
0x0000000100000d59 <main+197>:  lea    rdx,[rbp-0x20]

you’ll see a call to the strcpy_chk function in your disassembled code

Protected call to strcpy

0x0000000100000d50 <main+192>:  mov    edx,0x8
0x0000000100000d55 <main+197>:  call   0x100000daa <dyld_stub___strcpy_chk>
0x0000000100000d5a <main+202>:  lea    rdx,[rbp-0x20]

This will wreck havoc on any of your simple buffer overflow exploiting examples so make sure you disable it when you compile. Happy exploiting!