'Signing' a compressed file system for app verification

[BillS]

A friend asked me, "How do you know that someone didn't replace your KNOPPIX compressed file system with a tainted one containing malware and such?"

I do verify portions of systems by appending a MD5 hash (plus a bit of arcania) to the end of some executables so that I can later verify they are the ones I supplied. The executables don't mind the extra 32 bytes,

But the clooped file system DOES seem to mind. I only hash as much as is needed to detect changes. I then append the MD5 hash and a time stamp to the end of the file.

I have been through the sourcecode in "cloop-2"; and not found anything that explicitly compares the end of compressed data with the file size. And I have compressed FSs with odd byte sizes, so there is no specific size multiple. So why the failures?

BillS

This is the result of loading and mounting a 'signed' compressed file.

root@lp2:/tmp# losetup /dev/cloop2 /tmp/KNOPPIX_NOV_29_04
root@lp2:/tmp# losetup -a
root@lp2:/tmp# head -4 KNOPPIX_NOV_29_04 
#!/bin/sh
#V2.0 Format
insmod cloop.o file=$0 && mount -r -t iso9660 /dev/cloop $1
exit $?
root@lp2:/tmp# mount -r -t iso9660 /dev/cloop2 /mnt/test
mount: wrong fs type, bad option, bad superblock on /dev/cloop2,
       or too many mounted file systems

And then the same file without the signature (simply lacking the 32 bytes):

root@lp2:/tmp# losetup /dev/cloop /extra/CompressedFS/KNOPPIX_NOV_29_04
root@lp2:/tmp# losetup -a
root@lp2:/tmp# mount -r -t iso9660 /dev/cloop /mnt/test
root@lp2:/tmp# losetup -a
root@lp2:/tmp# ls /mnt/test
bin   cdrom  etc     home  mnt   proc  sbin  tmp  var
boot  dev    floppy  lib   none  root  sys   usr  vmlinuz
root@lp2:/tmp# cmp KNOPPIX_NOV_29_04 /extra/CompressedFS/KNOPPIX_NOV_29_04
cmp: EOF on /extra/CompressedFS/KNOPPIX_NOV_29_04

[Doctor]

Does the cloop stuff seek to the end of the file anywhere?

I remember that GIF headers are at the front of a file and the real headers for a zip file are at the end, so a person could make a dual format file. Maybe it was the other way around with the headers, but the point is, maybe it explicitly looks at the end of the file for something important.

[BillS]

'cloop-2' code does 'lseek()' calls for either current position only, or to an absolute position using an offset value already stored in the compressed file index. And it is curious that the failure is in the mount (unless the losetup failed and didn't produce an error message.)

BillS

[Ghandalfar]

why not just put md5sum into md5.txt file?

If one can replace your Knoppix, then he can also generate new md5 for it.

[BillS]

Not that simple to bypass. Without going into details;
"appending a MD5 hash (plus a bit of arcania) to the end"

1) Maybe it is a SHA1 rather than an MD5
2) The hash is 'salted'; not simple to replicate
2) 'arcania' is made of things like epoch time and size and such
3) values are interleaved at a byte level.

Suffice to say, the substitution of one file for another is a more complex task.
Placing values in a seperate file leave the question of that external files integraty at question.


Nothing wholey self-contained can ever be secure.
Given enough resources, it can be comprimised.
But the quantity of resources to compromise it can
be forced to extremely high levels.

BillS

[BillS]

OK ... I found the check that stops the cloop load.
In 'compressed_loop.c' (AKA cloop.o) is a check if this is a block device, and that the end of the file matches the last offset. It is making the incorrect assumption that the offset is short, rather than the file has been extended.

 if(!isblkdev &&
    be64_to_cpu(clo->offsets[ntohl(clo->head.num_blocks)]) != inode->i_size)
  {
   printk(KERN_ERR "%s: final offset wrong (%Lu not %Lu)\n",
          cloop_name,
          be64_to_cpu(clo->offsets[ntohl(clo->head.num_blocks)]),
          inode->i_size);
   vfree(clo->zstream.workspace); clo->zstream.workspace=NULL;
   goto error_release_free_all;
  }

The result is a message in 'dmsg' of:

cloop: Initializing cloop v2.00
cloop: /tmp/KNOPPIX_NOV_29_04: 1600 blocks, 65536 bytes/block, largest block is 65562 bytes.
cloop: final offset wrong (39821172 not 39821204)

Those two numbers are the file sizes before and after signing.

BillS